CN115130505A - FOCS fault diagnosis method based on improved residual shrinkage network - Google Patents

Abstract

FOCS fault diagnosis method based on improved residual shrinkage network. 1. Reading an output current signal of a current optical fiber current sensor (FOCS); 2. constructing an improved depth residual shrinkage network, fusing a batch standardization layer and a correction linear unit in a residual shrinkage construction module for sharing a threshold value between two channels, combining the advantages of two models, adding a new module to enhance the chemical ability and the coverage range of the network, and performing network optimization reconstruction by the model with stronger data upper limit and generalization ability; 3. establishing an FOCS drift, transformation ratio deviation and fixed deviation characteristic model, and constructing a fault data set; 4. training the constructed improved depth residual error shrinkage network by taking the FOCS signal as input and the fault label as output, and continuously optimizing the improved depth residual error shrinkage network model to ensure that the model is converged; 5. and acquiring FOCS fault diagnosis information. The method is suitable for the field of optical fiber current sensing measurement, completes the fault early warning function of FOCS, and improves the reliability of FOCS.

Description

FOCS fault diagnosis method based on improved residual shrinkage network

Technical Field

The invention belongs to the technical field of fault detection of optical fiber current sensors, and particularly relates to an FOCS fault diagnosis method based on an improved residual shrinkage network.

Background

The optical fiber current sensor (FOCS) has the advantages of good insulating property, high reliability, wide band frequency domain, good transient characteristic and the like, and is widely applied to high-voltage direct-current transmission engineering. As a measuring element of a primary system, the stable and reliable operation of the system is an important guarantee for relay protection, measurement and control and electric energy metering. Due to the influence of complex environments such as high and low temperature, aging and vibration in the station, the FOCS has nonlinear errors such as drift and transformation ratio, so that the measurement result is deviated, the performance of the FOCS is degraded, even operation accidents are caused, and the output signal is abnormal.

In actual work, the FOCS is subjected to complex scene factors such as large temperature change, vibration, electromagnetic interference and the like in the converter station, so that the reliability of a measurement result is reduced; as an optical interferometer, the internal structure is complex, a large number of optical and electronic components are used, and the performance can be gradually degraded due to loss and aging of any link. On the other hand, faults such as short circuit and grounding generated by the power grid and switching operation of switches in the station also have certain influence on the FOCS measurement result. The FOCS is started late and has short operation time, so that the phenomena of unreasonable field construction, insufficient operation and maintenance experience and the like are caused, and the stable operation of the FOCS is also seriously influenced. Due to the influence factors, the measurement accuracy of the FOCS is reduced, and in a serious case, a fault may be caused, so that the safe and stable operation of the power grid is seriously influenced.

At present, FOCS research is mainly focused on the aspects of sensing mechanism analysis, optical fiber material processing technology, optical device temperature characteristics, noise processing, and the like. With the great application of FOCS in the direct current transmission project, the research on related fault mechanisms and fault diagnosis technologies gradually becomes a hotspot in the field, the fault diagnosis further becomes a comprehensive technology integrating multiple subjects such as mathematics, signal processing, computer science, artificial intelligence and the like, and deep learning also becomes a hotspot direction in the research of multiple fields. With the rapid development of the methods based on the deep learning theory, such as speech recognition, image recognition and the like, the learner applies the deep learning method to the fault diagnosis direction, and obtains a great deal of results. The FOCS output signal belongs to a typical nonlinear non-stationary time series signal, so that the deep learning method is also suitable for the field of FOCS fault diagnosis.

