Research of lighting system fault diagnosis method based on knowledge graph

4.1 Overall architecture of fault diagnosis knowledge graph

Knowledge graph can be divided into two levels according to its logical structure: schema layer and data layer [18] Among them, the data layer is expressed in the form of (head entity, relationship, tail entity), i.e. G = ( 𝐸𝑛𝑡𝑖𝑡𝑦 ℎ𝑒𝑎𝑑 , 𝑅𝑒𝑙𝑎𝑡𝑖𝑜𝑛 , 𝐸𝑛𝑡𝑖𝑡𝑦 𝑡𝑎𝑖𝑙 ) , or in the form of (entity, attribute, attribute value), i.e. G = ( 𝐸𝑛𝑡𝑖𝑡𝑦 , 𝑃𝑟𝑜𝑝𝑒𝑟𝑡𝑦 , 𝑉𝑎𝑙𝑢𝑒 ) as the basic expression of the data information in the form of ternary, and the ternary data related to the troubleshooting of the lighting field will be stored in the database to form a huge entity-relationship network, which will be able to form a knowledge “mapping” for the troubleshooting of the lighting system. “map”. The schema layer, on the other hand, is built on top of the data layer for storing the refined knowledge or concepts, which is the core of the knowledge graph. According to the difference in the construction order of the schema layer and the data layer, the construction of the knowledge graph can be categorised into three types, including top-down, bottom-up and hybrid construction approaches [19]. The top-down construction method refers to defining various concepts in the schema layer for knowledge mapping, and then obtaining entities and relationships that conform to these concepts and adding them to the knowledge base to form the data layer, which is clearer in the logic of realization, but the workload is larger; the bottom-up construction method refers to extracting knowledge from a large amount of data in the lighting field directly, and selecting entities with a higher degree of confidence to join them to build the data layer, on the basis of which the data layer can be constructed. The bottom-up approach refers to extracting knowledge directly from a large amount of data in the lighting field and selecting entities with higher confidence to be added to the knowledge base to construct a data layer, on the basis of which abstract concepts are summarized and generalized to construct a pattern layer, but when the distribution of the data is not even enough, the abstracted pattern layer will be incomplete and unrepresentative.

In order to construct a high-quality knowledge graph for lighting system troubleshooting, this paper adopts a hybrid knowledge graph construction method, and the construction process is shown in Fig. 2. Firstly, the top-down approach is adopted, combining the professional vocabulary and expert knowledge in the field of lighting, and the pattern layer is constructed for the overall architecture of the lighting system and the needs of troubleshooting, so as to guide the extraction of knowledge and construct the data layer; then the bottom-up approach is adopted, and the knowledge extraction is carried out by using the method of deep learning, and the ternary obtained by the extraction is mapped to the relevant concept nodes, which can supplement and update the pattern layer of the knowledge graph; finally, the construction and improvement of the knowledge graph of lighting system troubleshooting is realised. pattern layer; and finally realise the construction and improvement of the knowledge graph of lighting system fault diagnosis.

OPEN IN VIEWER

First, the overall architecture of the lighting system is analyzed to determine the networking mode between devices, and the schema layer of the knowledge graph is designed in a top-down manner for the information of troubleshooting work orders, national standard documents on lighting safety and equipment operation data to determine the devices included in the lighting system, the binding relationship between devices, and the possible failures of the devices, the causes of the failures and the solutions. Then, under the guidance of the schema layer, the bottom-up approach is used to construct the data layer, construct the BERT-BiLSTM-CRF model and use the fault diagnosis dataset to train the model, use the trained model to extract knowledge from the lighting system fault diagnosis data, and form the entity and relationship triples to be finally stored in the Neo4j graph database. While the schema layer guides and specifies the data layer, the data layer simultaneously complements and updates the schema layer.

The schema layer is the knowledge organisation structure of the lighting system troubleshooting knowledge graph, which mainly describes the relationship information between entities and entities within the lighting system through the form of concepts. The schema layer of the knowledge graph is generally stored in the ontology library, where ontology refers to the conceptual template of the knowledge, which has the advantages of a strong hierarchical structure and a small degree of redundancy. In this paper, under the guidance of experts in the field of lighting, combined with the actual situation of lighting system troubleshooting, we analyse the national standard documents of the lighting industry, the basic information of each device in the system, the troubleshooting work orders and other data, and refine the types of concepts, the relationships between concepts and the relevant attributes that are meaningful in lighting system troubleshooting, so as to form the schema layer of the lighting system troubleshooting knowledge mapping.

