Fault Diagnosis Method for Complex Industrial Systems Based on Multi-Criteria Fusion

Abstract

The present invention discloses a fault diagnosis method for complex industrial systems based on multi-criterion fusion. A fault self-diagnosis model is established with multiple highly sensitive characteristic parameters as inputs and the corresponding operating states and membership degrees as outputs. The quotient between each highly sensitive characteristic parameter of the subsystem to be diagnosed and the corresponding highly sensitive characteristic parameters of other reference subsystems is calculated. A fault mutual-diagnosis model is established with the relative characteristic parameters as inputs and the corresponding operating states and membership degrees as outputs. The highly sensitive characteristic parameters of the status signals collected online are input into the fault self-diagnosis model, the relative characteristic parameters of the subsystem to be diagnosed are calculated and input into the fault mutual-diagnosis model to generate operating state evaluation values in the form of Picture fuzzy numbers for each decision candidate, and a fusion evaluation value is obtained by using the weighted average method. Diagnosis is carried out based on the differences in status information between similar subsystems, which is beneficial to eliminating random factors inside complex industrial systems and improving the diagnosis accuracy and stability.

Description

Technical Field

[0001] The present invention belongs to the field of fault diagnosis, and specifically relates to a method for multi-fault diagnosis of complex industrial systems, especially for complex industrial systems in which two or more similar subsystems operate under the same or similar working conditions respectively. Background Art

[0002] Fault diagnosis methods are becoming increasingly comprehensive, intelligent and diverse, and there are also many classification methods. According to the information source of fault diagnosis, it can be divided into two categories: self-fault diagnosis and mutual-fault diagnosis. Among them, the self-fault diagnosis method formulates fault diagnosis criteria based on the state information of the monitored object itself, and the mutual-fault diagnosis method formulates fault diagnosis criteria based on the state information of the same type of monitored objects under the same or similar operating conditions. At present, for the fault diagnosis of complex systems with multiple similar subsystems, whether it is for local faults or global faults, most of them are based on self-fault diagnosis criteria. For example:

[0003] (1) In 2022, Gao Sujie et al. published a method combining local mean decomposition permutation entropy and BP neural network in the paper "Fault Diagnosis Method of Planetary Gearbox Based on LMD Permutation Entropy and BP Neural Network" in the magazine "Journal of Mechanical Transmission" (Vol. 46, No. 10), which effectively solved the problems of poor discrimination of fault feature vectors and low diagnostic success rate in the fault diagnosis process of multi-stage planetary gearboxes. However, it only uses the global vibration signals of multiple measuring points to diagnose possible single local faults, and cannot distinguish the working conditions of each internal multi-planetary gear, and is even less effective for complex local fault combinations.

[0004] (2) In 2018, Zhou, H. et al. published a method for diagnosing motor torque faults of four-wheel independent drive vehicles in the paper "Motor torque fault diagnosis for four-wheel independent motor-drive vehicle based on unscented Kalman filter" in the magazine "IEEE Transactions on Vehicular Technology" (Vol. 67, No. 3), which can identify the fault situation where there is a difference between the actual output torque of the motor and the torque controlled by the vehicle control unit for the motor. Although it is a consideration of global torque faults, it only considers the correlation between the fault parameters of each drive wheel and the torque fault, and does not further analyze the differences between the wheels.

[0005] Due to the existence of many intricate, strongly correlated and strongly coupled relationships among subsystems or subunits within modern complex industrial facility systems, and the presence of uncertainty factors and uncertain information, faults with characteristics such as randomness, secondary nature, concurrency, and propagation occur frequently. Traditional fault self-diagnosis methods for single equipment, single subsystems or subunits are difficult to discover and consider the correlation relationships within complex systems, resulting in limited diagnostic accuracy and a relatively high probability of misdiagnosis and missed diagnosis. Summary of the Invention

[0006] To address the above problems, the present invention proposes a fault diagnosis method for complex industrial systems based on multi-criterion deep fusion, which simultaneously uses self-diagnosis and mutual-diagnosis criteria for fusion diagnosis. Therefore, it can utilize and consider the correlation within complex industrial systems during the diagnosis process, and can dynamically adjust the diagnostic results to accurately and effectively identify and diagnose faults.

[0007] The fault diagnosis method for complex industrial systems based on multi-criterion deep fusion proposed by the present invention adopts the following technical solutions:

[0008] Step 1): Collect and organize the state signals of each subsystem in different operating states in the complex industrial system, and determine and extract multiple highly sensitive characteristic parameters of the state signals of each subsystem.

