CN117499199A - VAE-based information enhanced decoupling network fault diagnosis method and system - Google Patents

Abstract

The invention belongs to the field of industrial fault diagnosis technology, and in particular relates to a VAE-based information enhanced decoupling network fault diagnosis method and system, which uses statistics to convert residual signals into scalars, and calculates the threshold of normal sample statistics. The online sample statistical value is compared with the threshold to achieve fault detection; a new pre-mapping operation is performed on the VAE input and a new post-mapping operation is performed on the output to achieve decoupling of network input and output and help find the specific source of system fault signals. location to achieve isolation of fault signals. This invention fixes the trained IEDN-VAE parameters, calculates the gradient of the residual signal statistics and the input fault sample x ^f , and updates x ^f through backpropagation until the sample returns to the normal domain, which is recorded as is the prototype of the fault sample in the normal domain, by comparing x ^f with Calculate an estimate of the failure.

Description

A VAE-based information-enhanced decoupled network fault diagnosis method and system

Technical Field

The present invention belongs to the technical field of industrial fault diagnosis, and in particular, relates to a VAE-based information enhanced decoupled network fault diagnosis method and system.

Background Art

Modern industrial systems are becoming increasingly large and complex. Once a system fails, it will pose a serious threat to the economic benefits of industrial production and the life safety of operators. In addition, if minor faults in the system cannot be detected and eliminated in time, they will gradually accumulate and spread to other parts along the system flow and evolve into major system failures. Therefore, accurate, timely and efficient fault diagnosis is particularly important to improve the safety and stability of modern industrial production processes.

Fault diagnosis generally includes three subtasks: fault detection, fault isolation, and fault estimation. Fault detection determines whether a fault has occurred in the system. Fault isolation can quickly and effectively locate the cause of the fault, greatly reducing the fault recovery time. Fault estimation can further determine the type of fault and the size of the fault, and reconstruct the fault signal in the observed data. In general, the relationship between fault detection, fault isolation, and fault estimation is progressive and closely related. Fault detection is the first step in fault diagnosis, and fault isolation and fault estimation are subsequent extended tasks.

In fault detection tasks, residual generators have been widely used. Residual signal generation methods can be divided into model-based methods, shallow learning-based methods, and deep learning-based methods. Model-based methods require the establishment of an accurate mathematical model or physical model of the system under study in advance, but this requirement is usually difficult to meet in today's large and complex industrial processes. At the same time, a lot of expert experience needs to be applied when designing fault detection schemes. Therefore, this method has insurmountable defects in practical applications and is greatly constrained; shallow learning-based methods extract compressed features by projecting data into a specific low-dimensional latent variable space, and the extracted features have clear mathematical and physical meanings. However, it is still difficult to process large-scale data and fit complex system behaviors; deep learning-based residual generation methods have complex multi-layer nonlinear structures, which can fit sufficiently complex process system models, and have the advantages of extracting complex nonlinear dynamic features and freely identifying system parameters. Variational autoencoders are one of the typical representatives and have attracted much attention in the field of industrial fault diagnosis.

After the system fault is detected, it is necessary to further implement fault isolation. Existing fault isolation methods are mainly based on two ideas: model decoupling and contribution evaluation. Model-based fault isolation methods usually construct a decoupled transfer matrix to structurally separate the effects of different types of faults on the residual signal; contribution evaluation-based methods achieve fault isolation by analyzing the contribution of each input quantity to the generated statistics. Although some fault isolation methods have been proposed, there is a lack of fault decoupling structures based on deep networks. Due to the complex multi-layer nonlinear structure of deep networks, and their opaque and difficult to interpret characteristics, there are huge challenges in implementing fault isolation tasks. Although attribution methods based on explainable artificial intelligence have achieved some success in fault isolation tasks, they are easily affected by the "smearing effect" and cannot analyze the contribution of residuals to test statistics. Therefore, applying the decoupling idea to the fault isolation method of deep learning can solve the above dilemma.

Fault estimation is the ultimate goal of fault diagnosis. After fault isolation, the variables will not interfere with each other's fault signal estimation, which can provide the correct direction, help reduce reconstruction noise, and improve the accuracy of fault estimation. Currently, the most common method is still model-based fault estimation method (such as sliding mode observer). This type of method has little research on fault estimators with decoupling capabilities, resulting in difficulty in ensuring the accuracy of fault estimation. In addition, deep network-based methods for fault estimation of complex nonlinear systems are still very scarce.

