A failure mode prediction method, apparatus, and computer readable medium

Abstract

Embodiments of the present application relate to natural language processing technology, and particularly relate to a fault mode prediction method and device and a computer readable medium. A model generation method includes: obtaining m target fault modes and a set of first keywords corresponding to each target fault mode, the first keywords are based on knowledge of a specific field and are used to describe the characteristics of the corresponding target fault mode; obtaining at least one historical fault case of the specific field; obtaining at least one second keyword from the at least one historical fault case; in the process of training a fault mode prediction model, adding each second keyword to m sets of first keywords respectively, so that the sum of m mutual informations is maximum, each mutual information corresponds to a target fault mode, and is the mutual information between the first keyword corresponding to the target fault mode added with the second keyword and the target fault mode; outputting m target fault modes and keywords used to describe the characteristics of the target fault modes.

Description

Technical Field

[0001] This invention relates to the field of Natural Language Processing (NLP) technology, and more particularly to a fault mode prediction method, apparatus, and computer-readable medium. Background Technology

[0002] Equipment fault diagnosis is becoming increasingly important for customers. Although each fault has its own characteristics, classifying faults, identifying common phenomena from historical fault cases, and determining different fault modes are still very beneficial for using the knowledge learned from historical fault cases to solve new problems.

[0003] In a typical approach to failure mode prediction using historical failure cases, domain experts collect similar historical failure cases, use their domain knowledge to identify common phenomena or deduce possible phenomena, and ultimately summarize failure modes with specific failure phenomena, as well as methods for resolving failures exhibiting these patterns. When a new failure occurs, the domain expert can compare the phenomena of the new failure with existing failure modes to accurately determine the new failure mode. Once determined, solutions to existing failure modes can be used to resolve the new failure. Therefore, identifying the failure mode of a new failure is highly beneficial and important for customers.

[0004] However, relying solely on domain experts for failure mode prediction is quite difficult. First, it requires extensive domain knowledge to analyze and correctly classify historical failure cases, a task not everyone can accomplish. Second, the sheer number of historical failure cases can be overwhelming, making it impossible for domain experts to analyze and classify them, or to exhaustively enumerate all possible failure modes.

[0005] One labor-saving approach is to use Natural Language Processing (NLP) for failure mode prediction. In the field of NLP, failure mode prediction can be considered a topic modeling task. Given a set of documents (historical failure cases), the topic model can generate topic (failure mode) labels, thus semantically associating documents with the same labels.

[0006] Latent Dirichlet Allocation (LDA) is a traditional topic model that has proven effective in solving many tasks. LDA assumes that documents are generated from a mixture of topics, each of which can be represented as a probability distribution of words. Each word is generated by multiple topics with varying probabilities. The goal of LDA is to determine which topics can generate the actual words in a document. LDA is an unsupervised model that treats each document as a combination of words. The topic labels it generates are relatively consistent with human understanding.

[0007] Currently, there are other topic modeling algorithms that consider word order, such as neural topic models. Neural topic models also treat documents as a distribution of topics, and topics as a distribution of words. The difference is that in neural topic models, words are generated sequentially, and neural networks (RNNs or LSTMs) extract the relevance between words. Similarly, parameters are calculated by neural networks instead of Gibbs sampling.

[0008] LDA and neural topic models are successful in performing topic modeling tasks in many general domains. However, they are less effective for specific domain tasks. Because the entire process is unsupervised, we cannot control the algorithm to generate the topics we want.

[0009] In summary, relying solely on domain experts is insufficient for fault mode prediction; and relying solely on NLP algorithms for fault mode prediction in a specific domain cannot generate the desired topics. Therefore, there is currently no method that is both labor-saving and effective for fault mode prediction in a specific domain. Summary of the Invention

[0010] This invention provides a fault mode prediction method, apparatus, and computer-readable medium. Using historical fault cases and limited domain knowledge as input, a fault mode prediction model can be generated using machine learning algorithms. This model can effectively predict the fault modes of new faults.

