A fault diagnosis method

Abstract

The application discloses a kind of fault diagnosis methods.The fault diagnosis method includes obtaining diagnostic rule table, and establishes fault diagnosis model according to diagnostic rule table;Establish fault reasoning model and reinforcement learning model, fault reasoning model is based on fault diagnosis model and carries out uncertainty reasoning, to obtain possible first fault reason according to fault phenomenon, and the probability value that first fault reason appears is reasoned out, and fault reasoning model compares the probability value that first fault reason appears with preset threshold value:if the probability value that first fault reason appears meets preset threshold value, then first fault reason is as the result of fault reasoning model;If the probability value that first fault reason appears does not meet preset threshold value, then reinforcement learning model updates diagnostic rule table.According to the fault diagnosis method of the application, it has the ability of automatic learning and automatic updating, has more complete fault diagnosis knowledge and more accurate fault diagnosis capability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, in particular to a fault diagnosis method. BACKGROUND

[0002] The fault diagnosis method is mainly aimed at how to detect, separate and identify the fault occurred in the system, that is, to determine whether the fault occurs, locate the part and type of the fault occurrence, and determine the size and time of the fault occurrence, etc. Effective fault diagnosis method enables people to take measures to prevent the further development of the fault at the appropriate time, thereby avoiding the occurrence of accidents, improving the utilization rate of equipment, reducing maintenance costs, thereby reducing production costs and improving the comprehensive competitiveness of enterprises. SUMMARY

[0003] A series of simplified concepts are introduced in the summary part, which will be further described in detail in the specific embodiment part. The summary part of the present application does not mean to try to limit the key features and necessary technical features of the claimed technical solution, nor to try to determine the protection scope of the claimed technical solution.

[0004] In order to at least partially solve the above problems, according to the first aspect of the present application, a fault diagnosis method is provided, characterized in that the fault diagnosis method comprises:

[0005] Obtaining a diagnosis rule table, and establishing a fault diagnosis model according to the diagnosis rule table;

[0006] Establishing a fault reasoning model and a reinforcement learning model, the fault reasoning model is based on the fault diagnosis model to perform uncertainty reasoning, to obtain a possible first fault cause according to a fault phenomenon, and to reason a probability value of the first fault cause appearing,

[0007] The fault reasoning model compares the probability value of the first fault cause appearing with a preset threshold value:

[0008] If the probability value of the first fault cause appearing meets the preset threshold value, the first fault cause is taken as the result of the fault reasoning model;

[0009] If the probability value of the first fault cause appearing does not meet the preset threshold value, the reinforcement learning model updates the diagnosis rule table.

[0010] According to the fault diagnosis method, the fault diagnosis method comprises: acquiring a diagnosis rule table, and establishing a fault diagnosis model according to the diagnosis rule table; establishing a fault reasoning model and a reinforcement learning model, the fault reasoning model performs uncertain reasoning based on the fault diagnosis model to obtain a possible first fault cause according to a fault phenomenon, and reason out a probability value of occurrence of the first fault cause, and the fault reasoning model compares the probability value of occurrence of the first fault cause with a preset threshold value: if the probability value of occurrence of the first fault cause meets the preset threshold value, the first fault cause is taken as a result of the fault reasoning model; if the probability value of occurrence of the first fault cause does not meet the preset threshold value, the reinforcement learning model updates the diagnosis rule table. In this way, the automatic learning and automatic updating capability is obtained, the problems such as rule contradiction, matching conflict and combination explosion in the complex model adjustment process are avoided, the requirement is reduced, the self-learning and self-adaptive capability is obtained, the good extraction and induction capability is obtained, the more complete fault diagnosis knowledge and the more accurate fault diagnosis capability are obtained, and the self-learning and self-adaptive process of the fault diagnosis model realizes rule mining, the new diagnosis rule table is formed through the reinforcement learning, and then the fault diagnosis model is further optimized.

[0011] Optionally, in the diagnosis rule table, a same fault phenomenon corresponds to multiple fault causes, and different fault causes correspond to different confidence degrees respectively, and / or,

[0012] In the diagnosis rule table, a same fault cause corresponds to multiple possible fault phenomena, and different fault phenomena correspond to different confidence degrees respectively.

[0013] Optionally, the fault reasoning model reasons out the probability value of occurrence of the first fault cause according to the different confidence degrees.

[0014] Optionally, the reinforcement learning model updates the diagnosis rule table by using a reinforcement learning algorithm, and the reinforcement learning algorithm comprises:

[0015] A Q(s, a) table is established, in which, an environment state s of the reinforcement learning model comprises all values of the confidence degrees in the diagnosis rule table;

[0016] An action a of the reinforcement learning model is set according to the Q(s, a) table, and the action a is an adjustment of the confidence degree value;

[0017] The reinforcement learning model adjusts the environment state s to a new environment state s' through the action a, and calculates a reward r;

[0018] The diagnosis rule table is updated according to the new environment state s'.

[0019] Optionally, the reinforcement learning algorithm further comprises: updating the Q(s, a) table according to the new environment state s' based on Bellman equation:

[0020] Q(s, a) = Q(s, a) + a[r + max a′ Q(s', a') - Q(s, a)].

[0021] Optionally, the action a comprises increasing and decreasing the credibility.