Through the above analysis, the problems and defects of the existing technology are as follows: In terms of fault detection, model-based fault detection methods are usually difficult to establish accurate mathematical and physical models in today's large and complex industrial processes, and cannot meet actual needs; fault detection methods based on shallow learning have weak ability to process large-scale data and are difficult to fit complex system behaviors. In terms of fault isolation, there is a lack of fault decoupling structure based on deep networks, and the existing fault isolation methods are easily affected by the "smearing effect" and cannot analyze the contribution of residuals to test statistics. In terms of fault estimation, there is little research on fault estimation with decoupling capabilities, which makes it difficult to ensure the accuracy of fault estimation.

Summary of the invention

In view of the problems existing in the prior art, the present invention provides a VAE-based information enhanced decoupled network fault diagnosis method and system.

The present invention is implemented as follows: a VAE-based information enhancement decoupling network fault diagnosis method, the VAE-based information enhancement decoupling network fault diagnosis method comprising:

S1: Input the normal state data x into IEDN-VAE and get the output Realize decoupling and reconstruction of input;

S2: Construct a residual generator r related to the input and output observations;

S3: convert the residual signal into a scalar signal using T ² statistics, and determine the threshold J _th of normal samples;

S4: Calculate the ^T2 test statistic of the online sample residual signal, determine the status of the online sample, and complete the fault detection task;

S5: The residual signal is further used to locate the cause variable of the fault in the input sample to complete the fault isolation task;

S6: Based on the results of fault isolation, a fault estimator based on gradient descent strategy is constructed to complete the fault estimation task.

Furthermore, the information enhancement decoupling network in S1 includes:

S101, input data: the input is recorded as x, x = (x ₁ , ..., x _m ); where m represents the number of sensors collecting input data; the input data of the offline process are all in a normal state, that is,

in, represents the entire observation space, and They represent the normal domain space and fault domain space with dimension m respectively;

Diagonalize x:

Diag{x}＝[χ ₁ ,…,χ _m ] ^T =X;

Each matrix row vector As a new sample, The kth element of satisfies:

S102, pre-mapping P _pre : X _pre =XΓ=[χ ₁ ,…,χ _m ] ^T Γ. Wherein Γ is a circulant matrix, which can ensure that X _pre is fair to each variable;

The circulant matrix can be determined by its first row. For example, the circulant matrix with the first row [1,2,3,4] can be constructed as follows

Select the Circulation Matrix That is: X _pre =[χ ₁ ,…,χ _m ] ^T Γ _pre ;

Among them, 1<n<m, if n is too small, it will affect the ability of _Xpre to inherit the effective information of input X, and if n is too large, it will have a negative impact on the generalization ability of the model;

S103, VAE reconstruction: The VAE network is a typical unsupervised neural network, which consists of two parts: the encoder and the decoder, both of which are composed of fully connected layers; the relationship between any two adjacent layers (l-1, l) can be written as:

x ^(l) =σ ^(l) (w ^(l-1,l) x ^(l-1) +b ^(l) );

Where x ^(l-1) and x ^(l) represent the values of the l-1th layer and the lth layer respectively, w represents the weight coefficient, b represents the bias, and σ represents the activation function;

X _pre is used as the input of VAE, and the output is recorded as

represents the mapping function of the VAE network, and θ _VAE represents the network parameters;

S104, post mapping P _post :

and Γ _ik represent The elements of the i-th row and k-th column of Γ are concatenated with the function mapping defined above. The entire IDN is described as follows:

in, That is, the decoupled reconstruction of the input data x;

S105, the process of training IEDN-VAE includes two parts: internal loss l _in and external loss l _out . _{l in} represents the reconstruction loss of the internal VAE, and l _out represents the decoupled reconstruction loss of IEDN-VAE:

After sufficient training with normal data samples, l _in and l _out will satisfy:

Furthermore, the construction method of the residual generator in S2 includes:

in, represents the optimal model parameters after IEDN-VAE training;

The role of the residual generator R is to quantify the reconstruction deviation of IEDN-VAE, where r is a deviation vector:

Further, the method for determining the normal sample threshold _Jth in S3 includes:

First, the Hotelling statistic (T ² ) is used to convert the S2 generated residual signal into a scalar:

Where r ⁿ (t) represents the residual signal of the tth normal sample, and N represents the number of training samples in the offline process;

The randomness of the normal sample ^T2 can be measured by the threshold To describe, we use the kernel density estimation method to determine The value of; in the kernel density estimation method, The probability density at is expressed as follows:

Among them, k(·) represents the kernel function, ρ represents the The bandwidth with significant influence is when m _T2 = 1 in ρ. The cumulative density function at is calculated as follows:

The above formula describes Does not exceed a specific boundary (threshold ) is the confidence level, δ represents the confidence level, which is usually taken as 99.5%.