[0011] Firstly, a model generation method is provided for generating a fault mode prediction model. This method involves obtaining m target fault modes and a set of first keywords corresponding to each target fault mode, where m is a positive integer. Each set of first keywords is based on knowledge from a specific domain and is used to describe the features of the corresponding target fault mode. At least one historical fault case from the specific domain is obtained. At least one second keyword is obtained from the at least one historical fault case. During the training of the fault mode prediction model, each second keyword is added to the m sets of first keywords, where a second keyword can only be added to one set of first keywords, to maximize the sum of the m mutual information. Each mutual information corresponds to a target fault mode and is the mutual information between the first keyword corresponding to the target fault mode for which the second keyword has been added and the target fault mode itself. Finally, m target fault modes are output, and for each target fault mode, a set of keywords describing the features of that target fault mode is output.

[0012] Secondly, a fault mode prediction method is provided. This method receives new fault cases from a specific domain; inputs the new fault cases into a fault mode prediction model; and outputs the target fault mode to which the new fault cases belong, obtained by running the fault mode prediction model. The fault mode prediction model is trained using the method provided in the first aspect.

[0013] Thirdly, a model generation apparatus is provided for generating a fault mode prediction model, including a module for executing the method provided in the first aspect, such as a program module.

[0014] Fourthly, a fault mode prediction apparatus is provided, including a module for performing the method provided in the second aspect, such as a program module.

[0015] Fifthly, a model generation apparatus is provided for generating a failure mode prediction model, comprising: at least one memory configured to store computer-readable code; and at least one processor configured to invoke the computer-readable code to perform the steps provided in the first aspect.

[0016] A sixth aspect provides a fault mode prediction apparatus, comprising: at least one memory configured to store computer-readable code; and at least one processor configured to invoke the computer-readable code to perform the steps provided in the second aspect.

[0017] A seventh aspect is a computer-readable medium storing computer-readable instructions that, when executed by a processor, cause the processor to perform the steps provided in the first or second aspect.

[0018] In this model generation process, a topic model is trained using information theory methods. Domain experts can easily incorporate domain knowledge into the topics and apply it to the mutual information calculation process to generate the desired topics representing failure modes. When a new failure occurs, the trained failure mode prediction model can be used for failure diagnosis, providing solutions or similar cases. No assumptions are made regarding the distribution of keywords and topics, thus making it applicable to a wider range of corpora. Unlike ordinary topic models, the failure mode prediction model trained using the method of this invention can perform tasks specific to a particular domain.

[0019] Optionally, when calculating the mutual information, each first keyword has a first factor, and each second keyword has a second factor. The second factor of each second keyword in a set of first keywords is smaller than the first factor of each first keyword in that set, and the larger the first factor, the greater the mutual information between the set of first keywords with added second keywords and the relevant target fault mode. In this way, the influence of domain knowledge in calculating mutual information can be controlled through factors. During training, the target topic can learn features highly correlated with the fault mode, thus becoming more representative of the target fault mode. Here, each set of first keywords is equivalent to an anchor point set in its relevant target fault mode, so that the actually generated topic conforms to the preset fault mode. Simultaneously, the first keyword also attracts more co-occurring second keywords to join the same fault mode, thereby maximizing the mutual information. These second keywords supplement the description of the features of the target fault mode.

[0020] For any of the above aspects, optionally, when obtaining the m groups of first keywords during the model generation process, they can be obtained from historical failure cases in the specific domain that are labeled with their respective target failure modes; or they can be input by experts in the specific domain. Domain experts can easily interact with the model, inputting domain knowledge into the model and controlling the model to generate the desired topics, resulting in a highly interpretable model.

[0021] Alternatively, for any of the above aspects, a trained failure mode prediction model can be used to label the target failure mode for each historical failure case. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the structure of the model generation device provided in an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the fault mode prediction device provided in an embodiment of the present invention.

[0024] Figure 3 A flowchart of a model generation method provided in an embodiment of the present invention.

[0025] Figure 4 A flowchart of a fault mode prediction method provided in an embodiment of the present invention.

[0026] Figure 5 This is a flowchart of an optional embodiment of the present invention.

[0027] List of reference numerals in the attached diagram:

[0028] 11: Model generation device; 12: Failure mode prediction device; 20: Failure mode prediction program

[0029] 111: At least one memory module; 112: At least one processor; 113: Communication module.