[0022] Optionally, the fault diagnosis model is updated according to the updated diagnosis rule table, and the fault reasoning model performs uncertainty reasoning based on the updated fault diagnosis model to derive a possible second fault cause according to the fault phenomenon and reason a probability value of occurrence of the second fault cause.

[0023] Optionally, the fault reasoning model compares the probability value of occurrence of the second fault cause with a preset threshold.

[0024] Optionally, if the probability value of occurrence of the second fault cause meets the preset threshold, the second fault cause is taken as a result of the fault reasoning model, and the reinforcement learning model obtains a positive reward.

[0025] Optionally, if the probability value of occurrence of the second fault cause does not meet the preset threshold, the reinforcement learning model continues to update the diagnosis rule table, and the reinforcement learning model obtains a punishment reward.

[0026] Optionally, if the probability value of occurrence of the second fault cause is negative, the reinforcement learning model obtains a doubled punishment reward.

[0027] Optionally, the fault reasoning model reasons the probability value of occurrence of the first fault cause according to the different credibility by using the following formula:

[0028]

[0029] Optionally, the diagnosis rule table is acquired according to existing fault cases and / or expert knowledge. BRIEF DESCRIPTION OF DRAWINGS

[0030] The following drawings for the present application are hereby incorporated as part of the present application for the purpose of understanding the present application. The embodiments of the present application and its description shown in the drawings are used to explain the devices and principles of the present application. In the drawings,

[0031] Figure 1 a flow chart of a fault diagnosis method according to a preferred embodiment of the present application;

[0032] Figure 2a flowchart of a reinforcement learning algorithm according to a preferred embodiment of the present application; and

[0033] Figure 3 a visualized graph of a fault diagnosis model according to a preferred embodiment of the present application. DETAILED DESCRIPTION

[0034] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. It will be apparent, however, to one skilled in the art that the present application can be practiced without one or more of these specific details. In other instances, well-known features have not been described in detail so as not to unnecessarily obscure the present application.

[0035] In order to fully understand the present application, detailed descriptions will be given in the following description with reference to the accompanying drawings so as to explain the present application. It is apparent that the practice of the present application is not limited to the particular details described hereafter. The preferred embodiments of the present application are described in detail as follows, however, the present application can have other embodiments in addition to these detailed descriptions, and should not be construed as being limited to the embodiments presented herein.

[0036] It is to be understood that the terms used herein are merely for describing specific embodiments and are not intended to limit the present application, and the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. When the terms "comprise" and / or "include" are used in the specification, it is indicated that there are the features, numbers, steps, operations, elements, and / or components, but it does not exclude the presence or addition of one or more other features, numbers, steps, operations, elements, components, and / or combinations thereof. The terms "upper," "lower," "front," "rear," "left," "right," and similar expressions used in the present application are for illustrative purposes only, and are not limiting.

[0037] The ordinal numbers such as "first" and "second" used in the present application are merely used to identify the components, and do not have any other meaning, such as a particular order, etc. Also, for example, the term "first component" does not by itself imply the existence of a "second component", and the term "second component" does not by itself imply the existence of a "first component".

[0038] Hereinafter, specific embodiments of the present application will be described in greater detail with reference to the accompanying drawings, which illustrate representative embodiments of the present application, and are not intended to limit the present application.

[0039] As Figure 1As shown, the present application provides a fault diagnosis method of a preferred embodiment, the fault diagnosis method provided by the present application can accurately detect, separate and identify the fault occurred in the system, that is, determine whether the fault occurs, locate the position and type of the fault, and determine the size and time of the fault. The fault diagnosis method of the present application has the ability of automatic learning and automatic updating, avoids the problems of rule contradiction, matching conflict, combination explosion and the like in the complex model adjustment process, reduces the requirements, has the self-learning and self-adaptive ability, has good extraction and induction ability, has more complete fault diagnosis knowledge and more accurate fault diagnosis ability.

[0040] Specifically, the fault diagnosis method comprises the following steps:

[0041] Obtaining a diagnosis rule table.

[0042] The diagnosis rule table comprises at least one fault phenomenon and at least one fault cause, one fault phenomenon can correspond to one fault cause, and one fault phenomenon can also correspond to multiple fault causes. The diagnosis rule table further comprises at least one credibility, the credibility refers to the probability value of the fault cause corresponding to the fault phenomenon. The credibility (also referred to as rule strength) can represent the degree of certainty that the fault phenomenon is caused by the corresponding fault cause.

[0043] The structure of the diagnosis rule table is shown in the following table (1):

[0044] Table (1)

[0045]

[0046] In table (1), multiple fault phenomena and multiple fault causes are included, and the same fault phenomenon can correspond to multiple fault causes. For example, fault phenomenon A can correspond to three fault causes C, D and E, fault phenomenon A can be caused by fault cause C, or fault phenomenon A can be caused by fault cause D, or fault phenomenon A can be caused by fault cause E. Different fault causes also correspond to different credibilities. For example, the credibility of fault phenomenon A caused by fault cause C is ω1. The credibility of fault phenomenon A caused by fault cause D is ω2. The credibility of fault phenomenon A caused by fault cause E is ω3.

[0047] Fault phenomenon B can correspond to two fault causes C and F, fault phenomenon B can be caused by fault cause C, or fault phenomenon can be caused by fault cause F. The credibility of fault phenomenon B caused by fault cause C is ω4. The credibility of fault phenomenon B caused by fault cause F is ω5.