Further, the method for determining the online sample status in S4 includes:

After offline training is completed, the learned The value will be used as a logical decision condition in the online process to determine whether the system is in a normal state. After setting the confidence level, It can be calculated by the inverse function of the above formula;

In the online process, the state of the input sample of IEDN-VAE is unknown, and the test statistic T ² (r) of the residual signal of the unknown state sample needs to be calculated first; after setting the confidence level, the state of the online sample is determined by the following rules:

Furthermore, the method of further applying the residual signal to the fault isolation task in S5 includes:

According to S1, the output of IEDN-VAE It is only related to the input _xi and has nothing to do with other input variables. Therefore, when the residual generator defined by S2 is further applied to the fault isolation task:

in, Indicates the cause variable of the fault, represents the i-th variable of r ^f , express The i-th diagonal element of ; I is an m×m identity matrix.

Further, the method for completing the fault estimation task in S6 includes:

The offline process has trained the fault detection model well and has fully learned the knowledge of normal state samples. Since the task of fault estimation is to reconstruct the fault signal f added to the normal sample ^xn , and ^xf is observable, the normal sample before the fault signal is added (denoted as ) is particularly important for estimating the value of f; therefore, a reasonable reconstruction is designed The loss function is:

The IEDN-VAE parameters obtained by S1 are Fixed, minimized by optimizing the input sample According to the chain rule, when updating the input for the kth time, The gradient with respect to the input is expressed as:

in:

The input sample for the k+1th time is

Input samples are iterated continuously until Stop, the input sample at this time is already in a normal state, and the obtained input sample can be regarded as the normal sample corresponding to the original fault input sample ^xf

Finally, the fault signal f added to the normal sample ^xn can be calculated:

at this time:

Another object of the present invention is to provide a VAE-based information enhanced decoupled network fault diagnosis system using the VAE-based information enhanced decoupled network fault diagnosis method, and the VAE-based information enhanced decoupled network fault diagnosis system includes:

Reconstruction module: used to input normal state data x into IEDN-VAE and obtain output Realize decoupling and reconstruction of input;

Residual generator construction module: used to construct a residual generator r related to the input and output observations;

Threshold determination module: converts the residual signal into a scalar signal using ^T2 statistics and determines the threshold _Jth of normal samples;

Fault detection module: used to calculate the ^T2 test statistic of the online sample residual signal, determine the status of the online sample, and complete the fault detection task;

Fault isolation module: used to further use the residual signal to locate the cause variable of the fault in the input sample and complete the fault isolation task;

Fault estimation module: It is used to build a fault estimator based on the gradient descent strategy according to the results of fault isolation to complete the fault estimation task.

Another object of the present invention is to provide a computer device, the computer device comprising a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the processor executes the steps of the VAE-based information enhanced decoupled network fault diagnosis method.

Another object of the present invention is to provide an information data processing terminal, which is used to implement the VAE-based information enhanced decoupled network fault diagnosis system.

In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

First, the VAE-based fault diagnosis method provided by the present invention is a data-driven intelligent method. An observer is designed to simulate the behavior of the system in a normal state through reconstruction deviation. Once a fault signal appears in the system, the observer will be sensitive to the intolerable reconstruction deviation; and the decoupled design can help us find the specific location where the intolerable deviation behavior occurs in the reconstruction vector, that is, to achieve the isolation of the fault signal; and then by fixing the parameters of the deep network, the gradient descent algorithm is used to iteratively update the input sample to minimize the reconstruction deviation value, and finally find the corresponding value of the tested fault sample in the normal domain, and calculate the estimated value of the fault signal. In summary, the decoupled network based on VAE can distinguish the influence of each fault variable while performing fault detection to achieve fault isolation, and then update the fault input sample to find its corresponding value in the normal domain to achieve fault estimation.

Second, the VAE-based fault diagnosis method provided by the present invention proposes a new front-mapping and back-mapping method for fault isolation tasks, which can help build a fault decoupling structure based on a deep network and better realize the decoupling reconstruction of input data.

The present invention provides a VAE-based fault diagnosis method, which proposes an explainable fault estimation method for fault estimation tasks. The parameters of the trained deep network are fixed, the reconstruction loss and the gradient of the input sample are calculated, and the input is updated by back propagation according to the chain rule. The sample data in the fault domain is pulled back to the normal domain, and finally the initial fault sample and the normal domain sample are compared to calculate the estimated value of the fault.

