A Fault Diagnosis Fusion Method Based on Deep Belief Network and Correlation Model

Abstract

The present invention belongs to the technical field of fault diagnosis of industrial control systems, and specifically relates to a fault diagnosis fusion method based on a deep belief network and a correlation model. The method comprises the following steps: step one, establishing a D matrix model; step two, weighting the D matrix obtained in step one to obtain a weight matrix W; step three, normalizing the weight matrix W obtained in step two with normal distribution; step four, using the normalized weight matrix obtained in step three as the initial weight matrix of the DBN network; step five: inputting test information, outputting fault mode, and completing the fusion fault diagnosis of the D matrix and the DBN network. The present invention establishes a fusion scheme between the correlation model fault diagnosis and the deep belief network fault diagnosis, which improves the convergence speed and prediction accuracy of the DBN network, improves the limitations of the correlation model diagnosis, and has a wider applicability of the fusion diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial control system fault diagnosis, and in particular relates to a fault diagnosis fusion method based on a deep belief network and a correlation model. Background Art

[0002] In order to make the control system work properly in various harsh working environments, strict requirements are also put forward for its reliability. Therefore, the testing and diagnosis of nuclear power control systems have received more and more attention.

[0003] Since product testing and fault diagnosis exist throughout the entire life cycle of the product system, fault diagnosis technology plays a vital role in the maintenance and support of equipment.

[0004] Fault diagnosis technology as a discipline originated in the 1960s. The corresponding test equipment has also gone through a process of development from manual to semi-automatic and then to fully automated, and the intelligence level of test technology has been continuously improved.

[0005] Fault diagnosis technology mainly studies how to detect, isolate and locate faults in the system, that is, to determine whether a fault has occurred, locate the component or module where the fault occurred, and distinguish the type of fault.

[0006] The fault phenomena of nuclear power industrial control systems are complex and diverse. For example, if a parameter of a component exceeds the tolerance range, it can be defined as a fault state.

[0007] In order to improve the accuracy and speed of control system fault diagnosis, reduce the false alarm rate and missed alarm rate, and determine the exact location of the fault, the mainstream trend is to use intelligent algorithms to diagnose equipment systems.

[0008] Deep learning is the main method to promote the development of intelligent diagnosis. The system obtains a network model by training existing data samples, and then diagnoses the system faults. When the system acquires self-learning ability, it can continuously acquire new knowledge from the environment, thereby self-improving the system learning diagnosis.

[0009] Therefore, it is urgent to develop a fault diagnosis method that combines the deep belief network and the correlation model to make up for the high complexity of the deep belief network and improve the prediction accuracy, thereby improving the availability of the control system and reducing the system's life cycle cost. Summary of the invention

[0010] The technical problem to be solved by the present invention is to provide a fault diagnosis fusion method based on a deep belief network and a correlation model, so as to accurately detect, isolate and locate faults and reduce the complexity of fault diagnosis.

[0011] In order to achieve this purpose, the technical solution adopted by the present invention is:

[0012] A fault diagnosis fusion method based on a deep belief network and a correlation model comprises the following steps:

[0013] Step 1: Establish the D matrix model

[0014] Acquire the testability correlation information of the tested object and establish the testability model of the D matrix;

[0015] Step 2: weight the D matrix obtained in step 1 to obtain the weight matrix W;

[0016] Step 3: Normalize the weight matrix W obtained in step 2

[0017] Step 4: The normalized weight matrix obtained in step 3 is used as the initial weight matrix of the DBN network;

[0018] Step 5: Input test information, output fault mode, and complete the fusion fault diagnosis of D matrix and DBN network.

[0019] Furthermore, in the fault diagnosis fusion method based on deep belief network and correlation model as described above, in step 2, corresponding D matrices are established under multiple typical states, each state has a specific probability of occurrence, and its probability is converted into a weight, and multiple D matrices under different states are weighted averaged to obtain a weight matrix W containing test and fault correlation information.