[0030] 21: Model Generation Program 211: Acquisition Module 212: Training Module

[0031] 213: Output module 214: Marking module

[0032] 121: At least one memory module; 122: At least one processor; 123: Communication module.

[0033] 22: Fault Mode Prediction Model Program 221: Receiving Module 222: Running Module

[0034] 223: Output Module

[0035] 300: Model generation method S301~S306: Steps of method 300

[0036] 400: Fault Mode Prediction Method S401~S403: Steps of Method 400

[0037] 51: Historical Failure Cases 52: Domain Devices 53: Domain Knowledge

[0038] 54: Experts define keywords for target failure modes

[0039] 55: Experts label the failure modes of historical failure cases.

[0040] 56: Expert-initiated model training process 57: Fault mode prediction model

[0041] 58: Update the keywords for the target failure mode.

[0042] 59: Output a list of target fault modes and keywords for each target fault mode.

[0043] 60: Historical Fault Cases 61: Fault Diagnosis Process Detailed Implementation

[0044] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed merely to enable those skilled in the art to better understand and implement the subject matter described herein, and are not intended to limit the scope, applicability, or examples set forth in the claims. The function and arrangement of the elements discussed may be changed without departing from the scope of the embodiments of the invention. Various processes or components may be omitted, substituted, or added as needed in the various examples. For example, the described methods may be performed in a different order than described, and steps may be added, omitted, or combined. Furthermore, features described in some examples may be combined in other examples.

[0045] As used herein, the term "comprising" and its variations are open terms meaning "including but not limited to". The term "based on" means "at least partially based on". The terms "one embodiment" and "an embodiment" mean "at least one embodiment". The term "another embodiment" means "at least one other embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other definitions, whether explicit or implicit, may be included below. Unless explicitly indicated by the context, the definition of a term shall remain consistent throughout the specification.

[0046] This invention provides a model generation method and apparatus for generating a fault mode prediction model; a fault mode prediction method and apparatus for performing fault prediction using the fault mode prediction model; and a computer-readable medium. During model generation, a topic model is trained using information theory methods. Domain experts can conveniently incorporate domain knowledge into the topics to generate desired topics representing fault modes. When a new fault occurs, the trained fault mode prediction model can be used for fault diagnosis, providing solutions or similar cases.

[0047] The information theory method used in the embodiments of the present invention will be introduced below.

[0048] Let X be a discrete random variable taking a finite number of possible values. Let H(X) denote the entropy of X. Then the mutual information of two random variables X1 and X2 can be defined as:

[0049] I(X1,X2)=H(X1)+H(X2)-H(X1,X2)…(1)

[0050] In information theory, mutual information can represent the correlation between variables. In particular, if X1 is a word and X2 is a topic, then the mutual information between X1 and X2 can be used to measure the degree of matching between the word and the topic; the greater the mutual information, the higher the degree of matching.

[0051] Furthermore, let X GLet X be a set of discrete random variables, then X G Multivariate mutual information can be defined as:

[0052] I(X G )=∑ i∈G H(X i )-H(X G (2)

[0053] The random variable X can be calculated using the following formula (3). G Joint mutual information based on another variable Y:

[0054] I(X G Y) = I(X) G )-I(X G |Y)…(3)

[0055] So, how can we use mutual information from information theory to generate the fault mode prediction model in this invention and perform fault mode prediction?

[0056] First, find topics that match a set of keywords.

[0057] In formula (3) above, X can be used. G Let Y represent a set of keywords, and let Y represent the topic we want to learn. Thus, when the joint mutual information I(X) is... G When the match between keywords X and topic Y is maximized, the set of keywords X that best matches topic Y has been found. G Due to conditional mutual information I(X) G |Y) takes a non-negative value when I(X) G When |Y) is 0, I(X) G The maximum value of Y means that Y can explain all the values ​​in X. G Dependencies in the data.