Third, whether the technical solution of the present invention overcomes technical bias: Fault diagnosis technology monitors the presence or absence of fault signals, locates the root causes and reconstructs the size by detecting, isolating and identifying system faults. Due to its multi-layer nonlinear structure, deep learning has the ability to simulate complex process characteristics, which can greatly improve the diagnostic effect when used for fault diagnosis tasks. However, the multi-layer nonlinear structure of the deep network is generally regarded as a "black box model", which cannot provide corresponding evidence and support for its output or decision-making process. Therefore, people are skeptical about its output results and believe that the results of the deep network lack interpretability. In the present invention, by making full use of the system normal domain knowledge learned by the deep network model, the input fault samples are updated based on the gradient descent algorithm, and the fault domain samples are migrated back to the normal domain. The input found by this method can enable the trained deep network to achieve the best reconstruction performance. The final input can be reasonably considered to be the prototype of the fault sample in the normal domain. It can explain the prototype of specific knowledge extracted by the deep network during the learning process and enhance the understanding of the learning representation of multi-layer nonlinear structures. Therefore, this method can help improve the transparency of deep learning, make the learning representation of deep networks explainable, and help people understand the decision-making behavior and prediction results of deep networks, thereby promoting their safer use.

Fourth, this VAE-based Information Enhanced Decoupled Network (IEDN-VAE) fault diagnosis approach involves multiple steps, each with its own unique technical advancement.

S1: Decoupled reconstruction: By using the IEDN-VAE model, this method is able to achieve decoupled reconstruction of normal state data. This is a significant technical advancement because it enables us to extract valuable information from complex data and better understand the normal operating state of the system.

S2: Residual Generator: This method can accurately identify abnormal behavior of the system by constructing a residual generator related to the input and output observations. This is an important technical advancement because it provides an effective way to monitor the health of the system.

S3: Threshold determination: By using statistics to transform the residual signal into a scalar signal and determining the threshold of normal samples, this is a significant technological advancement because it provides the system with a reliable judgment standard to effectively identify faults.

S4: Fault detection: Calculate the test statistic of the residual signal of the online sample, determine the state of the online sample, and complete the fault detection task. This is an important technological advancement because it allows fault detection to be performed in real time or near real time, thereby discovering and handling faults as early as possible.

S5: Fault Isolation: The residual signal is further used to locate the cause variable of the fault in the input sample to complete the fault isolation task. This is a significant technical advancement because it provides a systematic method to determine the source of the fault and perform targeted repairs.

S6: Fault Estimation: Based on the results of fault isolation, a fault estimator based on gradient descent strategy is constructed to complete the fault estimation task. This is an important technical advancement because it provides an effective way to estimate the impact of faults and thus develop more effective maintenance strategies.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

FIG1 is a flow chart of a fault diagnosis method of an information enhanced decoupled network based on VAE provided by an embodiment of the present invention;

FIG2 is a curve diagram of the detection of additive faults by IEDN provided in an embodiment of the present invention;

FIG3 is a schematic diagram of estimation results of a typical additive fault provided by an embodiment of the present invention.

DETAILED DESCRIPTION

In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

In view of the problems existing in the prior art, the present invention provides a fault diagnosis method and system for an information enhanced decoupled network based on VAE. The present invention is described in detail below with reference to the accompanying drawings.

As shown in FIG1 , the fault diagnosis method of the information enhanced decoupled network based on VAE provided by the embodiment of the present invention includes:

a) Offline process

S1: Input the normal state data x into IEDN-VAE (Information Enhancement Decoupling Network) and get the output Realize decoupling and reconstruction of input.

S2: Construct a residual generator r related to the input and output observations.

S3: Use T ² statistics to convert the residual signal into a scalar signal, and determine the threshold J _th of normal samples.

b) Online process

S4: Calculate the ^T2 test statistic of the online sample residual signal, determine the status of the online sample, and complete the fault detection task.

S5: The residual signal is further used to locate the cause variable of the fault in the input sample to complete the fault isolation task.

S6: Based on the results of fault isolation, a fault estimator based on gradient descent strategy is constructed to complete the fault estimation task.

The information enhancement decoupling network described in S1 includes:

S101, input data: the input is denoted as x, x = (x ₁ ,…, x _m ). Wherein, m represents the number of sensors collecting input data;

It should be pointed out that the input data of the offline process are all in normal state, that is,

In the above formula, represents the entire observation space, and They represent the normal domain space and fault domain space with dimension m respectively;

Diagonalize x:

Diag{x}＝[χ ₁ ,…,χ _m ] ^T =X

Each matrix row vector As a new sample, The kth element of satisfies:

S102, pre-mapping P _pre : X _pre =XΓ=[χ ₁ ,…,χ _m ] ^T Γ. Wherein Γ is a circulant matrix, which can ensure that X _pre is fair to each variable;