[0009] Step 2): Using the multiple highly sensitive characteristic parameters as inputs, and the corresponding operating states and membership degrees as outputs, establish the fault self-diagnosis models of each subsystem, and obtain the test accuracies of each fault self-diagnosis model.

[0010] Step 3): Calculate the quotient between each highly sensitive characteristic parameter of the subsystem to be diagnosed and the corresponding highly sensitive characteristic parameters of other reference subsystems, and use the exponential function characteristics for significant processing as relative characteristic parameters. Using the relative characteristic parameters as inputs, and the corresponding operating states and membership degrees as outputs, establish the fault mutual-diagnosis models of each subsystem, and obtain the test accuracies of each fault mutual-diagnosis model, which together with the test accuracies of the fault self-diagnosis models in Step 2) constitute the full model accuracy matrix.

[0011] Step 4): Normalize each element in the full model accuracy matrix row by row to obtain the initial fusion weights.

[0012] Step 5): Online collect and organize the state signals of each subsystem during the first diagnostic time period in the actual operation process, extract multiple highly sensitive characteristic parameters of the state signals of each subsystem, and input them into the fault self-diagnosis models to output the corresponding membership degrees, forming a membership degree vector.

[0013] Step 6): According to Step 3), calculate the relative characteristic parameters of the subsystem to be diagnosed, and input them into the fault mutual diagnosis model to obtain all membership vectors of each subsystem, which together with the membership vectors in Step 5) form a membership vector matrix;

[0014] Step 7): Take each operating state as a decision candidate, and take all the fault diagnosis models composed of the fault self-diagnosis model and the fault mutual diagnosis model as evaluation attributes. Generate the operating state evaluation values in the form of Picture fuzzy numbers for each decision candidate of each subsystem from the membership vector matrix, and construct a Picture fuzzy matrix from the operating state evaluation values;

[0015] Step 8): Based on the initial fusion weights, use the weighted average method to obtain the fusion evaluation value of the operating state evaluation value, and obtain the fusion evaluation result of the fault diagnosis in the first diagnosis time period according to the fusion evaluation value.

[0016] Step 9): Use the fusion evaluation result in the first diagnosis time period to update the initial fusion weights to obtain the fusion weights in the second diagnosis time period, and repeat the similar methods in Steps 5) to 9) to obtain the online fault diagnosis result of the complex industrial system.

[0017] The beneficial effects of the present invention are as follows:

[0018] 1. On the basis of diagnosing according to the information of the subsystem itself, the present invention proposes to diagnose according to the differences in the state information between similar subsystems, which is more conducive to eliminating the random factors inside the complex industrial system and improving the diagnosis accuracy and stability.

[0019] 2. The relative characteristic parameters proposed by the present invention significantly quantify and characterize the differences between the operating information of similar subsystems, thereby establishing a group of fault diagnosis models based on the mutual diagnosis criterion.

[0020] 3. Based on the Picture fuzzy theory, the present invention more accurately characterizes the evaluation results output by the fault diagnosis model, and proposes a fusion decision-making method for multi-criterion diagnosis, which effectively takes into account the evaluation result information based on multiple criteria and improves the credibility and accuracy of the diagnosis results.

[0021] 4. The present invention effectively optimizes the weight distribution by online rolling adjustment of the fusion weights, and further improves the accuracy and stability of online diagnosis. Description of the Drawings

[0022] The following further describes and explains the technical solutions of the present invention in detail in conjunction with the drawings and specific embodiments:

[0023] Figure 1It is the flow chart for establishing a group of multi-criterion fault diagnosis models in the first stage of the present invention;

[0024] Figure 2 is Figure 1 the flow chart of the relative characteristic parameter calculation method in

[0025] Figure 3 It is the flow chart of the online fault diagnosis method based on multi-criterion fusion in the second stage of the present invention;

[0026] Figure 4 is Figure 3 the flow chart of the multi-criterion-based evaluation result information fusion decision-making method in

[0027] Figure 5 is Figure 3 the flow chart of the update calculation method of the fusion weight matrix in Specific implementation manners

[0028] The complex industrial system targeted by the present invention belongs to modern complex industrial facilities or equipment. The characteristics of this complex industrial system are that there are two or more similar or alike subsystems inside, which operate under the same or similar working conditions respectively. They are relatively independent of each other while also influencing and correlating with each other. They have both static similarity and dynamic interactivity. For example, multi-wheel motor drive systems, multi-stage planetary gear systems, and so on.

[0029] The present invention includes two stages. The first stage is the stage of establishing a diagnosis model based on multi-criteria, and the second stage is the online diagnosis stage of deep multi-criteria fusion.

