Construction method and diagnosis method of data set for fault diagnosis of clean energy equipment

Abstract

The invention relates to the field of machinery and the field of natural language processing, in particular to a construction method of a data set for fault diagnosis of clean energy equipment and a diagnosis method. The construction method comprises the following steps: S1, acquiring an original corpus; s2, analyzing and preprocessing the original corpus, and converting the original corpus into sentence-level embedded representation; s3, defining a fault type and marking the fault type; s4, calculating the centroid characteristics of all fault text instances in each fault type, calculating the similarity between each fault text instance and the centroid characteristics of the fault text instance, and selecting representative fault text instances according to the similarity and storing the representative fault text instances in a memory library; and S5, enhancing the data in the memory bank, expanding the corpus scale, guiding to generate additional supplementary samples for fault types with small sample sizes, and taking the data in the memory bank as a data set for fault diagnosis of the clean energy equipment. According to the method, a data set which is rich in fault type and wide in coverage for the clean energy equipment can be constructed, and the fault diagnosis efficiency and accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the fields of mechanics and natural language processing, and in particular to a method for constructing a data set for clean energy equipment fault diagnosis and a method for diagnosing the fault. Background Art

[0002] As the global awareness of sustainable development and environmental protection increases, the development and utilization of clean energy has received unprecedented attention. Clean energy equipment, such as wind turbines, solar panels and hydroelectric power generation equipment, not only provide us with environmentally friendly energy, but also the stability and efficiency of their operation are crucial. However, these equipment will inevitably have various faults during long-term operation, such as damage, leakage, and reduced efficiency. If these faults cannot be diagnosed and handled in a timely and accurate manner, they will greatly reduce the operating efficiency and life of the equipment, and may even lead to serious safety accidents.

[0003] At present, the fault diagnosis of clean energy equipment mainly relies on the experience judgment and on-site inspection of professional technicians. This method is not only time-consuming and labor-intensive, but the accuracy and efficiency of diagnosis depend largely on the professional level and experience of technicians. With the development of artificial intelligence and big data technology, the use of these technologies for fault diagnosis has gradually become a research hotspot. In particular, natural language processing (NLP) technology has shown great potential in processing and analyzing large amounts of text data. By analyzing and learning large amounts of text data such as historical fault reports and maintenance records, the accuracy and efficiency of fault diagnosis can be effectively improved.

[0004] However, since the clean energy field is relatively new, there are relatively few related fault reports and text materials, and the formats are different, which poses a huge challenge to building an effective fault diagnosis model. At present, the main methods to alleviate data scarcity include synonym replacement, sentence reorganization, remote supervision, etc. Text translation methods such as synonym replacement and sentence reorganization may cause changes in the original semantics, especially in texts in complex or professional fields. And this method does not consider the context of the entire sentence or paragraph, resulting in the generated new data being contextually inappropriate or inconsistent. In addition, these methods generate new data within a limited range, which limits the improvement of data set diversity. Remote supervision methods use external knowledge bases to automatically annotate data, but may introduce incorrect labels and increase the noise of the data set. And without sufficient context, remote supervision may have difficulty accurately identifying or eliminating entity ambiguity.

[0005] Therefore, how to extract effective information from limited, unstructured text data and construct a high-quality dataset for clean energy equipment fault diagnosis has become an urgent problem to be solved. Summary of the invention

[0006] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a method for constructing a data set for clean energy equipment fault diagnosis, which can construct a data set with rich fault types and wide coverage for clean energy equipment, solve the problem of scarce corpus, and thus improve the efficiency and accuracy of fault diagnosis.

[0007] In order to solve the above technical problems, the technical solution of the present invention is: a method for constructing a data set for clean energy equipment fault diagnosis, comprising:

[0008] S1, collects the fault texts of clean energy equipment as the original corpus;

[0009] S2, analyze and preprocess each original corpus and convert it into sentence-level embedding representation;

[0010] S3, define the fault type and annotate the embedding representations at each sentence level to form a fault text instance. The annotation content includes the entity of the faulty device or component and the corresponding fault type relationship;

[0011] S4, calculate the centroid features of all fault text instances in each fault type, calculate the similarity between each fault text instance and the centroid feature of its class, select representative fault text instances based on the similarity and save them in the memory library, when there is a new type of fault, calculate the centroid features and similarity, select representative fault text instances to update the memory library;

[0012] S5, enhance the data in the memory bank and expand the corpus size. Then, for the fault types with fewer samples in the memory bank, guide the generation of additional supplementary samples. Finally, use the data in the memory bank as a data set for clean energy equipment fault diagnosis.