[0048] As can be seen from Table (1), one fault cause can correspond to multiple possible fault phenomena. For example, fault cause C corresponds to fault phenomena A and B, fault cause C corresponds to possible fault phenomenon A, and fault cause C also corresponds to possible fault phenomenon B. Different fault phenomena correspond to different confidence levels. The confidence level of fault cause C corresponding to possible fault phenomenon A is ω1. The confidence level of fault cause C corresponding to possible fault phenomenon B is ω4.

[0049] The diagnosis rule table can be obtained according to existing fault cases, which include fault phenomena, fault causes and confidence levels. The existing fault cases can be summarized into the diagnosis rule table. The diagnosis rule table can be obtained according to expert knowledge. Experts summarize the corresponding content into the diagnosis rule table according to their own experience.

[0050] In the diagnosis rule table, one fault phenomenon corresponds to multiple fault causes, and different fault causes correspond to different confidence levels.

[0051] In the diagnosis rule table, one fault cause corresponds to multiple possible fault phenomena, and different fault phenomena correspond to different confidence levels.

[0052] The diagnosis rule table can not only be obtained from existing fault cases, but also be given by experts, and the combination of the two can better realize the combination of expert knowledge and fault cases, and eliminate the heavy work of repeatedly manually adjusting model parameters by human experts in the process of creating, maintaining and upgrading the fault diagnosis model.

[0053] For ease of understanding, the following Table (2) takes a steam turbine as an example for illustration:

[0054] Table (2)

[0055]

[0056] Table (2) includes multiple fault phenomena and multiple fault causes. The same fault phenomenon can correspond to multiple fault causes. For example, the fault phenomenon of low-speed shaft oil film vortex (eddy current) can correspond to two fault causes: low-speed shaft 1# sliding bearing oil film vortex (comprehensive) and low-speed shaft 2# sliding bearing oil film vortex (comprehensive). The fault phenomenon of low-speed shaft oil film vortex (eddy current) can be caused by the fault cause of low-speed shaft 1# sliding bearing oil film vortex (comprehensive), or the fault phenomenon of low-speed shaft oil film vortex (eddy current) can be caused by the fault cause of low-speed shaft 2# sliding bearing oil film vortex (comprehensive). Different fault causes also correspond to different credibility. For example, the credibility of the fault phenomenon of low-speed shaft oil film vortex (eddy current) being caused by the fault cause of low-speed shaft 1# sliding bearing oil film vortex (comprehensive) is 0.3 (i.e., the probability value is 30%). The fault phenomenon, low-speed shaft oil film vortex (eddy current), can be caused by the fault cause, low-speed shaft 2# sliding bearing oil film vortex (comprehensive), with a credibility of 0.3 (i.e., a probability value of 30%).

[0057] The fault phenomenon of low-speed shaft rubbing (eddy current) can correspond to two fault causes: local friction (middle) of low-speed shaft sliding bearing #1 and local friction (middle) of low-speed shaft sliding bearing #2. The fault phenomenon of low-speed shaft rubbing (eddy current) can be caused by the fault cause of local friction (middle) of low-speed shaft sliding bearing #1, or the fault phenomenon of low-speed shaft rubbing (eddy current) can be caused by the fault cause of local friction (middle) of low-speed shaft sliding bearing #2. Different fault causes also correspond to different credibility levels. For example, the credibility level of the fault phenomenon of low-speed shaft rubbing (eddy current) being caused by the fault cause of local friction (middle) of low-speed shaft sliding bearing #1 is 0.3 (i.e., a probability value of 30%). The credibility level of the fault phenomenon of low-speed shaft rubbing (eddy current) being caused by the fault cause of local friction (middle) of low-speed shaft sliding bearing #2 is 0.3 (i.e., a probability value of 30%).

[0058] The fault phenomenon of low-speed shaft #1 sliding bearing failure (vibration) can correspond to two fault causes: full-circle friction (intermediate) and localized friction (intermediate). The fault phenomenon of low-speed shaft #1 sliding bearing failure (vibration) can be caused by the fault cause of full-circle friction (intermediate), or it can be caused by the fault cause of localized friction (intermediate). Different fault causes also correspond to different confidence levels. For example, the confidence level for the fault phenomenon of low-speed shaft #1 sliding bearing failure (vibration) being caused by the fault cause of full-circle friction (intermediate) is 0.4 (i.e., a probability value of 40%). The confidence level for the fault phenomenon of low-speed shaft #1 sliding bearing failure (vibration) being caused by the fault cause of localized friction (intermediate) is 0.4 (i.e., a probability value of 40%).