[0058] When keywords come from a set of documents, that set of documents may cover multiple topics. To obtain the different topics within this set of documents, we need to find a list of topics Y1, Y2, ... Y... m Each topic has its own set of keywords X G1 ,X G2 …X Gm Here, the keywords are restricted to different topics; that is, a keyword can only belong to one topic and cannot belong to other topics. We want to maximize the sum of the following mutual information:

[0059]

[0060] Based on the calculations above, we obtain a list of topics and keywords for each topic. Here, each topic represents a failure mode, and the keywords for each topic can be used to describe the characteristics of that failure mode.

[0061] However, such calculations alone still yield random topics. To enable the model to generate topics related to specific failure modes, in this embodiment of the invention, domain knowledge is further incorporated into the model and applied to the mutual information calculation process.

[0062] Furthermore, domain knowledge is incorporated into the model and used for mutual information calculation. This includes, but is not limited to, the following two optional implementation methods:

[0063] Method 1

[0064] Domain experts provide a list of keywords based on their domain knowledge. This list of keywords describes the characteristics of a failure mode. By substituting this list of keywords into the above calculation process, a model can be trained that can generate topics about the failure mode.

[0065] Method 2: Provide documents related to the failure mode (e.g., historical failure cases), and the documents will automatically generate a list of keywords. Then, use the generated list of keywords to train the model, and the resulting model will be able to generate topics about the failure mode.

[0066] The training process described above is a semi-supervised, interactive process. Domain experts can freely change the list of keywords and retrain the model. Furthermore, during model training, when calculating mutual information, keywords can be assigned a factor α (α > 1), and different keywords can have different factor values. In this way, the influence of domain knowledge in calculating mutual information can be controlled through factor α. During training, the target topic can learn features highly correlated with the failure mode, thus becoming more representative of the target failure mode.

[0067] The trained failure mode prediction model can serve various users in a specific domain. Domain experts can easily examine the keywords generated by the model for each topic and interpret the failure modes described by each topic. If a topic does not match the expected failure mode, the domain expert can modify the pre-set list of keywords to constrain the model and observe the changes. The resulting failure mode prediction model can help users understand what kind of failure is occurring and provide solutions for the same failure mode or similar cases for reference.

[0068] This invention makes no assumptions about the distribution of keywords and topics, thus it is applicable to a wider range of corpora. Unlike ordinary topic models, the fault mode prediction model trained using the method of this invention can perform tasks in a specific domain. Furthermore, domain experts can easily interact with the model, inputting domain knowledge and controlling the model to generate desired topics, resulting in a highly interpretable model.

[0069] The model generation device 11 and the fault mode prediction device 12 provided in the embodiments of the present invention will be described below.

[0070] The model generation apparatus 11 provided in this embodiment of the invention can be implemented as a network of computer processors to execute the model generation method 300 in this embodiment of the invention. The model generation apparatus 11 can also be as follows: Figure 1 The illustrated single-unit computer includes at least one memory 111, which includes a computer-readable medium, such as random access memory (RAM). The device 11 also includes at least one processor 112 coupled to the at least one memory 111. Computer-executable instructions are stored in the at least one memory 111 and, when executed by the at least one processor 112, can cause the at least one processor 112 to perform the steps described herein. The at least one processor 112 may include a microprocessor, an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a central processing unit (CPU), a graphics processing unit (GPU), a state machine, etc. Embodiments of computer-readable media include, but are not limited to, floppy disks, CD-ROMs, magnetic disks, memory chips, ROMs, RAMs, ASICs, configured processors, all-optical media, all magnetic tapes or other magnetic media, or any other media from which a computer processor can read instructions. Furthermore, various other forms of computer-readable media can send or carry instructions to a computer, including routers, private or public networks, or other wired and wireless transmission devices or channels. Instructions may include code in any computer programming language, including C, C++, C++, Visual Basic, Java, and JavaScript.

[0071] When executed by at least one processor 112 Figure 1 At least one memory 111 shown may contain a model generation program 21, causing at least one processor 112 to execute the model generation method 300 described in this embodiment of the invention. The model generation program 21 may include:

[0072] - The acquisition module 211 is configured to: acquire m target fault modes and a set of first keywords corresponding to each target fault mode, where m is a positive integer, each set of first keywords is based on knowledge of a specific domain and is used to describe the characteristics of the corresponding target fault mode; acquire at least one historical fault case in a specific domain; and acquire at least one second keyword from at least one historical fault case.