[0013] Furthermore, S2 specifically includes:

[0014] S21, preprocessing text: removing noise, standardizing text, and restoring words to their dictionary form using part-of-speech restoration and part-of-speech tagging;

[0015] S22, corpus analysis: Analyze the entities and corresponding relationship types in the corpus, reorganize the long text into multiple short texts, each short text contains at least one entity and one fault type;

[0016] S23, generate sentence-level representation: use the Transformer model to generate sentence-level embedding representations to capture richer contextual information;

[0017] x k =BERT([CLS],t t1 ,t t2 ,t t3…,t tn ,[SEP])

[0018] Where [CLS] is the start tag, [SEP] is the end tag, and each long text is reorganized into a short text set T = {t t1 ,t t2 ,t t3 …,t tn}.

[0019] Furthermore, in S3, the fault types defined include abnormal temperature, damage, leakage, wear, aging, reduced efficiency, defects, abnormal pressure, corrosion, cracks, sticking, scaling, poor contact, low flow, looseness, loss, software failure vibration, blockage, rust, deformation, and structural fatigue;

[0020] Furthermore, the formula for calculating the centroid features of all fault text instances in each fault type is:

[0021]

[0022] where c j is the average feature vector of category j, i.e., the centroid feature; n j is the total number of samples in category j; x k is the kth sample in category j, and f() represents the feature representation of each text generated using the pre-trained model BERT;

[0023] The formula for calculating the similarity between each fault text instance and the centroid feature of its class is:

[0024]

[0025] where x i represents the i-th sample in category j, and rep(x) represents the normalized representation of x.

[0026] Furthermore, in S5, for the fault types with fewer samples in the memory library, a prompt template is used to guide the language model to generate additional supplementary samples. The format of the prompt template is:

[0027] Input: Relation Type: <fault type label>

[0028] Output: head entity: <subject>, tail entity: <object>

[0029] Context: <text1><text2><text3>…<textn>.

[0030] Furthermore, in S5, before enhancing the data in the memory library, [S] and [\S] are respectively inserted before and after the head entity of each fault text instance, and [O] and [\O] are respectively inserted before and after the tail entity to mark the starting positions of the head entity and the tail entity.

[0031] The present invention also relates to a clean energy equipment fault diagnosis method, the method comprising:

[0032] The first step is to construct a clean energy equipment fault diagnosis dataset using a clean energy equipment fault diagnosis dataset construction method;

[0033] In the second step, similarity calculation is used to select samples that are highly relevant to known tasks from the clean energy equipment fault diagnosis data set to expand the training data;

[0034] The third step is to use the expanded data to train the fault diagnosis model;

[0035] The fourth step is to use the trained fault diagnosis model to diagnose the fault data.

[0036] Furthermore, in the third step, during the training process, the total loss function is used for optimization, and the formula is:

[0037] L total =L ce +L1+L2;

[0038] Among them, L total represents the total loss function; L ce Represents the loss function used for relation classification tasks; L1 represents the positive and negative boundary loss function; L2 represents the negative class penalty boundary loss function;

[0039]

[0040] D represents the training data set, R represents the set of all possible relationship types in the task, δ(y i ,r j ) represents the Kronecker function. When the actual relationship y i Equal to the predicted relationship r j Returns 1 if it is, otherwise returns 0. θ (x i ),r l ) is a function Indicates using cosine similarity to calculate text entity pair x i and the relationship j The similarity between

[0041]

[0042] m1 is the boundary value, which is used to ensure that the score of the correct relationship is higher than the minimum interval of the score of the wrong relationship, t i For a given sample x i The correct relation index of

[0043]

[0044] s i is divided by t i The index of the wrong label with the highest external similarity, and m2 is the boundary value.

[0045] After adopting the above technical solution, the method for constructing a data set for clean energy equipment fault diagnosis in the invention makes full use of natural language processing technology. By collecting, analyzing and processing document resources such as industrial fault reports publicly available on the Internet, a data set with rich fault types and wide coverage for clean energy equipment is constructed. This data set can not only improve the efficiency and accuracy of fault diagnosis, but also has important theoretical and practical significance for promoting the intelligent maintenance and management of clean energy equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flowchart of a method for constructing a data set for clean energy equipment fault diagnosis according to the present invention;

[0047] Figure 2 This is a schematic diagram of the original corpus sample;

[0048] Figure 3 It is a structural schematic diagram of the clean energy equipment fault diagnosis method of the present invention. DETAILED DESCRIPTION