[0059] The fault phenomenon of the low-speed shaft 2# sliding bearing failure (vibration) can correspond to two fault causes, i.e., the low-speed shaft 2# sliding bearing whole-circle friction (intermediate) and the low-speed shaft 2# sliding bearing local friction (intermediate). The fault phenomenon of the low-speed shaft 2# sliding bearing failure (vibration) can be caused by the fault cause of the low-speed shaft 2# sliding bearing whole-circle friction (intermediate), or the fault phenomenon of the low-speed shaft 2# sliding bearing failure (vibration) can be caused by the fault cause of the low-speed shaft 2# sliding bearing local friction (intermediate). Different fault causes also correspond to different reliabilities. For example, the reliability of the fault phenomenon of the low-speed shaft 2# sliding bearing failure (vibration) caused by the fault cause of the low-speed shaft 2# sliding bearing whole-circle friction (intermediate) is 0.4 (i.e., the probability value is 40%). The reliability of the fault phenomenon of the low-speed shaft 2# sliding bearing failure (vibration) caused by the fault cause of the low-speed shaft 2# sliding bearing local friction (intermediate) is 0.4 (i.e., the probability value is 40%).

[0060] It can also be seen from Table (2) that the same fault cause can also correspond to multiple possible fault phenomena. For example, the fault cause of the low-speed shaft shaft imbalance (comprehensive) corresponds to the generated fault phenomena of the low-speed shaft 1# sliding bearing whole-circle friction (intermediate) and the low-speed shaft 2# sliding bearing whole-circle friction (intermediate), the fault cause of the low-speed shaft shaft imbalance (comprehensive) corresponds to the possible fault phenomenon of the low-speed shaft 1# sliding bearing whole-circle friction (intermediate), and the fault cause of the low-speed shaft shaft imbalance (comprehensive) also corresponds to the possible fault phenomenon of the low-speed shaft 2# sliding bearing whole-circle friction (intermediate). Different fault phenomena also correspond to different reliabilities. The reliability of the fault cause of the low-speed shaft shaft imbalance (comprehensive) corresponding to the possible fault phenomenon of the low-speed shaft 1# sliding bearing whole-circle friction (intermediate) is 0.5 (i.e., the probability value is 50%). The reliability of the fault cause of the low-speed shaft shaft imbalance (comprehensive) corresponding to the possible fault phenomenon of the low-speed shaft 2# sliding bearing whole-circle friction (intermediate) is 0.5 (i.e., the probability value is 50%).

[0061] A fault diagnosis model is established according to the diagnosis rule table.

[0062] The diagnosis rule table is input into a data processing device, such as an industrial computer or the like, and a fault diagnosis model is established in the data processing device, thereby facilitating subsequent fault diagnosis and data processing. The fault diagnosis model is a white box model, which can be established by expert knowledge without fault samples, and the fault diagnosis model has interpretability.

[0063] A fault reasoning model is established, and the fault reasoning model performs uncertain reasoning based on the fault diagnosis model.

[0064] The fault reasoning model can perform uncertain reasoning based on a large amount of data and logical relationships in the fault diagnosis model, thereby performing fault diagnosis. For example, Figure 3The fault phenomenon and the fault cause of the fault diagnosis model are visualized. One fault phenomenon can correspond to multiple fault causes, and one fault cause can also cause multiple possible fault phenomena. A fault phenomenon can also be a fault cause of another fault phenomenon.

[0065] Specifically, the fault phenomenon whole-circle friction can be caused by the fault cause temperature anomaly or the fault cause vibration anomaly, and thus the fault phenomenon whole-circle friction can also cause the fault phenomenon poor lubrication or the fault phenomenon imbalance. The fault phenomenon imbalance can be caused by the fault cause whole-circle friction or the fault cause shaft center locus imbalance. The fault phenomenon shafting misalignment can be caused by the fault cause shaft center locus misalignment or the fault cause bearing misalignment.

[0066] A fault cause can also be a fault phenomenon. The fault cause bearing misalignment can also be a fault phenomenon caused by the fault cause local friction or the fault cause oil film whirl. The fault cause local friction can also be a fault cause of the oil film whirl. The fault cause local friction can cause different fault phenomena, such as the fault cause local friction can also cause the fault phenomenon oil film whirl or the fault phenomenon collision. The fault cause local friction can also be a fault phenomenon caused by the fault cause vibration anomaly or the fault cause shaft center locus friction. The fault cause oil film whirl can be a fault phenomenon caused by the fault cause local friction or the fault cause shaft center locus oil film whirl, and the fault cause oil film whirl can also cause the fault phenomenon looseness. The fault phenomenon looseness can also be caused by the fault cause vibration anomaly.

[0067] The credibility in the diagnosis rule table is an uncertainty measure from a fault phenomenon to a fault cause. -1≤ω≤1. Taking Table (1) as an example, when the fault phenomenon A appears and the fault cause C is true, ω1=1; when the fault phenomenon A and the fault cause C are irrelevant, ω1=0; when the fault phenomenon A appears and the fault cause C is false, ω1=-1. ω1≥0 indicates that the fault phenomenon A supports the fault cause C, and ω1<0 indicates that the fault phenomenon A does not support the fault cause C.

[0068] The fault reasoning model performs uncertainty reasoning based on the credibility of the fault diagnosis model, obtains a possible first fault cause according to a fault phenomenon, and reasons a probability value of the first fault cause.