[0073] - Training module 212 is configured to add each second keyword to m groups of first keywords during the training of the fault mode prediction model. Each second keyword can only be added to one group of first keywords to maximize the sum of m mutual information. Each mutual information corresponds to a target fault mode and is the mutual information between the first keyword corresponding to the target fault mode with the added second keyword and the target fault mode.

[0074] - Output module 213 is configured to output m target fault modes, and for each target fault mode, output a set of keywords to describe the characteristics of that target fault mode.

[0075] Optionally, when calculating mutual information, the training module 212 provides that each first keyword has a first factor and each second keyword has a second factor; the second factor of each second keyword in a set of first keywords is less than the first factor of each first keyword in the set, and the larger the first factor, the greater the mutual information between the set of first keywords with added second keywords and the relevant target fault mode.

[0076] Optionally, when the acquisition module 211 acquires the first keyword of the m groups, it is specifically configured to acquire it from historical failure cases in a specific field that are labeled with the target failure mode, or to be input by an expert in a specific field.

[0077] Optionally, it may also include a labeling module 214, configured to: label the target failure mode to which each historical failure case belongs using a trained failure mode prediction model.

[0078] Optionally, the device 11 may further include a communication module 113, which is connected to at least one processor 112 and at least one memory 111 via a bus, for the device 11 to communicate with external devices.

[0079] It should be mentioned that embodiments of the present invention may include those having different... Figure 2 The apparatus with the architecture shown is merely exemplary and is used to explain the method 300 provided in the embodiments of the present invention.

[0080] In addition, the above modules can also be regarded as functional modules implemented by hardware, used to implement various functions involved in the model generation device 11 when executing the model generation method. For example, the control logic of each process involved in the method can be pre-burned into a chip such as a field-programmable gate array (FPGA) or a complex programmable logic device (CPLD), and these chips or devices can execute the functions of the above modules. The specific implementation method can be determined according to engineering practice.

[0081] The fault mode prediction device 12 provided in this embodiment of the invention can be implemented as a network of computer processors to execute the fault mode prediction method 400 in this embodiment of the invention. The fault mode prediction device 12 can also be as follows: Figure 2 The illustrated single-unit computer includes at least one memory 121, which includes a computer-readable medium, such as random access memory (RAM). The device 12 also includes at least one processor 122 coupled to the at least one memory 121. Computer-executable instructions are stored in the at least one memory 121 and, when executed by the at least one processor 122, can cause the at least one processor 122 to perform the steps described herein. The at least one processor 122 may include a microprocessor, an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a central processing unit (CPU), a graphics processing unit (GPU), a state machine, etc. Embodiments of computer-readable media include, but are not limited to, floppy disks, CD-ROMs, magnetic disks, memory chips, ROMs, RAMs, ASICs, configured processors, all-optical media, all magnetic tapes or other magnetic media, or any other media from which a computer processor can read instructions. Furthermore, various other forms of computer-readable media can send or carry instructions to a computer, including routers, private or public networks, or other wired and wireless transmission devices or channels. Instructions may include code in any computer programming language, including C, C++, C++, Visual Basic, Java, and JavaScript.

[0082] When executed by at least one processor 122 Figure 2 At least one memory 121 shown may contain a fault mode prediction program 22, causing at least one processor 122 to execute the fault mode prediction method 400 described in this embodiment of the invention. The fault mode prediction program 22 may include:

[0083] - Receiver module 221 is configured to receive new fault cases in a specific domain;

[0084] - Module 222 is configured to input new failure cases into a failure mode prediction model;

[0085] - Output module 223 is configured to output the target failure mode to which the new failure case obtained from the failure mode prediction model is located.

[0086] The fault mode prediction model can be trained by the model generation device 11.

[0087] Optionally, the device 11 may further include a communication module 123, which is connected to at least one processor 122 and at least one memory 121 via a bus, for the device 12 to communicate with external devices.