[0049] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0050] Embodiment 1

[0051] like Figure 1 , Figure 2 As shown, a method for constructing a data set for clean energy equipment fault diagnosis includes:

[0052] S1, collects the fault texts of clean energy equipment as the original corpus;

[0053] Specifically: collect public industrial failure reports on the Internet and summarize the failure texts of common clean energy equipment as original corpus;

[0054] S2, analyze and preprocess each original corpus and convert it into sentence-level embedding representation;

[0055] S3, define the fault type and annotate the embedding representations at each sentence level to form a fault text instance. The annotation content includes the entity of the faulty device or component and the corresponding fault type relationship;

[0056] S4, calculate the centroid features of all fault text instances in each fault type, calculate the similarity between each fault text instance and the centroid features of its class, select representative fault text instances based on the similarity and save them in the memory library, when there is a new type of fault, calculate the centroid features and similarity, select representative fault text instances to update the memory library; in general, the similarity is greater than the threshold, as a representative fault text instance, the threshold can be 0.8;

[0057] S5, first enhance the data in the memory bank and expand the corpus size, then guide the generation of additional supplementary samples for the fault types with fewer samples in the memory bank, and finally use the data in the memory bank as a data set for clean energy equipment fault diagnosis.

[0058] In this embodiment, S2 may specifically include:

[0059] S21, preprocessing text: removing noise, standardizing text, and restoring words to their dictionary form using part-of-speech restoration and part-of-speech tagging;

[0060] Among them, noise includes irrelevant symbols, incorrect formats, etc.; use NLP libraries (such as Spacy) to standardize text, such as identifying and unifying terms and abbreviations;

[0061] S22, corpus analysis: Analyze the entities and corresponding relationship types in the corpus, reorganize the long text into multiple short texts, each short text contains at least one entity and one fault type;

[0062] The corpus is analyzed in a manually supervised manner, and the content of the analysis is the entities in the corpus (entities of faulty equipment or components) and the corresponding relationship types (fault types).

[0063] S23, Generate sentence-level representation: Use the Transformer model to generate sentence-level embedding representations to capture richer contextual information and improve the performance of subsequent model training.

[0064] x k =BERT([CLS],t t1 ,t t2 ,t t3 …,t tn ,[SEP])

[0065] Where [CLS] is the start tag, [SEP] is the end tag, and each long text is reorganized into a short text set T = {tt1 ,t t2 ,t t3 …,t tn}.

[0066] The Transformer model may be BERT or other models.

[0067] In this embodiment, in S3, the defined fault types include abnormal temperature, damage, leakage, wear, aging, reduced efficiency, defects, abnormal pressure, corrosion, cracks, sticking, scaling, poor contact, low flow, looseness, loss, software failure vibration, blockage, rust, deformation, and structural fatigue.

[0068] Among them, manual labeling can be used for the initial situation when the data volume is small, and named entity recognition (NER) technology can be used to automatically detect key entities in the text, such as equipment parts, for the later situation when the data volume is large.

[0069] In this embodiment, the formula for calculating the centroid feature of all fault text instances in each fault type may be:

[0070]

[0071] where c j is the average feature vector of category j, i.e., the centroid feature; n j is the total number of samples in category j; x k is the kth sample in category j, f() represents the use of the pre-trained model BERT to generate the feature representation of each text; function f(x k ) can be defined as using the BERT model to extract each text x k The feature embedding representation of f(x k )=BERT(x k ).

[0072] The formula for calculating the similarity between each fault text instance and the centroid feature of its class can be:

[0073]

[0074] where x i represents the i-th sample in category j, and rep(x) represents the normalized representation of x obtained from the last layer of BERT. This formula attempts to ensure that sentences with the same entity pairs have a higher cosine similarity than sentences with different entity pairs.

[0075] In this embodiment, in S5, before enhancing the data in the memory library, [S] and [\S] may be respectively inserted before and after the head entity of each fault text instance, and [O] and [\O] may be respectively inserted before and after the tail entity to mark the starting positions of the head entity and the tail entity.

[0076] Among them, in S5, the large language model can be used to enhance the data in the memory bank. The format of the fault text amplification template for enhancing the data is:

[0077] Text 1:…[S]head entity[\S]…[O]tail entity[\O]

[0078] Text 2:…[S]head entity[\S]…[O]tail entity[\O]

[0079] Text 3:…[S]head entity[\S]…[O]tail entity[\O]

[0080] …

[0081] Text N:...[S]head entity[\S]...[O]tail entity[\O]

[0082] The large prediction model can be GPT2, GPT3.5, or T5, etc.