Compared with the traditional neural network model, the convolutional neural network has the greatest characteristic that the local connection and the convolutional kernel parameter share two parts, and the characteristic enables the parameter quantity of the convolutional neural network to be greatly reduced and the training speed to be accelerated. Local connections refer to the connection of each layer of neurons to only the region of neurons above it, also known as the local receptive field. It not only reduces the number of weight parameters that need to be trained, but also enhances the connection between neighboring data. Weight sharing refers to using the same weight among multiple local connections in a convolutional neural network. Since each local part has various features such as the same edge information, it is only necessary to perform a convolution operation on the entire input using one convolution kernel. Compared with the traditional classification model, the convolutional neural network training process is end-to-end, and aims to integrate the extraction and classification of features into one structure, automatically extract the features of data through convolution and pooling operation, and finally perform classification output through a full connection layer, so that the identification error caused by manual feature extraction is avoided.

The process of convolutional neural network training is that local features of input layers are extracted, then the local features are combined to obtain global features, namely, a plurality of convolutional layers are combined, the number of layers of the convolutional layers is in a direct proportion relation with the extracted global features, and high-level features can be extracted, so that a deeper network has more excellent information extraction capability. However, as the network structure deepens, the gradient vanishing phenomenon occurs only by increasing the number of network layers through stacking of the network, that is, when the gradient is reversely propagated, repeated multiplication operations make the gradient infinitely small, and the performance tends to be saturated, and the phenomenon is called a degradation problem. The generation of degeneration problems makes deeper neural networks difficult to train. The deep residual error network is provided by the Rogowski, Microsoft, Asian institute of technology, on the VGG19 model, and the degradation problem is solved by introducing a residual error unit. The depth residual shrinking network is essentially the integration of the depth residual network, soft thresholding and attention mechanism, and has the characteristics of no need of manually screening characteristics, adaptation to strong noise, high redundant signals and the like, so that the depth residual shrinking network is widely concerned and improved by scholars. The method has the advantages that part of the noise-related characteristics and the threshold self-learning capability can be removed. Two models, DRSN-CS and DRSN-CW, respectively. The DRSN-CS has the advantages that part of noise-related features and threshold self-learning capability can be removed, but the performance is weak. Because each channel of the DRSN-CW has an independent threshold, the performance is better than that of the DRSN-CS network, but the DRSN-CW has excessive training parameters, large training amount and resource occupation. Therefore, two RSBUs modules are stacked again, a BN layer and a ReLU layer are added and combined, and the FOCS fault diagnosis model based on the improved deep residual error shrinkage network is provided, so that the training amount is reduced, and a better result is obtained.

Disclosure of Invention

Aiming at the FOCS fault diagnosis, the invention provides an FOCS fault diagnosis method based on an improved residual shrinkage network, two RSBUs modules are stacked again, a BN layer and a ReLU layer are added and combined, and an FOCS fault diagnosis model based on an improved deep residual shrinkage network is provided, so that the training amount is reduced, and a better result is obtained.

In order to achieve the purpose, the invention adopts the technical scheme that:

the fault diagnosis method of the FOCS fault diagnosis model of the improved residual shrinkage network is characterized by comprising the following specific steps of:

step 1: respectively taking output signals of FOCS normal state, drift deviation fault, transformation ratio deviation fault and fixed deviation fault to construct a fault data set;

and 2, step: dividing the data set: dividing a fault data set into a training set, a verification set and a test set, wherein the test set is only used as model evaluation and does not participate in training, and the method is as follows;

the experimental data sampling frequency is 4kHz, the sampling time of each section of signal is 5s, namely 20000 data sampling points of each piece of data are respectively collected, and drift fault signals, transformation ratio fault signals, fixed deviation fault signals and normal output signals are collected, so that 4 types of experimental data are counted;

dividing fault data according to four categories and marking corresponding labels, wherein 0 represents drift deviation fault, 1 represents variable ratio deviation fault, 2 represents fixed deviation fault, and 3 represents normal signal;

and step 3: the building, training and optimizing processes of the fault diagnosis model are as follows: constructing an improved depth residual shrinkage network model, setting network hyper-parameters and configuring network model structure parameters, and continuously optimizing the model to ensure that the model converges;

the improved depth residual error shrinkage network comprises the following steps:

(3-1) improving a FOCS fault diagnosis model of a deep residual error shrinkage network, wherein the overall structure comprises a convolution layer Conv, a batch normalization layer BN, a residual error shrinkage unit RSBUs, a global pooling layer GAP, a full connection layer FC and an activation function ReLU, the residual error shrinkage unit is divided into a shared threshold type among channels and different threshold types per channel, and comprises one RSBU-CW and 4 RSBU-CS, wherein the RSBU-CW is used for acquiring characteristic threshold parameters in different channels;

(3-2) wherein the residual contracting unit fuses the residual unit, the soft threshold function and SENE, and is realized by adding SENEt into the residual unit and taking the soft threshold function as a weighting function;

for the RSBU-CS type residual shrinkage unit, the set threshold is actually the product of the average value of the absolute values of the calculated characteristic diagram and a coefficient alpha, the value range of alpha is [0,1], the threshold calculated by the module is a positive number, the output can not be fully 0, and the calculation formula expression of the RSBU-CS threshold tau is as follows:

in the formula, i, j, c represent three axes of the characteristic diagram, x _i,j,c Representing a characteristic diagram, and averaging represents averaging operation;

for RSBU-CW type residuesA difference contraction unit for setting the threshold value to calculate the absolute value of the feature map and the vector alpha _c The threshold value calculated by this module is thus a vector representing the corresponding contraction threshold value for each channel of the signature, RSBU-CW threshold value τ _c The formula expression is calculated as:

in the formula, i, j, c represent three axes of the characteristic diagram, x _i,j,c Representing a characteristic diagram, and averaging represents averaging operation;

and 4, step 4: the method comprises the steps of model evaluation indexes, loading a model, verifying on a verification set, and comprehensively evaluating a network model by using various model evaluation indexes, wherein the method comprises the following steps;

in order to comprehensively evaluate the application condition of the model in the FOCS fault diagnosis, the detection capability of the fault diagnosis model is judged by using the accuracy, the confusion matrix and the F1 index, and whether the network is converged is judged by using the Loss curve trend;

firstly, introducing a confusion matrix concept, wherein the confusion matrix is a square matrix and is used for counting and displaying the number of real TP, true inverse TN, false positive FP and false inverse FN in a prediction result of a classifier;

meanwhile, the three parameters in the confusion matrix are accuracy, precision and F1;

wherein, TP is a true case, which indicates that the true value is positive and the predicted value is positive; FP represents a false positive case, representing that the actual value is negative but the predicted value is positive; FN represents false negative example, which means true value is positive but predicted value is negative; TN represents a true-inverse example, representing that the true value is negative but the predicted value is negative;

the accuracy rate refers to the probability of a sample being predicted correctly in all the calculation samples, and the calculation formula is as follows:

Accuracy＝(TP+TN)/(TP+FP+TN+FN)

calculating an F1 index, namely calculating a precision rate and a recall rate, wherein the precision rate and the recall rate are a pair of contradictory indexes, and when the precision rate result is high, the recall rate result is low, and vice versa;

the F1 index is a weighted harmonic mean of the two, assuming that both are equally important. The calculation formulas of the three indexes are respectively as follows:

Precision＝TP/(TP+FP)

Recall＝TP/(TP+FN)

F1＝(2×Precision×Recall)/(Precision+Recall)

the maximum values of the three indexes are all 100 percent, and the minimum value is 0 percent. The closer the three index values are to 1, the better the model effect is.