OPEN IN VIEWER

The lighting system fault diagnosis knowledge mapping schema layer is shown in Fig. 3, which mainly consists of each device included in the lighting system, the corresponding fault name, the corresponding fault cause, the corresponding solution, and the interrelationship between them.

The concept of “lighting system”, as the centre of the schema layer, is the core of the knowledge graph for lighting system fault diagnosis. In practice, a lighting system usually contains many various types of lighting equipment and management control equipment, which are defined as “equipment” in the schema layer. Therefore, the concept of “lighting system” and the concept of “device” naturally belong to the inclusion relationship. Each management and control device may carry multiple lighting fixtures, e.g., a DALI gateway may carry up to 64 fixtures, which are bound to the DALI gateway. Therefore, there should be a binding relationship between “devices” in the schema layer, and this binding relationship between devices can also better respond to the occurrence of faults when the fault location to accurately locate the fault, thereby reducing the difficulty of troubleshooting.

Each device has its corresponding faults, and these faults are unified and abstracted into the concept of “corresponding fault name”, so the relationship between the concept of “device” and the concept of “corresponding fault name” is called fault name. Therefore, when the concept of “equipment” points to the concept of “corresponding fault name”, the relationship is called fault name. The significance of the lighting system fault diagnosis knowledge graph is not only to indicate that a fault has occurred in a device, but also to point out the cause of the fault and the corresponding solution. Therefore, the schema layer also contains the concepts of “corresponding fault cause” and “corresponding solution”. The concept of “corresponding fault name” points to the concept of “corresponding fault cause” as the relationship of fault cause, and the concept of “corresponding fault name” points to the concept of “corresponding solution” as the relationship of solution. The concept of “corresponding solution” points to the concept of “corresponding cause” as the relationship of the cause of the fault, and the concept of “corresponding fault name” points to the concept of “corresponding solution” as the relationship of the solution.

The concepts in the schema layer are extracted and abstracted from the entities in the data layer, and the relationships between the schema layer concepts also provide guidance for the relationships between the entities in the data layer [20]. Therefore, in this section, under the guidance of the schema layer of the knowledge graph, a deep learning approach is used to extract knowledge from the data related to lighting system troubleshooting so as to construct the data layer of the knowledge graph from the bottom up. Knowledge extraction generally includes two parts, entity extraction and relationship extraction [21], because lighting system fault diagnosis belongs to a specific domain, in which the types of relationships between entities are characterised by a limited number, and this paper constructs the schema layer of the knowledge graph in Section 4.2, which can provide a basis and a normative role for the types of inter-entity relationships in the data layer. Therefore, this paper optimises the knowledge extraction work when the types of inter-entity relationships are determined, and builds the BERT-BiLSTM-CRF model to focus on the extraction of entities in the troubleshooting-related data using deep learning, which is also called the named entity recognition task in the field of Natural Language Processing, so as to improve the accuracy of the entity recognition in the data layer, and to make the constructed Lighting System Troubleshooting knowledge graph more clear and explicit.

One of the more mainstream models in named entity recognition tasks is the BiLSTM-CRF model, which was first proposed by Lample et al. 2016.The BiLSTM-CRF model mainly consists of three layers, namely Word Embedding, BiLSTM and CRF [22]. Although the experimental results of Lample et al. show that the predictive effect of the BiLSTM-CRF model in the named entity recognition task has exceeded that of the CRF model, the direct use of the model in the named entity recognition task of lighting system fault diagnosis does not work well. The main reason is that the Chinese corpus is in a different form from the English corpus, the Chinese sentences do not have clear word boundaries, so it is necessary to transform the vectors for each word in the Word Embedding layer, and the lighting domain has a large number of jargons, and there is a lack of a unified Chinese annotation dataset, which all these problems greatly increase the difficulty of feature vector extraction [23].

Introducing the pre-trained language model into the neural network can improve the sequence labelling performance of the model [24], so in this paper, for the characteristics of the lighting system fault diagnosis data, we improve on the BiLSTM-CRF model by introducing the BERT pre-trained language model as the feature representation layer in the Word Embedding layer, and propose a BERT-BiLSTM-CRF model-based The BERT model is different from the traditional unidirectional language model in that it is based on the mechanism of bi-directional attention so that the context in the text is well captured in both directions, thus providing a strong guarantee for the subsequent tasks. The structure of previous pre-trained models is limited by the unidirectional language model, which can only capture the text information in one direction, resulting in a poor representation ability of the model. The BERT model adopts a deep bi-directional transformer structure to build the whole model and uses a new masked language model for pre-training, so it can integrate the deep feature information of the text context and has the ability of bi-directional linguistic representation, and can be quickly tuned with a small number of labelled samples of the model parameters under the new training task.