[0069] When multiple fault phenomena correspond to the same fault cause, the credibility of the multiple fault phenomena fusion is calculated. The fault reasoning model is a white box model. The fault reasoning model reasons based on the fault diagnosis model, and the rules and logic in the fault diagnosis model are intuitive and understandable. The fault reasoning model reasons the probability value of the occurrence of the first fault cause according to different credibilities. In this way, the accuracy of the probability value reasoned by the fault reasoning model can be improved. The fault reasoning model can reason the probability value of the occurrence of the first fault cause according to different credibilities using the following formula:

[0070]

[0071] As described above, different fault phenomena can be caused by the same fault cause. In Table (1), fault phenomenon A and fault phenomenon B can be caused by fault cause C. When fault phenomenon A and fault phenomenon B occur at the same time, the fusion probability of fault cause C is calculated using the calculation rule in the above formula. Assuming that ω1 and ω4 are greater than 0, when fault phenomenon A and fault phenomenon B occur at the same time, CF(C) = ω1 + ω4 - ω1ω4.

[0072] When fault phenomenon A and fault phenomenon B occur at the same time, the rule strengths of fault cause C, fault cause D, fault cause E, and fault cause F are CF(C) = ω1 + ω4 - ω1ω4, CF(D) = ω2, CF(E) = ω2, and CF(F) = ω5, respectively.

[0073] The fault reasoning model reasons the possible first fault cause of the fault phenomenon according to the above-mentioned uncertain reasoning method, and reasons the probability value of the occurrence of the first fault cause according to the above-mentioned formula. For example, referring to the content of Table (2), the fault reasoning model reasons that the possible first fault cause of causing the fault phenomenon of low-speed shaft oil film whirl (electro-vortex) is low-speed shaft 1# sliding bearing oil film whirl (comprehensive) according to the uncertain reasoning method, and reasons that the probability value of the occurrence of the first fault cause is 30% (credibility 0.3) according to the above-mentioned formula.

[0074] The fault reasoning model also compares the probability value of the occurrence of the first fault cause with a preset threshold value.

[0075] If the probability value of the occurrence of the first fault cause meets the preset threshold value, the first fault cause is taken as the result of the fault reasoning model.

[0076] The threshold value CFthresh is set, and when the probability value of the first fault cause reasoned by the fault reasoning model meets CFthresh, it is considered that there is such a fault, and it is taken as the result of the fault reasoning model.

[0077] For example, the preset threshold CFthresh is 20%~50%, the first fault cause of low-speed shaft 1# sliding bearing oil film vortex (comprehensive) is inferred by the fault reasoning model according to the uncertain reasoning method, and the probability value is 30% (confidence 0.3). The probability value 30% inferred by the fault reasoning model is greater than the preset threshold 20%, and it can be considered that the fault cause of the low-speed shaft oil film vortex (electric eddy current) causing the fault phenomenon is the low-speed shaft 1# sliding bearing oil film vortex (comprehensive), and the fault cause of the low-speed shaft 1# sliding bearing oil film vortex (comprehensive) is taken as the result of the fault reasoning model.

[0078] The reinforcement learning model is established. The reinforcement learning model can learn from the fault cases to automatically update the diagnostic rule table.

[0079] If the probability value of the first fault cause does not meet the preset threshold, the reinforcement learning model updates the diagnostic rule table.

[0080] The threshold CFthresh is set. When the probability value of the first fault cause inferred by the fault reasoning model does not meet CFthresh, it can be considered that the result inferred by the fault reasoning model is wrong, and the reinforcement learning model updates the diagnostic rule table.

[0081] For example, the preset threshold CFthresh is 40%~50%, the first fault cause of low-speed shaft 1# sliding bearing oil film vortex (comprehensive) is inferred by the fault reasoning model according to the uncertain reasoning method, and the probability value is 30% (confidence 0.3). The probability value 30% inferred by the fault reasoning model is less than the preset threshold 40%, and it can be considered that the probability value of the first fault cause inferred by the fault reasoning model is wrong, and the diagnostic rule table needs to be updated so that the fault reasoning model can infer the correct probability value again.

[0082] According to the fault diagnosis method, the fault diagnosis method comprises: acquiring a diagnosis rule table, and establishing a fault diagnosis model according to the diagnosis rule table; establishing a fault reasoning model and a reinforcement learning model, the fault reasoning model performs uncertain reasoning based on the fault diagnosis model to obtain a possible first fault cause according to a fault phenomenon, and reason out a probability value of occurrence of the first fault cause, and the fault reasoning model compares the probability value of occurrence of the first fault cause with a preset threshold value: if the probability value of occurrence of the first fault cause meets the preset threshold value, the first fault cause is taken as a result of the fault reasoning model; and if the probability value of occurrence of the first fault cause does not meet the preset threshold value, the reinforcement learning model updates the diagnosis rule table. In this way, the automatic learning and automatic updating capability is obtained, the problems such as rule contradiction, matching conflict and combination explosion in the complex model adjustment process are avoided, the requirement is reduced, the self-learning and self-adaptive capability is obtained, the good extraction and induction capability is obtained, the more complete fault diagnosis knowledge and the more accurate fault diagnosis capability are obtained, and the self-learning and self-adaptive process of the fault diagnosis model realizes rule mining, the new diagnosis rule table is formed through the reinforcement learning, and the fault diagnosis model is further optimized.

[0083] The updating of the diagnosis rule table by the reinforcement learning model is specifically as follows:

[0084] The reinforcement learning model updates the diagnosis rule table by using a reinforcement learning algorithm, so that the reinforcement learning model can automatically update the diagnosis rule table through self-learning of the fault phenomenon and the fault cause, thereby causing the fault reasoning model to re-reason out a correct probability value. The reinforcement learning model can update the credibility in the diagnosis rule table in the fault diagnosis model by using a reinforcement learning QLearning algorithm through automatic learning of reasoning cases, and optimize the current fault diagnosis model.