[0083] The above is an explanation of several structured relational statements. [S] and [O] mark the starting position of the head and tail entities respectively, and [\S] and [\O] mark the ending position of the head and tail entities respectively. There must be a fault relationship between the entity pairs, such as the clean energy industrial equipment, which can be expanded into more sentences with similar themes and fields, the same sub-object format, but diversified entities.

[0084] Specifically, the contextual learning ability of the large language model enables it to generate logically coherent and contextually consistent text when performing data augmentation or annotation. Using this ability, the template can be augmented through the fault text to guide the model to not only focus on the key information in the text (such as entities and relationships), but also maintain the coherence and logic of the overall text. This method is particularly suitable for processing complex text data and improving the accuracy and effectiveness of information extraction.

[0085] In this embodiment, in S5, for the fault types with a small number of samples in the memory library, a prompt template is used to guide the large language model to generate additional supplementary samples. The format of the prompt template is:

[0086] Input: Relation Type: <fault type label>

[0087] Output: head entity: <subject>, tail entity: <object>

[0088] Context: <text1><text2><text3>...<textn>

[0089] The large prediction model can be GPT2, GPT3.5, or T5, etc.

[0090] Specifically, by designing specialized prompt templates, the large language model can further complete the annotation of entities and relationships. For example, the prompt template can be used to directly guide the model to generate annotated text, which not only improves the efficiency of annotation, but also ensures the consistency and accuracy of the data. This method is particularly effective in scenarios that require a large amount of annotated data, greatly reducing the workload of manual annotation.

[0091] Finally, we conducted experiments on a relation extraction model (ML2D, a multi-level attention table filling relation extraction model) suitable for the clean energy industry equipment fault diagnosis field. The clean energy industry contains a large number of nested entities. The dataset obtained in this embodiment achieved a 98.78% named entity recognition rate on the ML2D model based on the inclusion of such entities, proving the availability of the dataset. Through this method, we successfully constructed a dataset for clean energy equipment fault diagnosis, given the scarcity of corpus in the clean energy field and the absence of related work.

[0092] Embodiment 2

[0093] like Figure 3 As shown, a clean energy equipment fault diagnosis method, the method comprising:

[0094] The first step is to construct a clean energy equipment fault diagnosis data set using the method for constructing a clean energy equipment fault diagnosis data set in Example 1;

[0095] In the second step, similarity calculation is used to select samples that are highly relevant to known tasks from the clean energy equipment fault diagnosis data set to expand the training data;

[0096] The screening method in this step is the same as step S4 in Example 1.

[0097] The third step is to use the expanded data to train the fault diagnosis model;

[0098] Among them, the fault diagnosis model is a relational extraction model, which can be of many types, such as the application number CN202311858816.3, and the patent name is the compressor fault diagnosis model in the compressor fault diagnosis and feedback method based on word pair tagging and joint extraction.

[0099] The fourth step is to use the trained fault diagnosis model to diagnose the fault data.

[0100] In this embodiment, in the third step, during the training process, the total loss function is used for optimization, and the formula is:

[0101] L total =L ce +L1+L2;

[0102] Among them, L total represents the total loss function; L ce Represents the loss function used for relation classification tasks; L1 represents the positive and negative boundary loss function, which increases the score difference between the correct category and the wrong category; L2 represents the negative class penalty boundary loss function, which imposes an additional penalty on the closest but wrong label;

[0103]

[0104] D represents the training data set, R represents the set of all possible relationship types in the task, δ(y i ,r j ) represents the Kronecker function. When the actual relationship y i Equal to the predicted relationship r j Returns 1 if it is, otherwise returns 0. θ (x i ),r l ) is function g, which means using cosine similarity to calculate text entity pair x i and the relationship j The similarity between

[0105]

[0106] m1 is the boundary value, which is used to ensure that the score of the correct relationship is higher than the minimum interval of the score of the wrong relationship, and is set to 0.2. i For a given sample x i This loss function penalizes the model when the score of the wrong relation plus the margin value m1 is still higher than or close to the correct relation score.

[0107]

[0108] s i is divided by t i The index of the wrong label with the highest external similarity, m2 is the boundary value, set to 0.4.