[0020] Furthermore, in the above-mentioned fault diagnosis fusion method based on deep belief network and correlation model, in step 2, the specific process is as follows:

[0021] Step 2 (1) Establish corresponding D matrices under various typical states, select the Kth typical state of the control system, and establish the D matrix D under this state m×n ;

[0022] Step 2 (2): Analyze the probability of occurrence of the state in the entire life cycle of the system and determine the probability as the weight w of the system state. k ;

[0023] Step 2 (3), analyze whether the typical state of the control system is analyzed. If not, go back to step 1; if yes, add all the value values ​​to get the value w;

[0024] Step 2 (4), select D in step 1 m×n The i-th row of the control system is D under all typical states (assuming p states) m×n d inij The value is averaged to get w' ij , expressed as

[0025] Step 2 (5), let j = j + 1 and calculate column by column to determine whether j is greater than n. If not, return to step 4; if so, proceed to step 6;

[0026] Step 2 (6), let i=i+1 and calculate row by row to determine whether i is greater than m. If not, return to step 4; if so, the weighted matrix W is obtained.

[0027] Furthermore, in the above-mentioned fault diagnosis fusion method based on deep belief network and correlation model, in step 3, the weight matrix W is assumed to obey N(μ,σ2) 分布 , normalize it to normal distribution, and the element Y=(X-μ) / σ in the matrix W obeys N(0,1) distribution.

[0028] Furthermore, in the fault diagnosis fusion method based on deep belief network and correlation model as described above, in step 4, the visible element of the DBN network is the information of the test point of the D matrix, and the weight matrix after normal distribution standardization obtained in step 3 is used as the initial weight matrix of the DBN network for network training.

[0029] The beneficial effects of the technical solution of the present invention are:

[0030] The present invention establishes a fusion scheme between correlation model fault diagnosis and deep belief network fault diagnosis, which improves the DBN network convergence speed and prediction accuracy, improves the limitations of correlation model diagnosis, and the applicability of fusion diagnosis is wider. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is the flow chart of the D matrix weighting scheme.

[0032] Figure 2 Schematic diagram of fault diagnosis fusion method. DETAILED DESCRIPTION

[0033] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] The correlation matrix (D matrix) describes the correlation between the system's faults and tests in a specific state. When establishing a correlation matrix, if a fault occurs in a certain component module, theoretically, its test effectiveness is equivalent at each test point accessible by the information flow. In other words, for a certain fault, the corresponding elements of the D matrix at the measurement points accessible by the information flow are all "1". However, in engineering practice applications, the signal may attenuate during the transmission process. In this case, if only one state of the D matrix is ​​used to implement the correlation model diagnosis of the entire system process, it has limitations and cannot fully reflect the characteristics of the entire system. Therefore, the elements corresponding to the measurement points where attenuation changes occur in the D matrix cannot be absolutely set to "1".

[0035] In response to the above problems, the present application proposes to consider starting from the results of testability prediction, and replace each predicted test detection rate with the element value corresponding to the test in the D matrix, so that the D matrix can be weighted. That is, by weighted averaging the D matrices under all typical states of the system, a weighted D matrix is ​​formed, whose elements are values ​​between the interval [0,1]. In step 2 of the present invention, corresponding D matrices are established under multiple typical states, and each state has a specific probability of occurrence. The probability is converted into a weight, and multiple D matrices under different states are weighted averaged to obtain a weight matrix W containing test and fault correlation information.

[0036] The technical solution of the present invention is a fault diagnosis method based on a deep belief network model and a correlation model. Its essence is to use the knowledge of the D matrix to determine the problem of the initial weight matrix. Starting from the D matrix, the test of the D matrix is ​​used as the input of the display layer of the DBN network (deep belief network). Each display element represents the information of a test point. The result is used as the initial weight value of the DBN network through the weighting of the D matrix, which can optimize the initial weight of the DBN network. Figure 2 To design the diagnostic fusion scheme, the data in the weighted D matrix is ​​then processed and calculated using the standard normal distribution as the initial weight matrix of the DBN.

[0037] Weight initialization in DBN is an often overlooked issue. Weight initialization is not equivalent to random weight initialization. The purpose of weight initialization is to break the symmetry of gradient update.