As a further improvement of the invention, the normal output signal x (t) of the FOCS fault diagnosis model is calculated according to the following formula:

wherein t is a sampling time, x (t) is a measured value at the time t, k is an actual value amplitude, ω is an angular frequency, and the stuck-at fault is determined by the accuracy of the measurement equipment, defined as a stuck-at fault, and has a mathematical model of:

in the formula: b represents a coefficient of constant deviation;

the variable ratio deviation fault is expressed as a proportional relation between the amplitude of an output signal of the FOCS and a normal signal, and because the output signal of the FOCS fault diagnosis model is in a proportional relation with the amplitude of a modulation signal, when a modulation loop has a fault to change the amplitude, the FOCS fault diagnosis model can have a variable ratio change, which is called as a variable ratio deviation fault, and the mathematical model of the variable ratio deviation fault is as follows:

in the formula: k is a radical of ₁ Representing a variation ratio deviation coefficient;

the drift deviation fault signal is represented by a gradually increasing drift variation quantity, defined as a drift deviation fault and belongs to a gradual change fault, and the mathematical model of the drift deviation fault signal is as follows:

in the formula: k is a radical of ₂ Denotes the drift constant, k ₂ t represents the drift bias signal.

Compared with the prior art, the invention has the beneficial effects that:

has the beneficial effects that: and the accuracy and the effectiveness of the fault diagnosis model are verified by utilizing the fault data set constructed by the output signals. And constructing an optical fiber current sensor fault diagnosis model based on the improved depth residual shrinkage network by using the constructed fault data set. And training is carried out by using the fault data set, and the overall recognition rate of the fault data set reaches 98.24%. The method is suitable for the field of fault diagnosis of the optical fiber current sensor, the fault diagnosis function of the optical fiber current sensor is completed, and the accuracy of fault diagnosis is improved.

Drawings

FIG. 1 is a flow chart of the disclosed method;

FIG. 2 is a diagram of a modified depth residual shrinkage network model in the method disclosed by the present invention;

FIG. 3 is a model test result confusion matrix;

fig. 4 is a graph of improved depth residual shrinkage network performance indicators.

Detailed Description

The invention is described in further detail below with reference to the following detailed description and accompanying drawings:

the invention discloses an FOCS fault diagnosis method based on an improved residual shrinkage network, wherein a flow chart of the disclosed method is shown in figure 1, a structure chart of an improved deep residual shrinkage network model is shown in figure 2, and the method comprises the following steps:

the FOCS fault diagnosis method based on the improved residual shrinkage network comprises the following steps:

step 1: respectively taking output signals of FOCS normal state, drift deviation fault, transformation ratio deviation fault and fixed deviation fault to construct a fault data set;

the FOCS normal output signal x (t) is calculated according to the following formula:

wherein t is the sampling time, x (t) is the measured value at the time t, k is the amplitude of the true value, ω is the angular frequency, and the stuck-at fault is represented by the fact that the difference between the measured value and the true value is a constant, which is determined by the accuracy of the measuring equipment. The fixed deviation fault is defined, and the mathematical model of the fault is as follows:

in the formula: b represents a fixed deviation coefficient.

The variable ratio deviation fault is represented by a proportional relation between the amplitude of an output signal of the FOCS and a normal signal, and due to the fact that the amplitude of the output signal of the FOCS is proportional to the amplitude of a modulation signal, when a modulation loop fails and the amplitude changes, the FOCS can be caused to have variable ratio change, namely the variable ratio deviation fault, and the mathematical model of the variable ratio deviation fault is as follows:

in the formula: k is a radical of ₁ The ratio deviation coefficient is represented.

The drift deviation fault signal is represented by a gradually increasing drift variation quantity, defined as a drift deviation fault and belongs to a gradual change fault, and the mathematical model of the drift deviation fault signal is as follows:

in the formula: k is a radical of ₂ Denotes the drift constant, k ₂ t represents the drift bias signal.

Step 2: dividing the data set: and dividing the fault data set into a training set, a verification set and a test set, wherein the test set is only used as model evaluation and does not participate in training.