The circulant matrix can be determined by its first row. For example, the circulant matrix with the first row [1,2,3,4] can be constructed as follows

The present invention selects the circulant matrix That is: X _pre =[χ ₁ ,…,χ _m ] ^T Γ _pre ;

Among them, 1<n<m, if n is too small, it will affect the ability of _Xpre to inherit the effective information of input X, and if n is too large, it will have a negative impact on the generalization ability of the model;

S103, VAE reconstruction: The VAE network is a typical unsupervised neural network, which consists of two parts: the encoder and the decoder, both of which are composed of fully connected layers. The relationship between any two adjacent layers (l-1, l) can be written as

x ^(l) =σ ^(l) (w ^(l-1,l) x ^(l-1) +b ^(l) )

Where x ^(l-1) and x ^(l) represent the values of the l-1th layer and the lth layer respectively, w represents the weight coefficient, b represents the bias, and σ represents the activation function;

X _pre is used as the input of VAE, and the output is recorded as

represents the mapping function of the VAE network, and θ _VAE represents the network parameters;

S104, post mapping P _post :

and Γ _ik represent and the element in the i-th row and k-th column of Γ. Concatenating the function mapping defined above, the entire IDN is described as follows:

in, That is, the decoupled reconstruction of the input data x;

S105, the process of training IEDN-VAE includes two parts: internal loss l _in and external loss l _out . _{l in} represents the reconstruction loss of the internal VAE, and l _out represents the decoupled reconstruction loss of IEDN-VAE:

After sufficient training with normal data samples, l _in and l _out will satisfy:

The method for constructing the residual generator described in S2 includes:

in, represents the optimal model parameters after IEDN-VAE training;

The role of the residual generator R is to quantify the reconstruction deviation of IEDN-VAE, where r is a deviation vector:

The method for determining the normal sample threshold _Jth described in S3 includes:

First, the Hotelling statistic (T ² ) is used to convert the S2 generated residual signal into a scalar:

Where r ⁿ (t) represents the residual signal of the tth normal sample, and N represents the number of training samples in the offline process;

The randomness of the normal sample ^T2 can be measured by the threshold To describe, the present invention uses the kernel density estimation method to determine The value of; in the kernel density estimation method, The probability density at is expressed as follows:

Among them, k(·) represents the kernel function, ρ represents the The bandwidth with significant influence is when m _T2 = 1 in ρ. The cumulative density function at is calculated as follows:

The above formula describes Does not exceed a specific boundary (threshold ), δ represents the confidence level, which is usually taken as 99.5%.

The method for determining the online sample status described in S4 includes:

After offline training is completed, the learned The value will be used as a logical decision condition in the online process to determine whether the system is in a normal state. After setting the confidence level, It can be calculated by the inverse function of the above formula;

In the online process, the state of the input sample of IEDN-VAE is unknown, and the test statistic T ² (r) of the residual signal of the unknown state sample needs to be calculated first; after setting the confidence level, the state of the online sample is determined by the following rules:

The method of further applying the residual signal to the fault isolation task described in S5 includes:

According to S1, the output of IEDN-VAE It is only related to the input _xi and has nothing to do with other input variables. Therefore, when the residual generator defined by S2 is further applied to the fault isolation task:

in, represents the cause variable of the fault, _ri ^f represents the i-th variable of r ^f , express It should be noted that I is an m×m identity matrix.

The method for completing the fault estimation task described in S6 includes:

The offline process has trained the fault detection model well and has fully learned the knowledge of normal state samples. Since the task of fault estimation is to reconstruct the fault signal f added to the normal sample ^xn , and ^xf is observable, the normal sample before the fault signal is added (denoted as ) is particularly important for estimating the value of f; therefore, a reasonable reconstruction is designed The loss function is:

The IEDN-VAE parameters obtained by S1 are Fixed, minimized by optimizing the input sample According to the chain rule, when updating the input for the kth time, The gradient with respect to the input is expressed as:

in:

The input sample for the k+1th time is

Input samples are iterated continuously until Stop, the input sample at this time is already in a normal state, and the obtained input sample can be regarded as the normal sample corresponding to the original fault input sample ^xf

Finally, the fault signal f added to the normal sample ^xn can be calculated:

at this time:

In order to prove the creativity and technical value of the technical solution of the present invention, this section provides application examples of the claimed technical solution on specific products or related technologies.

An application embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of a VAE-based information enhanced decoupled network fault diagnosis method.

An application embodiment of the present invention provides an information data processing terminal, which is used to implement an information enhancement decoupled network fault diagnosis system based on VAE.