[0109] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A method for constructing a data set for clean energy equipment fault diagnosis, characterized in that: include: S1, collects the fault texts of clean energy equipment as the original corpus; S2, analyze and preprocess each original corpus and convert it into sentence-level embedding representation; S3, define the fault type and annotate the embedding representations at each sentence level to form a fault text instance. The annotation content includes the entity of the faulty device or component and the corresponding fault type relationship; S4, calculate the centroid features of all fault text instances in each fault type, calculate the similarity between each fault text instance and the centroid feature of its class, select representative fault text instances based on the similarity and save them in the memory library, when there is a new type of fault, calculate the centroid features and similarity, select representative fault text instances to update the memory library; S5, enhance the data in the memory bank and expand the corpus size. Then, for the fault types with fewer samples in the memory bank, guide the generation of additional supplementary samples. Finally, use the data in the memory bank as a data set for clean energy equipment fault diagnosis.

2. The method for constructing a clean energy equipment fault diagnosis data set according to claim 1, characterized in that: S2 specifically includes: S21, preprocessing text: removing noise, standardizing text, and restoring words to their dictionary form using part-of-speech restoration and part-of-speech tagging; S22, corpus analysis: Analyze the entities and corresponding relationship types in the corpus, reorganize the long text into multiple short texts, each short text contains at least one entity and one fault type; S23, generate sentence-level representation: use the Transformer model to generate sentence-level embedding representations to capture richer contextual information; x k =BERT([CLS],t t1 ,t t2 ,t t3 …,t tn ,[SEP]) Where [CLS] is the start tag, [SEP] is the end tag, and each long text is reorganized into a short text set T = {t t1 ,t t2 ,t t3 …,t tn }.

3. The method for constructing a clean energy equipment fault diagnosis data set according to claim 1, characterized in that: In S3, the defined fault types include abnormal temperature, damage, leakage, wear, aging, reduced efficiency, defects, abnormal pressure, corrosion, cracks, sticking, scaling, poor contact, low flow, looseness, loss, software failure vibration, blockage, rust, deformation, and structural fatigue.

4. The method for constructing a clean energy equipment fault diagnosis data set according to claim 1, characterized in that: The formula for calculating the centroid feature of all fault text instances in each fault type is: where c j is the average feature vector of category j, i.e., the centroid feature; n j is the total number of samples in category j; x k is the kth sample in category j, and f() represents the feature representation of each text generated using the pre-trained model BERT; The formula for calculating the similarity between each fault text instance and the centroid feature of its class is: where x i represents the i-th sample in category j, and rep(x) represents the normalized representation of x.

5. The method for constructing a clean energy equipment fault diagnosis data set according to claim 1, characterized in that: In S5, for the fault types with fewer samples in the memory library, a prompt template is used to guide the language model to generate additional supplementary samples. The format of the prompt template is: Input: Relation Type: <fault type label> Output: head entity: <subject>, tail entity: <object> Context: <text1><text2><text3>…<textn>.

6. The method for constructing a clean energy equipment fault diagnosis data set according to claim 1, characterized in that: In S5, before enhancing the data in the memory library, [S] and [\S] are inserted before and after the head entity of each fault text instance, and [O] and [\O] are inserted before and after the tail entity to mark the starting positions of the head entity and the tail entity.

7. A clean energy equipment fault diagnosis method, characterized in that: Methods include: The first step is to construct a clean energy equipment fault diagnosis data set using the method for constructing a clean energy equipment fault diagnosis data set in any embodiment of claims 1 to 6; In the second step, similarity calculation is used to select samples that are highly relevant to known tasks from the clean energy equipment fault diagnosis data set to expand the training data; The third step is to use the expanded data to train the fault diagnosis model; The fourth step is to use the trained fault diagnosis model to diagnose the fault data.

8. The clean energy equipment fault diagnosis method according to claim 7, characterized in that: In the third step, during the training process, the total loss function is used for optimization, and the formula is: L total =L ce +L1+L2; Among them, L total represents the total loss function; L ce Represents the loss function used for relation classification tasks; L1 represents the positive and negative boundary loss function; L2 represents the negative class penalty boundary loss function; D represents the training data set, R represents the set of all possible relationship types in the task, δ(y i ,r j ) represents the Kronecker function. When the actual relationship y i Equal to the predicted relationship r j Returns 1 if it is, otherwise returns 0. θ (x i ),r l ) is function g, which means using cosine similarity to calculate text entity pair x i and the relationship j The similarity between m1 is the boundary value, which is used to ensure that the score of the correct relationship is higher than the minimum interval of the score of the wrong relationship, t i For a given sample x i The correct relation index of s i is divided by t i The index of the wrong label with the highest external similarity, and m2 is the boundary value.

Images