Data set partitioning, comprising the steps of:

(2-1) the sampling frequency of experimental data is 4kHz, the sampling time of each signal is 5s, namely 20000 data sampling points are formed in each piece of data. And respectively collecting a drift fault signal, a transformation ratio fault signal, a fixed deviation fault signal and a normal output signal, and totaling 4 types of experimental data. And dividing the fault data according to the four classifications and marking corresponding labels. Wherein 0 represents drift deviation fault, 1 represents ratio deviation fault, 2 represents fixed deviation fault, and 3 represents normal signal;

and step 3: the building, training and optimizing processes of the fault diagnosis model are as follows: and (3) constructing an improved depth residual shrinkage network model, setting network hyper-parameters and configuring network model structure parameters, and continuously optimizing the model to ensure that the model achieves convergence.

The improved depth residual error shrinkage network comprises the following steps:

(3-1) improving the FOCS fault diagnosis model of the deep residual error shrinkage network, wherein the overall structure comprises a convolution layer Conv, a batch normalization layer BN, a residual error shrinkage unit RSBUs, a global pooling layer GAP, a full connection layer FC and an activation function ReLU. The Residual narrowing Unit is divided into an inter-Channel sharing threshold type (RSBU-CS) and a Channel-by-Channel different threshold type (RSUB-CW). Contains one RSBU-CW and 4 RSBU-CS. The RSBU-CW has the function of acquiring characteristic threshold parameters in different channels, and the RSBU-CS only has one group of thresholds, so that the number of parameters required by the network to be trained is reduced, and the training burden is greatly reduced. By combining the two residual error contraction units, a better feature extraction effect is obtained, and training parameters are reduced.

(3-2) wherein the residual contracting unit fuses the residual unit, the soft threshold function and the SEnet. The main realization is that the method is realized by adding SEnet into a residual unit and taking a soft threshold function as a weighting function.

For the RSBU-CS type residual shrinkage unit, the set threshold is actually the product of the average value of the absolute values of the calculated characteristic diagram and a coefficient alpha, the value range of alpha is [0,1], the threshold calculated by the module is a positive number, and the output can not be fully 0. The RSBU-CS threshold value calculation formula expression is as follows:

for the RSBU-CW type residual shrinking unit, the threshold is set to calculate the absolute value of the characteristic diagram and the vector alpha _c The threshold value calculated by this module is thus a vector representing the shrinkage threshold value for each channel of the feature map. The RSBU-CW threshold value calculation formula expression is as follows:

and 4, step 4: and (5) evaluating indexes of the model. Loading the model and verifying on the verification set, and performing comprehensive evaluation on the network model by using various model evaluation indexes;

and (5) evaluating indexes of the model.

(4-1) in order to comprehensively evaluate the application condition of the model in FOCS fault diagnosis, the Accuracy (Accuracy), the confusion matrix and the F1 index are used for evaluating the detection capability of the fault diagnosis model, and the Loss curve trend is used for judging whether the network is converged.

The confusion matrix concept is introduced first. The confusion matrix is an array, and is mainly used for counting and displaying the number of True (TP), True Negative (TN), False Positive (FP), and False Negative (FN) in the prediction result of the classifier. Meanwhile, the four parameters in the confusion matrix are the components of indexes such as accuracy, precision, F1 and the like.

Wherein, TP is a true case, which indicates that the true value is positive and the predicted value is positive; FP represents a false positive case, representing that the actual value is negative but the predicted value is positive; FN represents false negative example, which means that the true value is positive but the predicted value is negative; TN represents a true negative example, indicating that the true value is negative but the predicted value is negative.

The accuracy is the probability of being predicted to be correct in all the calculation samples, and the calculation formula is as follows:

Accuracy＝(TP+TN)/(TP+FP+TN+FN)

the F1 index is calculated by first calculating Precision (Precision) and Recall (Recall). The precision rate and the recall rate are a pair of contradictory indexes, and when the precision rate result is high, the recall rate result is low, and vice versa. The F1 index is a weighted harmonic mean of the two, assuming that both are equally important. The calculation formulas of the three indexes are respectively as follows:

Precision＝TP/(TP+FP)

Recall＝TP/(TP+FN)

F1＝(2×Precision×Recall)/(Precision+Recall)

the maximum values of the three indexes are all 100 percent, and the minimum value is 0 percent. The closer the three index values are to 1, the better the model effect is.