The embodiment of the present invention defines the performance evaluation index of fault detection and estimation:

The false alarm rate (FAR) and missed detection rate (MDR) are suitable for evaluating the fault detection (FD) performance of the observer and are defined as follows:

Where ^xo represents the observation value with unknown status in online testing.

Assume that N _c types of data sets are collected online, each of which represents a type of fault. There are N _i samples in the i-th data set, which consists of several normal data and fault data. Then, the total number of samples collected online can be expressed as It can be obtained that the average FAR and MDR of the i-th category can be expressed as:

Among them, FA, TA, MD, and RA represent false alarm, accurate alarm, missed detection, and correct detection, respectively.

In addition, the performance of fault estimation can be measured by the root mean square error (RMSE) and ARMSE, which can be expressed as:

in, and f represent the estimated fault signal and the actual fault signal added to the normal samples, respectively.

The embodiment of the present invention utilizes a continuous stirred tank reactor (CSTR) for simulation. In chemical processes, CSTR is a very common device. This process is a complex, highly nonlinear chemical reaction process and is widely used in actual industrial production. In the CSTR, the materials in the device can react continuously through the action of the agitator in the device, which promotes the balance of temperature and concentration in the reaction system. The model allows the present invention to simulate the operation process of the CSTR by controlling the feed temperature and concentration, the coolant inlet temperature, and other sensor measurements. Therefore, the manipulated variables of the CSTR system include the feed concentration _Ci and temperature _Ti and the coolant inlet temperature _Tci . The present invention can express the manipulated variable u and the response variable y as:

u＝[C _i T _i T _ci ] ^T

Wherein, the superscript s represents the variable measured by the sensor, C and T represent the concentration and temperature of the reactant, respectively; _Tc and _Qc represent the temperature and flow rate of the coolant, respectively.

In the CSTR model, four different types of additive faults are introduced, and the fourth type adds multiple fault signals. The details are shown in Table 1 below, where the subscript "0" represents the variable value before adding the fault signal:

Table 1 Introduction to faults in CSTR simulation

In the process of collecting samples, CSTR was simulated ten times in normal state, collecting 1201 samples each time, and CSTR was run 4 times additionally to collect the fault data set mentioned in Table 1 for online testing. The fault was introduced after the 200th sampling interval in the fault set. In this way, the present invention collected 12010 and 4804 samples for training and testing respectively, and the sample composition is shown in Table 2:

Table 2 Introduction to faults in CSTR simulation

Each simulation of the CSTR system lasts for 20 hours, and samples are collected every minute. The system runs 10 times in total. According to the model of the present invention, the hyperparameters are selected as follows: Nepoch = 20 iterations; batch size Nb = 16; dropout rate Dropout = 0.2; learning rate η = 0.0001; reconstruction loss factor λ = 10; allowable error ε = 5×10 ^-3 ; confidence 1-δ = 99.5%; the expected FAR is

Experimental comparison: The present invention constructs a residual-based deep autoencoder (DAE) and IEDN for comparison. The parameter settings of DAE are the same as those of IEDN.

Table 3 lists the mean and standard deviation of AFAR and AMDR after 10 independent repeated tests, where AFAR should not be higher than 0.5% and AMDR should be as small as possible. Although DAE has a lower AMDR, the AFAR exceeds 0.5%, which does not meet the expected value. Compared with the DAE model, the proposed IEDN can achieve the best FD performance, and the mean table of FD results is shown in Table 4 below:

Table 3 Mean (± standard deviation) of FD results in ten independent repeated experiments

Accordingly, Table 4 below shows the fault detection results of IEDN and DAE for four different types of faults. It can be seen that both models show high FD performance for additive faults.

Table 4 Mean (± standard deviation) of FD results in ten independent repeated experiments

Accordingly, FIG2 shows the detection curve of IEDN for additive faults; in FIG2, the present invention analyzes a multivariate fault (fault 4). The straight line represents samples marked as normal, the dotted line starting from sample 201 represents samples marked as faulty, and the long black dotted line represents the learning threshold, the upper limit range of which represents the fault state of the sample predicted by IEDN, and the lower limit range represents the normal area of the predicted sample. It is not difficult to see from FIG1 that the predicted sample results can well meet the fault detection index.

In addition to meeting the fault detection task, IEDN can also be used for fault estimation tasks. Figure 3 intuitively shows the estimation results of a typical additive fault. In Figure 3, the present invention analyzes a multivariate fault (fault 4), so there are two actual fault signals. The real fault signal is represented by a straight line, and the corresponding predicted fault signal is represented by a dotted line. It can be clearly seen from the results of Figure 3 that the predicted fault signal can track the real fault signal well, and the prediction result can meet the fault estimation index.