[0085] The QLearning algorithm mainly realizes model optimization through updating of a Q(s, a) table. As shown in FIG. 2, the reinforcement learning algorithm comprises the following steps: Figure 2

[0086] The Q(s, a) table is established.

[0087] The learning rate a and the decay coefficient g of the reinforcement learning model are selected. The Q(s, a) table is established according to the learning rate a and the decay coefficient g.

[0088] The Q(s, a) table value is initialized. In the Q(s, a) table, the environment state s of the reinforcement learning model comprises all values of the credibility in the diagnosis rule table.

[0089] ​The credibility value in the fault diagnosis model can be used as the initial environment state s of the reinforcement learning model. The possibility of all values of the credibility in the current diagnosis rule table can be used as the environment state s of the reinforcement learning model. The possibility of all values of the credibility in the current fault diagnosis model can be used as the environment state s of the reinforcement learning model. Different environment states s represent different fault diagnosis models.

[0090] The action a of the reinforcement learning model is set according to the Q(s, a) table, and the action a is an adjustment to the credibility value.

[0091] The action a of the reinforcement learning model is an adjustment to the credibility value. The action a can be selected according to the current Q(s, a) table value by using an ε-greedy strategy. The action a includes increasing and decreasing the credibility. In this way, the action a can increase and decrease a single credibility value. For example, the fault reasoning model infers that the probability value of the occurrence of the first fault reason, low-speed shaft 1# sliding bearing oil film vortex (synthesis), is 30% (credibility 0.3) according to the uncertain reasoning method. The action a of the reinforcement learning model can increase the credibility value, and increase the credibility value by 0.1. Of course, the action a of the reinforcement learning model can decrease the credibility value, and decrease the credibility value by 0.1.

[0092] The reinforcement learning model adjusts the environment state s to a new environment state s' through the action a, and calculates the reward r.

[0093] Taking table (1) as an example, the environment state s of the current reinforcement learning model is [ω1, ω2, ω3, ω4, ω5]. The adjustment amount τ of the action a to the environment state s is set each time. The action a executable by the reinforcement learning model is ω1±τ, ω2±τ, ω3±τ, ω4±τ, ω5±τ, a total of ten. Taking the execution of ω1+τ as an example, the new environment state s' after the execution of the action a is changed to [ω1', ω2', ω3', ω4', ω5'].

[0094] For example, the environment state s of the current reinforcement learning model is [0.3, 0.5]. The adjustment amount τ of the action a of the reinforcement learning model to the environment state s is τ=0.1. The action a executable by the reinforcement learning model is 0.3±0.1, 0.5±0.1. Taking the execution of 0.3+0.1 as an example, the new environment state s' after the execution of the action a is changed to [0.4, 0.6].

[0095] The reinforcement learning model can obtain a reward r according to a final calculation result. The reward r of the reinforcement learning model includes three parts r1, r2 and r3, and the total reward of the model is r = r1 + r2 + r3. r1 is a positive reward, representing the degree of coincidence between the output result of the current fault reasoning model and the fault case to be learned. If the output result of the fault reasoning model is consistent with the fault case, a reward can be obtained, otherwise no reward can be obtained. r2 is a penalty reward, and the reinforcement learning model will obtain a penalty each time it is adjusted. The more the adjustment times, the greater the cumulative penalty. r3 is a doubled penalty reward (i.e. an out-of-bounds reward), and if the new environment state s' after adjustment by the action a is an illegal state (such as a certain credibility CF(B, A) ≤ -1 or CF(B, A) ≥ 1 in the new environment state s'), a doubled penalty reward is received.

[0096] The diagnostic rule table is updated according to the new environment state s'.

[0097] The new environment state s' after the execution of the action a is changed to [0.4, 0.6], and the diagnostic rule table is updated according to the new environment state. The diagnostic rule table includes a large number of credibility values and logical relationships, and the new environment state will not directly replace the existing credibility. The new environment state can be added to the diagnostic rule table, or the credibility value can be adjusted and the logical relationship can be updated according to the actual situation in the diagnostic rule table, so as to update the diagnostic rule table.

[0098] The Q(s, a) table is updated based on the Bellman equation according to the new environment state s':

[0099] Q(s, a) = Q(s, a) + a[r + γmax a′ Q(s', a') - Q(s, a)].

[0100] The updated diagnostic rule table can update the Q(s, a) table, and then the reinforcement learning model can further self-learn, and further obtain the updated environment state s', s'', s''',... through the action a, so as to cyclically update the diagnostic rule table. Thus, the reinforcement learning model can continuously self-learn, and further continuously update the diagnostic rule table.

[0101] The fault diagnosis model is updated according to the new diagnostic rule table.

[0102] The reinforcement learning model can obtain a new environment state s' through the action a, so as to update the diagnostic rule table, and further update the fault diagnosis model. The fault reasoning model can perform uncertainty reasoning based on the updated fault diagnosis model to obtain possible second fault reasons according to the fault phenomenon, and infer the probability value of the occurrence of the second fault reason.