Fig. 3 shows a confusion matrix of the diagnosis results of the model in the test set, wherein 0 represents a drift deviation fault, 1 represents a ratio deviation fault, 2 represents a fixed deviation fault, and 3 represents a normal signal. From the results of the confusion matrix, the overall fault diagnosis accuracy was 98.24%. The normal signal and the drift fault signal are both 100% in diagnosis accuracy, and the fixed deviation fault classification is the most in error. For drift bias faults, 2 pieces of data are mistaken for a ratio-change bias fault, and 3 pieces of data are mistaken for a fixed bias fault. For a ratio-shift fault, 1 piece of data is misclassified as a stuck-at fault.

FIG. 4 shows the precision, recall, and F1 indices of the test set data, respectively. From the accuracy index, the accuracy rate of the model to the drift deviation fault and the normal signal is 100%, and the identification result to the fixed deviation fault is the worst and is 95.51%. From the recall rate index, the recall rate of the model for the variable ratio fixed deviation fault and the normal signal is 100%, and the recognition result for the drift deviation fault is the worst and is 94.14%. From the index F1, the value of the normal signal reaches 100%, and the drift deviation fault identification result is the worst, and is 96.97%. The result proves that the designed model integrally achieves better effect in the fault diagnosis process of the optical fiber current sensor.

The above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention in any way, but any modifications or equivalent variations made according to the technical spirit of the present invention are within the scope of the present invention as claimed.

Claims (2)

1. The fault diagnosis method of the FOCS fault diagnosis model of the improved residual shrinkage network is characterized by comprising the following specific steps of:

step 1: respectively taking output signals of FOCS normal state, drift deviation fault, transformation ratio deviation fault and fixed deviation fault to construct a fault data set;

step 2: dividing the data set: dividing a fault data set into a training set, a verification set and a test set, wherein the test set is only used as model evaluation and does not participate in training, and the method is as follows;

the experimental data sampling frequency is 4kHz, the sampling time of each section of signal is 5s, namely 20000 data sampling points of each piece of data are respectively collected, and drift fault signals, transformation ratio fault signals, fixed deviation fault signals and normal output signals are collected, so that 4 types of experimental data are counted;

dividing fault data according to four categories and marking corresponding labels, wherein 0 represents drift deviation fault, 1 represents transformation ratio deviation fault, 2 represents fixed deviation fault, and 3 represents normal signal;

and step 3: the building, training and optimizing processes of the fault diagnosis model are as follows: constructing an improved depth residual shrinkage network model, setting network hyper-parameters and configuring network model structure parameters, and continuously optimizing the model to ensure that the model converges;

the improved depth residual error shrinkage network comprises the following steps:

(3-1) improving a FOCS fault diagnosis model of a deep residual error shrinkage network, wherein the overall structure comprises a convolution layer Conv, a batch normalization layer BN, a residual error shrinkage unit RSBUs, a global pooling layer GAP, a full connection layer FC and an activation function ReLU, the residual error shrinkage unit is divided into a shared threshold type among channels and different threshold types per channel, and comprises one RSBU-CW and 4 RSBU-CS, wherein the RSBU-CW is used for acquiring characteristic threshold parameters in different channels;