Embodiment 1: Fault diagnosis of wind turbine generator set

In wind turbines, the IEDN-VAE fault diagnosis method can be applied to monitor and diagnose the operating status of wind turbines. The specific implementation scheme is as follows:

S1: Collect data on the normal operating status of the wind turbine, including wind speed, power generation, temperature, etc., and input them into the IEDN-VAE model for decoupling and reconstruction.

S2: Construct a residual generator to generate residuals of input and output observations.

S3: Convert the residual signal into a scalar signal and determine the threshold of normal samples through statistical analysis.

S4: During the operation of the fan, data is collected in real time, and the residual signal test statistics of the online samples are calculated to determine the operating status of the fan and realize fault detection.

S5: If a fault is detected, further analyze the residual signal to locate the cause variable of the fault, such as abnormal wind speed, excessive temperature, etc., to achieve fault isolation.

S6: Based on the results of fault isolation, the severity of the fault is estimated through the gradient descent strategy, and the fault estimation is completed to provide a basis for maintenance decisions.

Example 2: Fault diagnosis of industrial production line

In industrial production lines, the IEDN-VAE fault diagnosis method can be used to monitor and diagnose the operating status of equipment. The specific implementation scheme is as follows:

S1: Collect data on the normal operating status of the equipment, such as working speed, temperature, pressure, etc., and input them into the IEDN-VAE model for decoupling and reconstruction.

S2: Build a residual generator associated with the input and output observations.

S3: Convert the residual signal into a scalar signal and determine the threshold of normal samples through statistical analysis.

S4: During the operation of the equipment, data is collected in real time, and the residual signal test statistics of the online samples are calculated to determine the operating status of the equipment and realize fault detection.

S5: If a fault is detected, further analyze the residual signal to locate the cause variable of the fault, such as equipment overheating, abnormal pressure, etc., to achieve fault isolation.

S6: Based on the results of fault isolation, the impact of the fault is estimated through the gradient descent strategy, and the fault estimation is completed to provide a basis for subsequent maintenance or replacement decisions.

It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and/or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims (10)