[0030] The meaning of multi-criteria is that the information or criteria relied on in the diagnosis process are different, including self-diagnosis criteria and mutual-diagnosis criteria. The self-diagnosis criteria can be understood as that the subsystem relies on its own state information for fault diagnosis, and the mutual-diagnosis criteria can be understood as being based on other similar subsystems and relying on the differences between state information for diagnosis. Due to the diversity of reference subsystems, it is collectively called multi-criteria. As Figure 1 shown, the specific implementation steps of the stage of establishing a diagnosis model based on multi-criteria are as follows:

[0031] Step 1: Determine the number n of similar subsystems in the complex system according to the actual situation, where n≥2, and label each subsystem as S (1) , S (2) ,... S (n)Determine the status signals monitored by each subsystem. The status signals may include, for example, vibration signals, temperature signals, current signals, etc. Determine the m operating status categories included in the diagnostic results of a single subsystem, denoted as STAT1, STAT2, …, STATm, also denoted as STATi, where i = 1, 2, ..., m. Among them, STAT1 is the normal operating status, and the remaining operating statuses are simple fault statuses or complex fault statuses. Alternatively, without distinguishing specific fault statuses, the operating status categories can be simply divided into two categories: fault and normal.

[0032] Analyze and determine multiple highly sensitive characteristic parameters to be extracted from the status signals of each subsystem according to the actual situation. For example, extract the root mean square value, average peak value, skewness, kurtosis, and waveform stability from the vibration signal as highly sensitive characteristic parameters; extract the temperature rise value from the temperature signal as a highly sensitive characteristic parameter; extract the current covariance from the current signal as a highly sensitive characteristic parameter, etc. Sequentially label the highly sensitive characteristic parameters to be used by each subsystem as SP1, SP2, SP3, ... SP r , where r is the number of characteristic parameters.

[0033] Collect and organize the data of each status signal of each subsystem under different operating statuses, and associate the above-mentioned characteristic parameter data extracted within a short time slice with the corresponding operating status to form a sample. In this way, the data of multiple time slices form multiple samples, constituting the sample dataset of each subsystem. Randomly select 70% of the samples in the sample dataset of each subsystem as the training set of this subsystem, and the remaining samples as the test set of this subsystem, which are respectively used for the training and verification of the subsequent diagnostic model.

[0034] Step 2: For the first subsystem S (1) , use machine learning methods to establish a multi-label classification model, that is, a multi-input multi-output classification model, such as a neural network model, an artificial hydrocarbon network model, etc., and use the training set data of the first subsystem S (1) for supervised training. The input of the model is the characteristic parameters SP1, SP2, SP3, ... SP (1) of the first subsystem S r , and the output of the model is the operating status categories STAT1, STAT2, ... STATm and the membership degrees q STAT1 , q STAT2 ,... q STATm , q error , where the membership degrees q STAT1 , q STAT2 ,... q STATm represent the membership degrees of the subsystem S (1)Probability of being in each operating state, membership degree q error Indicates the subsystem S (1) The probability that the operating state result of does not belong to any preset state. The trained model is the fault self-diagnosis model M based on the self-diagnosis criterion 11 , simply referred to as the self-diagnosis model M 11 .

[0035] Similarly, using the same method as the first subsystem S (1) , for the remaining n - 1 subsystems S (2) ,...S (n) Similar multi-label classification models are respectively established to obtain the self-diagnosis models of the remaining n - 1 subsystems. All n self-diagnosis models are composed of the fault self-diagnosis model group M based on the self-diagnosis criterion ii (i = 1, 2,...n).

[0036] Use the test set data to verify the prediction accuracy of each of the above self-diagnosis models. The test accuracy δ of each self-diagnosis model M ii , δ ii , δ ii ∈[0, 1]. From the test accuracy rate δ ii the accuracy matrix D = (δ 11 , δ 22 ,...δ nn ) is obtained, (i = 1, 2,...n).