(3-2) wherein the residual contracting unit fuses the residual unit, the soft threshold function and SENE, and is realized by adding SENEt into the residual unit and taking the soft threshold function as a weighting function;

for the RSBU-CS type residual contraction unit, the set threshold value is actually the product of the average value of the absolute value of the calculated characteristic diagram and a coefficient alpha, the value range of the alpha is [0,1], the threshold value obtained by the calculation of the module is a positive number, the output can not be totally 0, and the calculation formula expression of the RSBU-CS threshold value tau is as follows:

in the formula, i, j, c represent three axes of the characteristic diagram, x _i,j,c Representing a characteristic diagram, and averaging represents averaging operation;

for the residual shrinking unit of RSBU-CW type, the set threshold is actually the absolute value of the calculated characteristic diagram and the vector alpha _c The threshold value calculated by this module is thus a vector representing the corresponding contraction threshold value for each channel of the signature, RSBU-CW threshold value τ _c The formula expression is calculated as:

in the formula, i, j, c represent three axes of the characteristic diagram, x _i,j,c Representing a characteristic diagram, and averaging represents averaging operation;

and 4, step 4: the method comprises the steps of model evaluation indexes, loading a model, verifying on a verification set, and comprehensively evaluating a network model by using various model evaluation indexes, wherein the method comprises the following steps of;

in order to comprehensively evaluate the application condition of the model in the FOCS fault diagnosis, the detection capability of the fault diagnosis model is judged by using the accuracy, the confusion matrix and the F1 index, and whether the network is converged is judged by using the Loss curve trend; firstly, introducing a confusion matrix concept, wherein the confusion matrix is a square matrix and is used for counting and displaying the number of real TP, true inverse TN, false positive FP and false inverse FN in a prediction result of a classifier;

meanwhile, the three parameters in the confusion matrix are accuracy, precision and F1;

wherein TP is a true case, which indicates that the true value is positive and the predicted value is positive; FP represents a false positive case, which represents that the actual value is negative but the predicted value is positive; FN represents false negative example, which means that the true value is positive but the predicted value is negative; TN represents a true-negative example, representing that the true value is negative but the predicted value is negative;

the accuracy is the probability of being predicted to be correct in all the calculation samples, and the calculation formula is as follows:

Accuracy＝(TP+TN)/(TP+FP+TN+FN)

calculating an F1 index, wherein precision rate and recall rate are required to be calculated firstly, the precision rate and the recall rate are a pair of contradictory indexes, when the precision rate result is high, the recall rate result is low, and vice versa;

the F1 index is a weighted harmonic mean of the two indexes, which is based on the assumption that the two indexes are equally important, and the calculation formulas of the three indexes are:

Precision＝TP/(TP+FP)

Recall＝TP/(TP+FN)

F1＝(2×Precision×Recall)/(Precision+Recall)

the maximum values of the three indexes are all 100 percent, and the minimum value is 0 percent. The closer the three index values are to 1, the better the model effect is.

2. The FOCS fault diagnosis method based on the improved residual shrinkage network as claimed in claim 1, wherein:

the normal output signal x (t) of the FOCS fault diagnosis model is calculated according to the following formula:

where t is the sampling time, x (t) is the measured value at time t, k is the true amplitude, ω is the angular frequency,

the stuck-at fault is represented by a measured value and a true value which are different by a certain constant, is determined by the accuracy of the measuring equipment, and is defined as the stuck-at fault, and the mathematical model of the stuck-at fault is as follows:

in the formula: b represents a coefficient of constant deviation;

the variable ratio deviation fault is represented by a proportional relation between the amplitude of an output signal of the FOCS and a normal signal, and because the amplitude of the output signal of the FOCS fault diagnosis model is proportional to the amplitude of a modulation signal, when the modulation loop has a fault to change the amplitude, the FOCS fault diagnosis model has a variable ratio change, which is called as a variable ratio deviation fault, and the mathematical model of the fault is as follows:

in the formula: k is a radical of ₁ Representing a variation ratio deviation coefficient;

the drift deviation fault signal is represented by a gradually increasing drift variation quantity, defined as a drift deviation fault and belongs to a gradual change fault, and the mathematical model of the drift deviation fault signal is as follows:

in the formula: k is a radical of ₂ Denotes the drift constant, k ₂ t represents the drift bias signal.

Images