[0088] It should be mentioned that embodiments of the present invention may include those having different... Figure 2 The apparatus with the architecture shown is merely exemplary and is used to explain the method 400 provided in the embodiments of the present invention.

[0089] In addition, the above modules can also be regarded as functional modules implemented by hardware, used to implement various functions involved in the fault mode prediction device 12 when executing the fault mode prediction method. For example, the control logic of each process involved in the method can be pre-burned into a chip such as a field-programmable gate array (FPGA) or a complex programmable logic device (CPLD), and these chips or devices can execute the functions of the above modules. The specific implementation method can be determined according to engineering practice.

[0090] like Figure 4 As shown, an exemplary method 300 according to an embodiment of the present invention includes the following steps:

[0091] -S301: Obtain m target fault modes and a set of first keywords corresponding to each target fault mode, where m is a positive integer, and each set of first keywords is based on knowledge of a specific domain and is used to describe the characteristics of the corresponding target fault mode.

[0092] A specific field can be the industrial field, the agricultural field, or a subfield of a larger field, such as the CNC machine tool technology field in the industrial field.

[0093] Taking the field of CNC machine tool technology as an example, method 300 can be used to train a fault mode prediction model to predict the fault mode of a CNC machine tool system, so as to realize the fault monitoring, prediction and maintenance of the CNC machine tool system.

[0094] Different target fault modes possess different characteristics, which distinguish them from each other, and the corresponding fault solutions typically differ as well. In step S301, there are a total of m target fault modes, each with a set of keywords to describe its characteristics. These keywords are based on knowledge of the specific domain and can describe the characteristics of the target fault mode. Here, this set of keywords is called the "first keyword" to distinguish it from keywords added later during subsequent model training.

[0095] Optionally, data can be obtained from historical failure cases in a specific domain that are labeled with the target failure mode. These historical failure cases can be manually labeled by experts in that specific domain based on their knowledge of that domain. In this embodiment of the invention, experts do not need to label all historical failure cases, but only label a small number of historical failure cases with the target failure mode, and train the model based on these few labeled cases. For each target failure mode, several words that contribute the most to the mutual information can be found from the cases already labeled with that failure mode, and these words can be used as the first keyword of that failure mode.

[0096] Alternatively, each primary keyword can be entered by an expert in that specific field, who will mark the target failure mode to which these keywords belong when entering them.

[0097] -S302: Obtain at least one historical failure case in a specific domain. In the field of natural language processing technology, these historical failure cases are processed as documents.

[0098] -S303: Obtain at least one second keyword from at least one historical failure case; wherein the method for obtaining the second keyword is similar to the aforementioned method for obtaining the first keyword from historical failure cases labeled with the target failure mode, and will not be repeated here. Here, in order to distinguish it from the keyword obtained in step S301, the keyword obtained from historical failure cases without the target failure mode is called the "second keyword". These keywords will be attempted to be added to the aforementioned sets of keywords in subsequent steps to calculate the mutual information with the target failure mode.

[0099] -S304: During the training of the fault mode prediction model, each second keyword is added to m groups of first keywords. Each second keyword can only be added to one group of first keywords to maximize the sum of the m mutual information. Each mutual information corresponds to a target fault mode, representing the mutual information between the first keyword and the target fault mode after the second keyword has been added (as shown in Formula 4 above). Here, each group of first keywords acts as an anchor point in its related target fault modes, ensuring that the generated topics conform to the preset fault modes. Simultaneously, the first keywords attract more co-occurring second keywords to join the same fault mode, maximizing the mutual information. These second keywords supplement the description of the target fault mode's features. Compared to previous general-domain topic modeling processes, the model training process incorporates the constraint of maximizing the sum of mutual information, guiding the training process and ensuring that the features described by the keywords corresponding to the final target fault modes generated by the model are closer to the features described by the first keywords.