[0037] Step 3: The similar subsystems in the complex industrial system have both static similarities and dynamic interactions. By studying the differences between subsystems, anomalies in the operation of subsystems can be clearly perceived. All subsystems are divided into subsystems to be diagnosed and other subsystems, and other subsystems are used as reference subsystems. Therefore, relative characteristic parameters are defined, and the operation state of the subsystem to be diagnosed is investigated by using the correlation between the subsystem to be diagnosed and other reference subsystems. For the convenience of understanding, taking the subsystem S (1) and the remaining reference subsystems S (k) (k = 2, 3,..., n) as an example for illustration, the calculation method of the relative characteristic parameter is as Figure 2 shown, and the specific description is as follows:

[0038] The relative characteristic parameter is used to characterize the differential characteristics between two subsystems, and is specifically quantified by examining the difference relationship of the characteristic parameters between the two subsystems with one subsystem as the reference. First, calculate the quotient between the corresponding characteristic parameters SP1, SP2, SP3,...SP (1) of the first subsystem S (k) and the reference subsystem S r :

[0039]

[0040] Among them, represents the quotient of the characteristic parameters corresponding to subsystem S (1) and subsystem S (k) (k = 2, 3,... n), which is simply referred to as the characteristic quotient. The subscript i corresponds one by one in order to the characteristic parameters SP1, SP2, SP3,... SP r The subscript corresponds one by one in order, and SP i (1) represents the value of the characteristic parameter of subsystem S (1) , and SP i (k) represents the value of the characteristic parameter of the benchmarked subsystem S (k) .

[0041] Since the differentiation of the calculation result data of the above characteristic quotient may not be obvious, the characteristics of the exponential function are used to significantly process the characteristic quotient as follows:

[0042]

[0043] In the formula, the base a > 1, and the value of a is flexibly determined according to the actual distribution of the characteristic quotient, represents the relative characteristic parameter of subsystem S (1) with respect to subsystem S (k) (k = 2, 3,... n) as the benchmark. The subscript i corresponds one by one in order to the subscripts of the characteristic parameters SP1, SP2, SP3,... SP r . Obviously, when the value of the characteristic parameter corresponding to subsystem S (1) is equal to that of the benchmark subsystem, the relative characteristic parameter When the characteristic parameter of subsystem S (1) is greater than the characteristic parameter of the benchmark subsystem, the relative characteristic parameter On the contrary, the relative characteristic parameter

[0044]

[0045] For the subsystem S (1) to be diagnosed, one by one, with the remaining subsystems S (k) (k = 2, 3,... n) as the benchmark, the different operating states of subsystem S (1) are combined with the different operating states of other benchmark subsystems S (k) . Based on the characteristic parameter data of each subsystem in different states, according to the definition of the above relative characteristic parameter, calculate the relative characteristic parameter (1) of subsystem S (k) and the benchmark subsystem S Form a sample data set of relative characteristic parameters. Randomly select 70% of the sample data set as the training set of relative characteristic parameters, and the remaining samples as the test set of relative characteristic parameters, which are used for the training and verification of the subsequent diagnostic model respectively.

[0046] For the subsystem S to be diagnosed (1) , one by one, taking the rest of the subsystems S (k) (k = 2, 3,... n) as the benchmark, using machine learning methods, establish a multi-label classification model respectively, that is, a classification model with multiple inputs and multiple outputs, such as a neural network model, an artificial hydrocarbon network model, etc., and use the relative characteristic parameter training set data of the subsystem S (1) benchmarked on S (k) for supervised training. The model input is the relative characteristic parameter and the output is the operation state categories STAT1, STAT2,…, STATm and the membership degrees q corresponding to the unrecognized situations STAT1 , q STAT2 ,... q STATm , q error , which respectively represent the probabilities of the subsystem S (1) being in each operation state and the result not belonging to any preset state. The trained model is the fault classification model M 1j based on the mutual diagnosis criterion, j = 2, 3,... n, which is also simply called the mutual diagnosis model, and the subscript j indicates benchmarking on the rest of the subsystems S (k) (k = 2, 3,... n).

[0047] Similarly, for the remaining n - 1 subsystems, establish fault mutual diagnosis models respectively based on the mutual diagnosis criterion, and obtain the fault mutual diagnosis model group M ij (i = 1, 2,... n; j = 1, 2,... n; i ≠ j).

[0048] Combine this fault mutual diagnosis model group M ij with the above-mentioned fault self-diagnosis model group M ii to jointly form a fault diagnosis model group based on multiple criteria, which is intuitively represented as follows:

[0049]

[0050] Among them, one row represents all the fault diagnosis models of a subsystem to be diagnosed based on multiple criteria.