1. A VAE-based information enhanced decoupled network fault diagnosis method, characterized by comprising: S1: Input the normal state data x into IEDN-VAE and get the output Realize decoupling and reconstruction of input; S2: Construct a residual generator r related to the input and output observations; S3: convert the residual signal into a scalar signal using T ² statistics, and determine the threshold J _th of normal samples; S4: Calculate the ^T2 test statistic of the online sample residual signal, determine the state of the online sample, and complete the fault detection task; S5: The residual signal is further used to locate the cause variable of the fault in the input sample to complete the fault isolation task; S6: Based on the results of fault isolation, a fault estimator based on gradient descent strategy is constructed to complete the fault estimation task. 2. The VAE-based information enhanced decoupling network fault diagnosis method according to claim 1, wherein the information enhanced decoupling network in S1 comprises: S101, input data: the input is recorded as x, x = (x ₁ , ..., x _m ); where m represents the number of sensors collecting input data; the input data of the offline process are all in a normal state, that is, in, represents the entire observation space, and They represent the normal domain space and fault domain space with dimension m respectively; Diagonalize X: X＝Diag{x}＝[χ ₁ ,…,χ _m ] ^T ; Each matrix row vector As a new sample, i＝1,2,…,m， The kth element of satisfies: S102, pre-mapping P _pre : X _pre =XΓ=[χ ₁ ,…,χ _m ] ^T Γ, where Γ is a circulant matrix, which can ensure that X _pre is fair to each variable; The circulant matrix can be determined by its first row. For example, the circulant matrix with the first row [1,2,3,4] can be constructed as follows Select the Circulation Matrix That is: X _pre = [χ ₁ ,…,χ _m ] ^T Γ _pre ; where 1<n<m, a value of n that is too small will affect X _pre's ability to inherit the valid information of input X, and a value of n that is too large will have a negative impact on the generalization ability of the model; S103, VAE reconstruction: The VAE network is a typical unsupervised neural network, which consists of two parts: the encoder and the decoder, both of which are composed of fully connected layers; the relationship between any two adjacent layers (l-1, l) can be written as: x ^(l) =σ ^(l) (w ^(l-1,l) x ^(l-1) +b ^(l) ); Where x ^(l-1) and x ^(l) represent the values of the l-1th layer and the lth layer respectively, w represents the weight coefficient, b represents the bias, and σ represents the activation function; X _pre is used as the input of VAE, and the output is recorded as F _VAE (·) represents the mapping function of the VAE network, and θ _VAE represents the network parameters; S104, post mapping P _post : and Γ _ik represent The elements of the i-th row and k-th column of Γ are concatenated with the function mapping defined above. The entire IDN is described as follows: in, That is, the decoupled reconstruction of the input data x; S105, the process of training IEDN-VAE includes two parts: internal loss l _in and external loss l _out . l _in represents the reconstruction loss of the internal VAE, and l _out represents the decoupled reconstruction loss of IEDN-VAE: After sufficient training with normal data samples, l _in and l _out will satisfy: 3. The VAE-based information enhanced decoupled network fault diagnosis method according to claim 1, characterized in that the construction method of the residual generator in S2 comprises: in, represents the optimal model parameters after IEDN-VAE training; The role of the residual generator R is to quantify the reconstruction deviation of IEDN-VAE, where r is a deviation vector: 4. The VAE-based information enhanced decoupled network fault diagnosis method according to claim 1, wherein the method for determining the normal sample threshold _Jth in S3 comprises: First, the Hotelling statistic (T ² ) is used to convert the S2 generated residual signal into a scalar: Where r ⁿ (t) represents the residual signal of the t-th normal sample, and N represents the number of training samples in the offline process; The randomness of the normal sample ^T2 can be measured by the threshold To describe, we use the kernel density estimation method to determine The value of; in the kernel density estimation method, The probability density at is expressed as follows: Among them, k(·) represents the kernel function, ρ represents the The bandwidth that has a significant impact, ρ The cumulative density function at is calculated as follows: The above formula describes Does not exceed a specific boundary (threshold ) is the confidence level, δ represents the confidence level, which is usually taken as 99.5%. 5. The VAE-based information enhanced decoupled network fault diagnosis method according to claim 1, wherein the method for determining the online sample status in S4 comprises: After offline training is completed, the learned The value will be used as a logical decision condition in the online process to determine whether the system is in a normal state. After setting the confidence level, It can be calculated by the inverse function of the above formula; In the online process, the state of the input sample of IEDN-VAE is unknown, and the test statistic of the residual signal of the unknown state sample needs to be calculated first. After setting the confidence level, the status of the online sample is determined by the following rules: 6. The VAE-based information enhanced decoupled network fault diagnosis method according to claim 1, wherein the method of further applying the residual signal to the fault isolation task in S5 comprises: According to S1, the output of IEDN-VAE It is only related to the input _xi and has nothing to do with other input variables. Therefore, when the residual generator defined by S2 is further applied to the fault isolation task: in, Indicates the cause variable of the fault, represents the i-th variable of r ^f , express The i-th diagonal element of ; I is an m×m identity matrix. 7. The VAE-based information enhanced decoupled network fault diagnosis method according to claim 1, wherein the method for completing the fault estimation task in S6 comprises: The offline process has trained the fault detection model well and has fully learned the knowledge of normal state samples. Since the task of fault estimation is to reconstruct the fault signal f added to the normal sample ^xn , and ^xf is observable, the normal sample before the fault signal is added (denoted as ) is particularly important for estimating the value of f; therefore, a reasonable reconstruction is designed The loss function is: The IEDN-VAE parameters obtained by S1 are Fixed, minimized by optimizing the input sample According to the chain rule, when updating the input for the kth time, The gradient with respect to the input is expressed as: in: The input sample for the k+1th time is Input samples are iterated continuously until Stop, the input sample at this time is already in a normal state, and the obtained input sample can be regarded as the normal sample corresponding to the original fault input sample ^xf Finally, the fault signal f added to the normal sample ^xn can be calculated: at this time: 8. A VAE-based information enhanced decoupled network fault diagnosis system using the VAE-based information enhanced decoupled network fault diagnosis method according to any one of claims 1 to 7, characterized in that the VAE-based information enhanced decoupled network fault diagnosis system comprises: Reconstruction module: used to input normal state data x into IEDN-VAE and obtain output Realize decoupling and reconstruction of input; Residual generator construction module: used to construct a residual generator r related to the input and output observations; Threshold determination module: converts the residual signal into a scalar signal using ^T2 statistics and determines the threshold _Jth of normal samples; Fault detection module: used to calculate the ^T2 test statistic of the online sample residual signal, determine the status of the online sample, and complete the fault detection task; Fault isolation module: used to further use the residual signal to locate the cause variable of the fault in the input sample and complete the fault isolation task; Fault estimation module: It is used to build a fault estimator based on the gradient descent strategy according to the results of fault isolation to complete the fault estimation task. 9. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the VAE-based information enhanced decoupled network fault diagnosis method as described in any one of claims 1 to 7. 10. An information data processing terminal, used for implementing the VAE-based information enhanced decoupled network fault diagnosis system as claimed in claim 8.