[0100] Optionally, when calculating mutual information, each first keyword has a first factor, and each second keyword has a second factor. The second factor of each second keyword in a set of first keywords is less than the first factor of each first keyword in that set, and the larger the first factor, the greater the mutual information between the set of first keywords with added second keywords and the relevant target fault mode. For example, the first factor of the first keyword can be set to a number greater than 1, while the second factor of the second keyword is set to 1 by default. As mentioned earlier, the influence of domain knowledge in calculating mutual information can be controlled through factors. During training, the target topic can learn features highly correlated with the fault mode, thus better representing the target fault mode. The fault mode prediction model can be trained through step S304. Because specific domain knowledge is incorporated into the model training process and the desired topic generated by the model is controlled, the trained model better reflects the characteristics of that specific domain, and the fault mode prediction results are more accurate.

[0101] Optionally, in step S305, the model can be controlled to output m target fault modes, and for each target fault mode, a set of keywords describing the characteristics of that target fault mode can be output. Optionally, in step S306, the trained fault mode prediction model can be used to label the target fault mode to which each historical fault case belongs. Optionally, experts in the specific domain can examine the target fault modes and corresponding keywords output in step S305, and examine the historical fault cases labeled with the target fault modes output in step S306. If inaccurate results are found, the first keyword output in step S301 or the labeling of a small number of historical fault cases can be adjusted based on knowledge of the specific domain.

[0102] Below, refer to Figure 4 This describes the fault mode prediction method 400 provided in an embodiment of the present invention. For example... Figure 4 As shown, the method may include the following steps:

[0103] -S401: Receive new fault cases in a specific domain;

[0104] -S402: Input new failure cases into the failure mode prediction model trained by method 300;

[0105] -S403: Outputs the target fault mode to which the new fault cases obtained from running the fault mode prediction model belong.

[0106] Figure 5 A flowchart of an optional embodiment of the present invention is shown, as follows: Figure 5 As shown, in step 55, an expert 52 in a specific domain, such as CNC machine tools, labels a small number of historical fault cases 51 with their respective target fault modes; or in step 54, the expert 52, based on knowledge 53 in the CNC machine tool domain, defines a set of keywords (i.e., the aforementioned m target fault modes) for some target fault modes (i.e., the aforementioned m first keywords). In step 56, the expert 52 initiates a model training process. The historical fault cases 51 and each target fault mode labeled with keywords are input into the fault mode prediction model 57 for model training. After training, on the one hand, a list of target fault modes and keywords 59 for each fault mode can be output; on the other hand, the historical fault cases 51 can be labeled with their respective target fault modes, resulting in historical fault cases 60 labeled with target fault modes. In step 62, the expert 52 can check results 59 and 60. If any incorrect target fault modes are found, steps 54 or 55 can be adjusted, domain knowledge can be re-inputted, and the model can be trained. In step 58, results 59 can be used to update the keywords defined by the expert 52 for various target fault modes. The trained model 57 can be used in the fault diagnosis process 61 of CNC machine tools.

[0107] Furthermore, embodiments of the present invention also provide a computer-readable medium storing computer-readable instructions. When executed by a processor, the computer-readable instructions cause the processor to perform the aforementioned fault mode prediction method. Embodiments of the computer-readable medium include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Optionally, the computer-readable instructions can be downloaded from a server computer or the cloud via a communication network.

[0108] It should be noted that not all steps and modules in the above processes and system structure diagrams are mandatory; some steps or modules can be omitted as needed. The execution order of each step is not fixed and can be adjusted as required. The system structure described in the above embodiments can be a physical structure or a logical structure. That is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or they may be jointly implemented by certain components in multiple independent devices.

[0109] In summary, the embodiments of the present invention provide a model generation method, apparatus, fault mode prediction method and apparatus, and a computer-readable medium. They do not make any assumptions about the distribution of keywords and topics, thus making them applicable to a wider range of corpora. Unlike ordinary topic models, the fault mode prediction model trained using the method of the embodiments of the present invention can perform tasks in a specific domain. Furthermore, domain experts can easily interact with the model, inputting domain knowledge and controlling the model to generate desired topics, resulting in a highly translatable model.