[0051] Use the test set data to verify the fault mutual diagnosis models of each subsystem, and derive the accuracy δ ij (i = 1, 2,... n; j = 1, 2,... n; i ≠ j) of each mutual diagnosis model, δ ij∈[0,1], corresponding one by one in order to the mutual diagnosis model M ij (i = 1, 2,...n; j = 1, 2,...n; i≠j), together with the aforementioned self-diagnosis model accuracy matrix D = (δ 11 , δ 22 ,...δ nn ), jointly constitute the full model accuracy matrix:

[0052]

[0053] Step Four: For the obtained full model accuracy matrix Z above, normalize each element in the full model accuracy matrix Z row by row, and then determine the initial fusion weights of each subsystem's fault diagnosis model based on different criteria, denoted as the initial weight matrix ω (0) As shown below, for the subsequent actual online fusion diagnosis process:

[0054]

[0055] In the formula,

[0056]

[0057] Among them, represents the initial fusion weight of the multi-criteria based fault diagnosis model M ij (i = 1, 2,...n; j = 1, 2,...n), the subscripts have a one-to-one correspondence, and one row of the initial weight matrix ω (0) represents the initial weights of each diagnosis model of a subsystem, and the sum of the elements in each row is 1, that is

[0058] As Figure 3 shown is the online diagnosis stage based on multi-criteria fusion in the second stage. The specific steps are as follows:

[0059] Step One: Reasonably select the time period length of a single diagnosis, online collect and organize the operation information of each subsystem during the actual operation in the first diagnosis time period, and extract the high-sensitivity characteristic parameter data of the aforementioned state monitoring signals.

[0060] Step Two: Input the characteristic parameters SP1, SP2, SP3,...SP (k) (k = 1, 2,...n) of each subsystem S r correspondingly into the fault self-diagnosis model M ii (i = 1, 2,...n) of each subsystem based on the self-diagnosis criterion, and respectively output the aforementioned membership values q STAT1 , q STAT2 ,...q STATm , qerror , which constitutes the membership vector q of the self-diagnosis model of each subsystem ii =(q STAT1 ,q STAT2 ,...q STATm ,q error ), i=1,2,...,n, corresponding to the fault self-diagnosis model group M based on the self-diagnosis criterion in sequence ii (i=1,2,...n).

[0061] Step 3: Investigate the subsystem S to be diagnosed (1) , respectively, with the remaining subsystems S (k) (k=2,3,...n) as the benchmark, calculate the relative characteristic parameters corresponding to the above characteristic parameters And input the corresponding mutual diagnosis model M based on mutual diagnosis criteria 1j (j=2,3,...n), the corresponding output is the subsystem S (1) The resulting membership vector q of all mutual diagnostic criterion models 1j =(q STAT1 ,q STAT2 ,...q STATm ,q error ),j=2,3,...n.

[0062] Similarly, for the remaining subsystems S to be diagnosed (k) (k=2,3,...n) perform model diagnosis and identification based on mutual diagnosis criteria, and obtain the result membership vector q ij (i=2,3,...n;j=1,2,...n;i≠j). At this point, all models are output and the membership vector matrix of all results is obtained. The intuitive representation is as follows:

[0063]

[0064] Among them, a row represents the result membership vector of all fault diagnosis models of a subsystem based on multiple criteria.

[0065] Step 4: The fuzzy set theory is used to handle fuzzy and uncertain problems and is applied to the field of multi-attribute decision-making. The traditional intuitionistic fuzzy set gives information on both the approval and opposition aspects for each element, while the Picture fuzzy set is extended to four aspects: approval, neutral, opposition, and abstention. A Picture fuzzy number is defined as p = (α, β, γ, ρ), where α, β, γ, ρ represent the approval rate, neutral rate, opposition rate, and abstention rate respectively, and α, β, γ, ρ ∈ [0, 1], and α + β + γ + ρ = 1. The Picture fuzzy number contains more information and considers more comprehensively, so that more accurate decisions can be made. The present invention converts the output result of the multi-criterion model into a running state evaluation value in the form of a Picture fuzzy number and performs in-depth fusion decision-making, such as Figure 4 as shown

[0066] First, evaluate and fuse the decision of the diagnosis results of the multi-criterion model for the subsystem S (1) in the first time period. Using the foregoing preset operating states STAT1, STAT2,... STATm as decision candidates, the subsystem S (1) Based on all the fault diagnosis models M 1j (j = 1, 2,... n) of the multi-criterion as evaluation attributes, generate a running state evaluation value in the form of a Picture fuzzy number for each candidate running state of the subsystem S(1) from the result membership degree vectors q 11 , q 12 ,... q 1n The specific method is as follows:

[0067] Define the running state evaluation value p (1) of different evaluation attributes of the subsystem S ij = (α ij , β ij , γ ij , ρ ij )(i = 1, 2,..., m; j = 1, 2,..., n), which is a Picture fuzzy number. Intuitively, it is the evaluation value of the possible running states STATi of the subsystem S (1) based on different criteria fault diagnosis models M 1j for the subsystem S (1) . The subscripts i, j of the evaluation value correspond to the running state and the fault diagnosis model. Among them, the approval rate α ij (i = 1, 2,..., m; j = 1, 2,..., n) represents the degree of approval of the model M 1j for the subsystem S (1) being in the state STATi at this time, and is equal to the membership degree q 1j corresponding to the state STATi in the result membership degree vector q STATi ; the neutral rate βij (i = 1, 2, ..., m; j = 1, 2, ..., n) represents the model M 1j for the subsystem S (1) At this time, the neutral degree in state STATi is equal to the sum of the membership degrees of all the remaining states that potentially contain the faults in STATi. The calculation method can also be flexibly determined according to the actual situation; the abstention rate ρ ij (i = 1, 2, ..., m; j = 1, 2, ..., n) represents the model M 1j for the subsystem S (1) At this time, the abstention degree in state STATi is equal to the membership degree q of the unrecognized situation in the sample in the result membership degree vector q 1j ; the opposition rate γ error ; the opposition rate γ ij (i = 1, 2, ..., m; j = 1, 2, ..., n) represents the model M 1j for the subsystem S (1) At this time, the opposition degree in state STATi, by definition, γ ij = 1 - α ij - β ij - ρ ij . Obviously, when the neutral rate β ij and the abstention rate ρ ij are 0, the Picture fuzzy number p ij degenerates into an intuitionistic fuzzy number.

[0068] After obtaining the evaluation values of the diagnostic model for each candidate operating state, a Picture fuzzy matrix for the first time period is constructed using the evaluation values:

[0069]

[0070] One row of the matrix represents the different evaluation values of all the fault diagnosis models based on multiple criteria of the subsystem S (1) for an operating state.

[0071] The evaluation results of different diagnostic models for the current operating state of the subsystem S (1) are different. It is necessary to fuse the evaluation values of different models for the same operating state, that is, to aggregate the Picture fuzzy numbers in the fuzzy matrix P (1) row by row. Based on the initial fusion weight Adopting the idea of weighted average, the fusion evaluation value of each candidate operating state STATi (i = 1, 2,...m) under the joint evaluation of the multi-criteria diagnostic model is calculated based on the aforementioned fusion weight which is also a Picture fuzzy number, where respectively represent the fusion approval rate, the fusion neutral rate, the fusion opposition rate and the fusion abstention rate, and the calculation methods are as follows:

[0072]

[0073] In the formula, is the element of the aforementioned initial fusion weight matrix ω (0) in the first row, that is, the initial fusion weight of each diagnostic model of subsystem S (1)

[0074] Calculate the comprehensive score of each candidate operating state from the fusion evaluation value and the effectiveness score For a candidate operating state, the greater the difference between the fusion approval rate and the fusion opposition rate, the higher the comprehensive score is considered, and it is more inclined to consider that subsystem S (1) is in this operating state during the current time period; the lower the fusion abstention rate, that is, the higher the sum of the other three items, the more effective the fusion evaluation result is considered.

[0075] According to the comprehensive score of the candidate operating state, sort each candidate operating state STAT1, STAT2,... STATm from high to low according to the score. When the comprehensive scores are the same, sort them from high to low according to the effectiveness score The operating state ranked first finally is considered to be the operating state of subsystem S (1) during the current time period, and this is the operating state of subsystem S (1) in the current time period, that is, the fault diagnosis result, namely the fusion evaluation result.

[0076] Since there are differences between the results output by different fault diagnosis models and the final fusion evaluation result, using this difference, the fusion weight in the next stage is updated in real time to make the distribution of the fusion weight adapt to the actual diagnosis process, so as to improve the accuracy of real-time diagnosis. The update method is as Figure 5 shown. The specific description is as follows. For the convenience of explaining the following update method of the fusion weight, mark the diagnostic result state in the current time period, that is, the fusion evaluation result of the fault diagnosis, as STATx, x ∈ {1, 2,... m}, indicating that STATx is one of {STAT1, STAT2,... STATm}, and it is only used for unknown reference.