[0103] The fault diagnosis model can be updated according to the updated diagnosis rule table. The fault reasoning model performs uncertainty reasoning again based on the updated fault diagnosis model, and obtains a possible second fault cause according to the fault phenomenon. The first fault cause and the second fault cause are both stored fault causes in the diagnosis rule table. The first fault cause and the second fault cause can be the same fault cause or different fault causes. The fault reasoning model also reasons a probability value of the second fault cause, so that the fault reasoning model can obtain a new probability value again, repeatedly perform reasoning diagnosis, and determine the correct fault cause corresponding to the fault phenomenon.

[0104] The fault reasoning model compares the probability value of the second fault cause with a preset threshold value.

[0105] As described above, the fault reasoning model performs uncertainty reasoning again, obtains a possible second fault cause, and reasons a probability value of the second fault cause. The fault reasoning model compares the probability value of the second fault cause with a preset threshold value.

[0106] If the probability value of the second fault cause meets the preset threshold value, the second fault cause is taken as the result of the fault reasoning model, and the reinforcement learning model obtains a positive reward.

[0107] If the probability value of the second fault cause reasoned by the fault reasoning model is consistent with a fault case to be learned, the second fault cause is taken as the result of the fault reasoning model, and the reinforcement learning model also obtains a positive reward.

[0108] The setting of the reward r of the reinforcement learning model follows |r1|>|r2|>|r3| (for example, r1=100, r2=-10, r3=-0.1). If the probability value of the second fault cause reasoned by the fault reasoning model is consistent with a fault case to be learned, the reward obtained by the action a is r=r1+r2.

[0109] A threshold value CFthresh is set. When the probability value of the second fault cause reasoned by the fault reasoning model meets CFthresh, it is considered that there is such a fault, which is taken as the result of the fault reasoning model and the loop is ended.

[0110] For example, the preset threshold value CFthresh is 40%-50%, and the fault reasoning model reasons that the probability value of the occurrence of the second fault cause is 40% (confidence 0.4) according to the uncertainty reasoning method. The probability value 40% reasoned by the fault reasoning model meets the range of the preset threshold value, and it is considered that the fault cause causing the fault phenomenon is the second fault cause, and the second fault cause is taken as the result of the fault reasoning model. At the same time, the action a of the reinforcement learning model is correct, the reinforcement learning model obtains a positive reward, and the reward obtained by the action a is r=r1+r2.

[0111] If the probability value of the second fault cause does not meet the preset threshold value, the reinforcement learning model continues to update the diagnosis rule table, and the reinforcement learning model obtains a penalty reward.

[0112] A threshold value CFthresh is set. When the probability value of the second fault cause inferred by the fault reasoning model does not meet CFthresh, it is considered that the result inferred by the fault reasoning model is wrong, and the reinforcement learning model updates the diagnosis rule table. And the reinforcement learning model obtains a penalty.

[0113] The step of continuing to update the diagnosis rule table by the reinforcement learning model is similar to the above step, which will not be repeated here. The probability value of the second fault cause does not meet the preset threshold value, and the new environment state s' adjusted by the action a of the reinforcement learning model is an error state, so the reward of the action a is r=r2. For example, the probability value of the second fault cause inferred by the fault reasoning model based on the new diagnosis rule table is 35%, which is not within the range of the preset threshold value CFthresh 40%-50%. Therefore, the new environment state s' adjusted by the action a is an error state, and the reward of the action a is r=r2.

[0114] If the probability value of the second fault cause is negative, the reinforcement learning model obtains a doubled penalty reward.

[0115] If the probability value of the second fault cause inferred by the fault reasoning model based on the updated fault diagnosis model is negative, it means that the action of the reinforcement learning model does not meet the conventional logic, so the reinforcement learning model will obtain a doubled penalty reward. That is, the new environment state s' obtained after the action a of the reinforcement learning model is adjusted is an illegal state, and the reward obtained by the action a is r=r2+r3.

[0116] In this way, the reinforcement learning model obtains different types of rewards in self-learning. By designing a reasonable reward and penalty mechanism, the reinforcement learning model can optimize itself in trial and error, thereby better updating the diagnosis rule table. The reinforcement learning model can use reinforcement learning algorithms to adjust and optimize the uncertain fault reasoning model, pursue the global optimum of the model, and the trained model can be migrated and deployed, without the need for further training next time.

[0117] According to the fault diagnosis method of the application, the reinforcement learning is used to optimize the uncertainty model (fault reasoning model), the reinforcement learning method is introduced into the fault reasoning model, not only the rules can be given by experts, but also the model can be updated adaptively according to the fault cases, the automatic learning and automatic updating ability according to the gradually accumulated fault cases is possessed, the combination of the expert knowledge and the fault cases is better realized, the heavy work of repeatedly manually adjusting the model parameters of the human experts in the fault diagnosis model creation, maintenance and upgrading process is saved, the problems such as rule contradiction, matching conflict and combination explosion in the complex model adjustment process are avoided, the requirement is reduced, the self-learning and self-adaptive ability is possessed, the good extraction and induction ability is possessed, the more complete fault diagnosis knowledge and the more accurate fault diagnosis ability are possessed, the process of self-learning and self-adaptation of the fault diagnosis model realizes the rule mining, the new diagnosis rule table formed through the reinforcement learning is further used to optimize the fault diagnosis model, and the problems of complicated establishment and maintenance and update of the fault reasoning model are solved and the pain point of huge workload is solved.