Claims

1. A model generation method (300) for generating a failure mode prediction model, comprising: - Obtain (S301) m target fault modes and a set of first keywords corresponding to each target fault mode, where m is a positive integer, and each set of first keywords is based on knowledge of a specific domain and is used to describe the characteristics of the corresponding target fault mode. - Obtain at least one historical failure case in the specific domain described in (S302); - Obtain at least one second keyword from the at least one historical fault case (S303); - During the training of the (S304) fault mode prediction model, each second keyword is added to m groups of first keywords. Each second keyword can only be added to one group of first keywords to maximize the sum of m mutual information. Each mutual information corresponds to a target fault mode, which is the mutual information between the first keyword corresponding to the target fault mode with the added second keyword and the target fault mode. - Output (S305) m target fault modes, and for each target fault mode, output a set of output keywords to describe the characteristics of the target fault mode.

2. The method as described in claim 1, characterized in that, - When calculating the mutual information, each first keyword has a first factor and each second keyword has a second factor; the second factor of each second keyword in a set of first keywords is smaller than the first factor of each first keyword in the set, and the larger the first factor, the greater the mutual information between the set of first keywords with the second keyword and the relevant target fault mode.

3. The method as described in claim 1, characterized in that, Get the first keyword of group m (S301), including: - Obtained from historical failure cases in the specific domain that are labeled with their respective target failure modes; or - Input by experts in the specific field.

4. The method of claim 1, further comprising: - Use the trained failure mode prediction model to label the target failure mode to which each historical failure case belongs (S306).

5. A failure mode prediction method (400), comprising: - Receive (S401) new fault cases in a specific area; - Input the new failure case into a failure mode prediction model (S402); - Output (S403) the target failure mode to which the new failure case belongs, obtained from the failure mode prediction model. The fault mode prediction model is trained using the method described in any one of claims 1 to 4.

6. A model generation apparatus (11) for generating a failure mode prediction model, comprising: - The acquisition module (211) is configured as follows: - Obtain m target fault modes and a set of first keywords corresponding to each target fault mode, where m is a positive integer, and each set of first keywords is based on knowledge of a specific domain and is used to describe the features of the corresponding target fault mode. - Obtain at least one historical failure case in the specific domain; - Obtain at least one second keyword from the at least one historical failure case; - The training module (212) is configured to add each second keyword to m groups of first keywords during the training of the fault mode prediction model. Each second keyword can only be added to one group of first keywords so that the sum of m mutual information is maximized. Each mutual information corresponds to a target fault mode and is the mutual information between the first keyword corresponding to the target fault mode with the added second keyword and the target fault mode. - The output module (213) is configured to output m target fault modes and, for each target fault mode, output a set of output keywords to describe the characteristics of that target fault mode.

7. The apparatus as claimed in claim 6, characterized in that, - When the training module (212) calculates the mutual information, each first keyword has a first factor and each second keyword has a second factor; the second factor of each second keyword in a set of first keywords is smaller than the first factor of each first keyword in the set, and the larger the first factor, the greater the mutual information between the set of first keywords with added second keywords and the related target fault mode.

8. The apparatus as claimed in claim 6, characterized in that, When the acquisition module (211) acquires the first keywords of m groups, it is specifically configured as follows: - Obtained from historical failure cases in the specific domain that are labeled with their respective target failure modes; or - Input by experts in the specific field.

9. The apparatus of claim 6, further comprising: - The labeling module (214) is configured to: use the trained failure mode prediction model to label the target failure mode to which each historical failure case belongs.

10. A fault mode prediction device (12), comprising: - The receiving module (221) is configured to receive new fault cases in a specific domain; - The running module (222) is configured to input the new failure case into a failure mode prediction model; - Output module (223) is configured to output the target failure mode to which the new failure case obtained from the failure mode prediction model belongs; The fault mode prediction model is trained using the method described in any one of claims 1 to 4.

11. A model generation apparatus (11), characterized in that, include: - At least one memory (111) is configured to store computer-readable code; - At least one processor (112) is configured to invoke the computer-readable code to perform the method as described in any one of claims 1 to 4.

12. A fault mode prediction device (12), characterized in that, include: - At least one memory (121) is configured to store computer-readable code; - At least one processor (122) is configured to invoke the computer-readable code to perform the method as described in claim 5.

13. A computer-readable medium, characterized in that, The computer-readable medium stores computer-readable instructions that, when executed by a processor, cause the processor to perform the method as described in any one of claims 1 to 5.

Images