[0077] The basis for updating the fusion weight is the degree of closeness between the evaluation value under the evaluation of each model based on multiple criteria and the fusion evaluation value. Since an evaluation value in the form of a Picture fuzzy number has four sub-values, it is proposed to regard the evaluation value and the fusion evaluation value as a kind of four-dimensional points. Based on the calculation method of the distance between points, calculate the Euclidean distance between the evaluation value of each model for the diagnostic result state STATx and the fusion evaluation value:​

[0078]

[0079] where p xj =(α xj , β xj , γ xj , ρ xj )(j = 1, 2,... n) represents the evaluation value of the fault diagnosis model M of the subsystem S(1) obtained from the aforementioned calculation based on different criteria for the result operation state STATx, 1j and represents the evaluation value of the result operation state STATx under multi-model evaluation. α xj , β xj , γ xj , ρ xj (1) respectively represent the approval rate, neutral rate, opposition rate, and abstention rate for the result operation state STATx, and respectively represent the fused approval rate, fused neutral rate, fused opposition rate, and fused abstention rate for the result operation state STATx.

[0080] Based on the difference in the Euclidean distance between the evaluation value of the result operation state STATx obtained from the calculation of each model and the final fused evaluation value, the initial fused weights (1) of each model for the subsystem S are corrected so that the weight of the model with a smaller Euclidean distance increases accordingly, and the weight of the model with a larger Euclidean distance decreases accordingly, thereby obtaining the fused weight matrix ω (1) for the next time period. The specific update calculation steps for the elements in the matrix are as follows:

[0081] (1) Using the normalization method, calculate the weight value of the Euclidean distance between the evaluation value of the result operation state STATx obtained from each model and the fused evaluation value:

[0082]

[0083] where λ j (j = 1, 2,... n) is the weight value of the Euclidean distance between the evaluation value of the result operation state STATx of the multi-criteria diagnosis model M (1) of the subsystem S 1j and the fused evaluation value. 0 ≤ λ j ≤ 1. The smaller the weight value of the Euclidean distance under a certain model evaluation, the closer the evaluation result based on this model is to the fused evaluation result, and it is considered that the performance of this model in actual diagnosis is relatively better.

[0084] (2) Use the weight value of the above Euclidean distance to update the fused weight matrix. The update calculation method is as follows:

[0085]

[0086] Among them, is the fused weight matrix ω for the updated second time period (1) The elements of the first row, that is, the subsystem S (1) The fused weight for the second time period of each fault diagnosis model based on multiple criteria is the aforementioned initial fused weight matrix ω (0) The elements of the first row, that is, the subsystem S (1) The initial fused weights of each fault diagnosis model based on multiple criteria; b is the update rate, which is used to control the aggressiveness of the fused weight update. The larger b is, the greater the update amplitude of the fused weight.

[0087] Step Five: Adopt a method similar to that in Step Four for the actual online diagnosis of the subsystem S (1) to fuse and evaluate the operating states of the remaining subsystems to be diagnosed S (k) (k = 2, 3,... n) in the first time period, that is, the diagnostic results of each subsystem, and also update the remaining row elements of the fused weight matrix ω (1) in the second time period to obtain the fused weight matrix for the second time period:

[0088]

[0089] Summarize the diagnostic results of all subsystems in the first time period and output the final diagnostic result of the complex industrial system in the first time period.

[0090] Step Six: Collect and organize the operating information of each subsystem during the actual operation in the next diagnostic time period, extract the characteristic parameter data of each state monitoring signal, and repeat the similar methods in Steps Two to Five. In this way, the fault diagnosis for subsequent diagnostic time periods t (t > 1) is carried out in a cyclic and rolling manner for each time period, so as to realize the online diagnosis process.

Claims

1. A fault diagnosis method for complex industrial systems based on multi-criteria fusion, characterized in that Including: Step 1): Collect and organize the state signals of each subsystem in the complex industrial system under different operating states, and determine and extract multiple highly sensitive characteristic parameters of the state signals of each subsystem; Step 2): Using the multiple highly sensitive characteristic parameters as inputs, and the corresponding operating states and membership degrees as outputs, establish the fault self-diagnosis models of each subsystem, and obtain the test accuracies of each fault self-diagnosis model; Step 3): Calculate the quotient between each highly sensitive characteristic parameter of the subsystem to be diagnosed and the corresponding highly sensitive characteristic parameters of other reference subsystems, and use the exponential function property for significant processing as the relative characteristic parameter. Using the relative characteristic parameter as the input, and the corresponding operating state and membership degree as the output, establish the fault mutual-diagnosis models of each subsystem, and obtain the test accuracies of each fault mutual-diagnosis model, which together with the test accuracies of the fault self-diagnosis models in Step 2) form the full model accuracy matrix; Step 4): Normalize each element in the full model accuracy matrix row by row to obtain the initial fusion weights; Step 5): Online collect and organize the state signals of each subsystem in the first diagnosis time period during the actual operation process, extract multiple highly sensitive characteristic parameters of the state signals of each subsystem, and input them into the fault self-diagnosis model to output the corresponding membership degrees, forming a membership degree vector; Step 6): According to Step 3), calculate the relative characteristic parameters of the subsystem to be diagnosed, and input them into the fault mutual-diagnosis model to obtain all membership degree vectors of each subsystem, which together with the membership degree vector in Step 5) form a membership degree vector matrix; Step 7): Using each operating state as a decision candidate, and using all the fault diagnosis models composed of the fault self-diagnosis model and the fault mutual-diagnosis model as evaluation attributes, generate the operating state evaluation values in the form of Picture fuzzy numbers for each decision candidate of each subsystem from the membership degree vector matrix, and construct a Picture fuzzy matrix from the operating state evaluation values; Step 8): Based on the initial fusion weights, use the weighted average method to obtain the fusion evaluation value of the operating state evaluation value, and obtain the fusion evaluation result of the fault diagnosis in the first diagnosis time period according to the fusion evaluation value; Step 9): Using the fusion evaluation result in the first diagnosis time period, update the initial fusion weights to obtain the fusion weights in the second diagnosis time period, and repeat Steps 5) to 9) to obtain the online fault diagnosis result of the complex industrial system.