[0118] Although example implementations have been described herein with reference to the accompanying drawings, it is to be understood that the example implementations are only exemplary and are not intended to limit the scope of the application. Various changes and modifications can be made therein by those ordinarily skilled in the art without departing from the scope and spirit of the application. All such changes and modifications are intended to be included within the scope of the application as claimed in the appended claims.

[0119] In the specification provided herein, a large number of specific details are explained. However, it can be understood that the embodiments of the application can be practiced without these specific details. In some examples, well-known methods, structures and techniques are not shown in detail in order not to obscure the understanding of the present specification.

[0120] Similarly, it is to be understood that, in the description of the example implementations of the application, the various features of the application are sometimes grouped together in a single implementation, figure or description of an implementation for the purpose of streamlining the disclosure and aiding in the understanding of one or more of the various aspects of the application. However, the method of the application should not be construed to require that the features be present in any single implementation, figure or description of an implementation. Rather, the various aspects of the application, as described in the claims, can be present alone or in any combination of one or more features. Thus, the method of the application is to be understood in its broadest and most generic form, and not limited to the specific examples of the application described above.

[0121] Those skilled in the art will appreciate that all features described herein (including all features and processes described in the accompanying claims, abstract and drawings) can be combined in any combination. Unless specifically stated otherwise, each feature disclosed in this specification (including any accompanying claims, abstract and drawings) can be replaced by alternative features that serve the same, equivalent or similar purpose, unless otherwise expressly stated.

[0122] Furthermore, those skilled in the art will recognize that references in the specification to "one embodiment", "an embodiment", "an example embodiment" etc., mean that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. The appearances of the phrase "in one embodiment" in various places in the specification are not necessarily all referring to the same embodiment.

[0123] The above description is only specific embodiments of the present application or specific explanations of the specific embodiments, the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. The protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A fault diagnosis method, characterized in that: The fault diagnosis method comprises: Obtain a diagnosis rule table, and establish a fault diagnosis model based on the diagnosis rule table, wherein: In the diagnosis rule table, the same fault phenomenon corresponds to multiple fault causes, and different fault causes correspond to different credibility levels, and / or, In the diagnosis rule table, the same fault cause corresponds to multiple possible fault phenomena, and different fault phenomena correspond to different credibility levels; Establishing a fault reasoning model and a reinforcement learning model, wherein the fault reasoning model performs uncertainty reasoning based on the fault diagnosis model to derive a possible first fault cause based on the fault phenomenon, and infers a probability value of the first fault cause based on the different credibility levels. The fault inference model compares the probability value of the occurrence of the first fault cause with a preset threshold: If the probability value of the occurrence of the first fault cause meets the preset threshold, the first fault cause is used as the result of the fault inference model; If the probability value of the occurrence of the first fault cause does not meet the preset threshold, the reinforcement learning model updates the credibility in the diagnosis rule table. The reinforcement learning model uses a reinforcement learning algorithm to update the diagnosis rule table, and the reinforcement learning algorithm includes: Establishing a Q(s, a) table, in which the environmental state s of the reinforcement learning model includes all values ​​of the credibility in the diagnostic rule table; Setting an action a of the reinforcement learning model according to the Q(s, a) table, wherein the action a is an adjustment of the credibility value; The reinforcement learning model adjusts the environment state s to a new environment state s' through action a and calculates the reward r; The diagnosis rule table is updated according to the new environmental state s'.

2. The fault diagnosis method according to claim 1, characterized in that: The reinforcement learning algorithm further includes: updating the Q(s, a) table based on the Bellman equation according to the new environment state s': Q(s,a)=Q(s,a)+α[r+γmax a′ Q(s′,a′)-Q(s,a)]。 3. The fault diagnosis method according to claim 1, characterized in that: The action a includes increasing and decreasing the credibility.

4. The fault diagnosis method according to claim 1, characterized in that: The fault diagnosis model is updated according to the updated diagnosis rule table, and the fault reasoning model performs uncertainty reasoning based on the updated fault diagnosis model to derive a possible second fault cause according to the fault phenomenon and infer a probability value of the occurrence of the second fault cause.

5. The fault diagnosis method according to claim 4, characterized in that: The fault inference model compares the probability value of the occurrence of the second fault cause with a preset threshold.

6. The fault diagnosis method according to claim 5, characterized in that: If the probability value of the occurrence of the second fault cause meets the preset threshold, the second fault cause is used as the result of the fault inference model, and the reinforcement learning model obtains a positive reward.

7. The fault diagnosis method according to claim 5, characterized in that: If the probability value of the occurrence of the second fault cause does not meet the preset threshold, the reinforcement learning model continues to update the diagnosis rule table, and the reinforcement learning model obtains a penalty reward.

8. The fault diagnosis method according to claim 7, characterized in that: If the probability value of the second fault cause is negative, the reinforcement learning model receives a double penalty reward.

9. The fault diagnosis method according to claim 1, characterized in that: The fault inference model uses the following formula to infer the probability value of the first fault cause according to the different credibility:

10. The fault diagnosis method according to claim 1, characterized in that: The diagnosis rule table is obtained based on existing fault cases and / or expert knowledge.

Images