[0038] The present invention provides a fault diagnosis fusion method based on a deep belief network and a correlation model, comprising the following steps:

[0039] Step 1: Establish the D matrix model

[0040] Acquire the testability correlation information of the tested object and establish the testability model of the D matrix;

[0041] Step 2: weight the D matrix obtained in step 1 to obtain the weight matrix W;

[0042] The specific process is as follows Figure 1 As shown:

[0043] Step 2 (1) Establish corresponding D matrices under various typical states, select the Kth typical state of the control system, and establish the D matrix D under this state m×n ;

[0044] Step 2 (2): Analyze the probability of occurrence of the state in the entire life cycle of the system and determine the probability as the weight w of the system state. k ;

[0045] Step 2 (3), analyze whether the typical state of the control system is analyzed. If not, go back to step 1; if yes, add all the value values ​​to get the value w;

[0046] Step 2 (4), select D in step 1 m×n The i-th row of the control system is D under all typical states (assuming p states) m×n d in ij The value is averaged to get w' ij , expressed as

[0047] Step 2 (5), let j = j + 1 and calculate column by column to determine whether j is greater than n. If not, return to step 4; if so, proceed to step 6;

[0048] Step 2 (6), let i=i+1 and calculate row by row to determine whether i is greater than m. If not, return to step 4; if so, the weighted matrix W is obtained.

[0049] Step 3: Normalize the weight matrix W obtained in step 2

[0050] Assume that the weight matrix W obeys N(μ,σ2) 分布 , normalize it to normal distribution, and the element Y=(X-μ) / σ in the matrix W obeys N(0,1) distribution.

[0051] Weight initialization in DBN is an often overlooked issue. Weight initialization is not equivalent to random weight initialization. The purpose of weight initialization is to break the symmetry of gradient update. The probability of hidden layer activation unit is [0,1]. When the element value in the weight matrix W is in the interval [0,1], it will be found that the probability output of the hidden layer activation unit is very close to 0 or 1, which makes the expression of probability invalid, that is, the hidden layer neurons are in a saturated state. So when this happens, a slight adjustment in the weight will only bring extremely slight changes to the activation value of the hidden layer neurons. And this slight change will also affect the remaining neurons in the network, and then bring about corresponding changes in the cost function. As a result, these weights will learn very slowly when we perform the gradient descent algorithm. In order to solve this problem, we standardize the normal distribution of the elements in the weight matrix generated by weighting. After initialization, the weight matrix W obeys a normal distribution with a mean of 0 and a variance of 1.

[0052] Step 4: The normalized weight matrix obtained in step 3 is used as the initial weight matrix of the DBN network;

[0053] The visible element of the DBN network is the information of the test point of the D matrix. The normalized weight matrix obtained in step 3 is used as the initial weight matrix of the DBN network for network training.

[0054] Step 5: Input test information, output fault mode, and complete the fusion fault diagnosis of D matrix and DBN network.

Claims

1. A fault diagnosis fusion method based on deep belief network and correlation model, characterized in that: The steps include: Step 1: Establish the D matrix model Acquire the testability correlation information of the tested object and establish the testability model of the D matrix; Step 2: weight the D matrix obtained in step 1 to obtain the weight matrix W; Step 3: Normalize the weight matrix W obtained in step 2 Step 4: The normalized weight matrix obtained in step 3 is used as the initial weight matrix of the DBN network; Step 5: Input test information, output fault mode, and complete the fusion fault diagnosis of D matrix and DBN network; In step 2, a corresponding D matrix is ​​established for a variety of typical states. Each state has a specific probability of occurrence. Its probability is converted into a weight value, and multiple D matrices under different states are weighted averaged to obtain a weight matrix W containing test and fault correlation information; In step 2, the specific process is as follows Step 2 (1) Establish corresponding D matrices under various typical states, select the Kth typical state of the control system, and establish the D matrix D under this state m×n ; Step 2 (2): Analyze the probability of occurrence of the state in the entire life cycle of the system and determine the probability as the weight w of the system state. k ; Step 2 (3), analyze whether the typical state of the control system is analyzed. If not, go back to step 1; if yes, add all the value values ​​to get the value w; Step 2 (4), select D in step 1 m×n The i-th row of the control system is D under all typical states (assuming p states) m×n d in ij The value is averaged to get w' ij , expressed as Step 2 (5), let j = j + 1 and calculate column by column to determine whether j is greater than n. If not, return to step 4; if so, proceed to step 6; Step 2 (6), let i=i+1 and calculate row by row to determine whether i is greater than m. If not, return to step 4; if so, obtain the weighted matrix W; In step 3, the weight matrix W is assumed to obey N(μ,σ 2 ) distribution, normalize it to normal distribution, and the element Y=(X-μ) / σ in the matrix W obeys N(0,1) distribution; In step 4, the visible element of the DBN network is the information of the test point of the D matrix, and the normalized weight matrix obtained in step 3 is used as the initial weight matrix of the DBN network for network training.

Images