2. The fault diagnosis method for a complex industrial system according to Claim 1, characterized in that: In Step 9), calculate the Euclidean distance between the operating state evaluation value of the fusion evaluation result in the current time period and the fusion evaluation value, increase the weight of the model with a smaller Euclidean distance accordingly, and decrease the weight of the model with a larger Euclidean distance accordingly, and update the initial fusion weights in real time.

3. The fault diagnosis method for a complex industrial system according to Claim 1, characterized in that: In step 3), the relative characteristic parameters are processed using an exponential function. The base a > 1, i = 1, 2,... r, k = 2, 3,... n, where n is the number of subsystems and r is the number of characteristic parameters.

4. The fault diagnosis method for complex industrial systems according to claim 1, characterized in that: The initial fusion weight described in step 4) δ ij is the test accuracy of each of the fault mutual diagnosis models, δ ij ∈[0, 1], i = 1, 2,... n, j = 1, 2,... n, i ≠ j, and n is the number of subsystems.

5. The fault diagnosis method for complex industrial systems according to claim 1, characterized in that: In step 7), a Picture fuzzy number p = (α, β, γ, ρ), where α, β, γ, ρ represent the approval rate, neutral rate, opposition rate, and abstention rate respectively, α, β, γ, ρ ∈ [0, 1], and α + β + γ + ρ = 1.

6. The fault diagnosis method for complex industrial systems according to claim 5, characterized in that: the operation state evaluation value p of different decision candidates ij = (α ij , β ij , γ ij , ρ ij ), and the fusion evaluation value respectively represent the fusion approval rate, the fusion neutral rate, the fusion opposition rate, and the fusion abstention rate is the initial fusion weight matrix ω of the first subsystem (0) The elements in the first row, and m is the number of operating state categories 7. The fault diagnosis method for complex industrial systems according to claim 6, characterized in that: According to the fusion evaluation value Calculate the comprehensive scores of each operating state and the effectiveness scores Based on the comprehensive scores Sort the candidate operating states in descending order of scores; when the comprehensive scores are the same, sort them according to the effectiveness scores in descending order. The operating state ranked first is the fusion evaluation result of the subsystem in the current time period.

8. The fault diagnosis method for complex industrial systems according to claim 2, characterized in that: The Euclidean distance mentioned above p xj =(α xj , β xj , γ xj , ρ xj ) is the evaluation value of the entire fault diagnosis model of the subsystem for the operating state STATx, is the fusion evaluation value of the operating state STATx, j = 1, 2,... n, x ∈ {1, 2,... m}, and m is the number of operating state categories, respectively represent the fusion approval rate, the fusion neutral rate, the fusion opposition rate, and the fusion abstention rate. α xj , β xj , γ xj , ρ xj respectively represent the approval rate, the neutral rate, the opposition rate, and the abstention rate.

9. The fault diagnosis method for complex industrial systems according to claim 8, wherein: Calculate the weights of the Euclidean distances between the evaluation values of each model for the operating state STATx and the fused evaluation value Based on λ j Weights, according to the formula Update the weights, Is the fused weight matrix ω for the updated second time period (1) The elements of the first row, Is the initial fused weight matrix ω (0) The elements of the first row, b is the update rate.

10. The fault diagnosis method for complex industrial systems according to claim 9, characterized in that: Normalize each element of the full model accuracy matrix row by row to determine the initial fusion weights of each subsystem and establish an initial weight matrix:

Images