In 2018, Devlin et al. of Google AI proposed a pre-trained language model BERT (Bidirectional Encoder Representations from Transformers, BERT) model by combining Transformer’s encoder and bi-directional language model [25], whose structure is shown in Fig. 4 is shown. In terms of language model training, the BERT model performs the two tasks of Masked token prediction and Next sentence prediction as the training objectives, so that the BERT model can learn the deep semantic information in the text, so as to express the text at a high level of abstraction.

OPEN IN VIEWER

In this paper, the BERT model is introduced as the feature representation layer on the basis of BiLSTM-CRF model, thus the BERT-BiLSTM-CRF model is constructed for knowledge extraction in the lighting domain, with the following two main advantages:

(1)

The BERT model maps the text into a low-dimensional dense real-valued vector space so that each dimension can represent a potential syntactic or semantic feature, and thus can represent each word in the text related to lighting system troubleshooting as a vector for a higher-level and more abstract expression, effectively solving the problem of the BiLSTM-CRF model in extracting knowledge in the lighting domain. effectively solves the problem of the BiLSTM-CRF model in extracting lighting domain knowledge, which is difficult to capture the deeper semantic information within the Chinese corpus;

The BERT-BiLSTM-CRF model consists of three models, BERT, BiLSTM and CRF, and the overall model structure is shown in Fig. 5.

Firstly, the text related to lighting system troubleshooting is inputted into the feature representation layer of the BERT model, and the important features of the text are extracted by combining the contextual information, and each word in the sentence is mapped into a word vector ( x 1 , x 2 , x 3 , x 4 ) ; then, the named entity recognition is performed by learning the long-range contextual feature information through the BiLSTM model, and the labels corresponding to all the word vectors are outputted, and each dimension of ( l 1 , l 2 , l 3 , l 4 ) represents the hitting value of the text categorised to other labels, so that the indexes at the time of classification are greatly improved. scoring value, which makes the metrics at the time of word segmentation improved greatly.

OPEN IN VIEWER

Finally, the tag sequence output from the BiLSTM model is processed using the CRF model. The CRF model takes into account the influence of the front and back tags can take into account the logic and order of the entities, for example, the tag B-EQU tags will not be followed by the tag I-REA tags, so as to obtain a globally optimal sequence of the entity category tags, which can further improve the accuracy of the named entities.

Using the BERT-BiLSTM-CRF model for knowledge extraction of data related to lighting system troubleshooting is a method that combines deep learning models with machine learning models, which not only solves the problem that the CRF model needs to manually extract textual features and cannot extract long textual features, but also solves the problem that the BiLSTM model performs poorly in the professional vocabulary of the lighting domain as well as in the entity extraction of Chinese datasets. extraction.

5.1 Experimental design and evaluation indicators

In order to verify the effectiveness of the knowledge extraction model constructed in this paper, the daily maintenance work orders of a lighting company, national standard documents of the lighting industry and expert experience documents are used as experimental objects, firstly, the YEDDA tool is used to produce the lighting system fault diagnosis dataset, and then the dataset file in .anns format is saved as a .txt file, and finally, the data for training the BERT-BiLSTM- CRF model for lighting system fault diagnosis named entity recognition task with a total of 7890 data. The dataset was randomly divided into training set, validation set and test set in the ratio of 7:2:1 for model training and testing.

The hardware environment for the experiment was 64-bit Windows 10 operating system, AMD Ryzen 7 5800H with Radeon Graphics 3.20 GHz processor, 16 GB RAM, and NVIDIA GeForce RTX 3060 graphics card. The software environment is TensorFlow 2.5.0 framework with Python 3.7.4 programming language for algorithm design. In the BERT-BiLSTM-CRF model, the number of Transformer layers of the Chinese language model pre-trained with BERT is 12, the dimension of hidden layer is 768, the number of multi-attention is 12, and the size of the total number of parameters is 110 M. Specific model training parameters are configured as shown in Table 2.

In order to verify the effect of the improved algorithm, the evaluation indexes used in this paper include Precision (P), Recall (R) and the comprehensive evaluation index F 1 value, and the calculation of each index is shown in Eqs (1), (2) and (3).

P = n 𝑇𝑃 n 𝑇𝑃 + n 𝐹𝑃(1) R = n 𝑇𝑃 n 𝑇𝑃 + n 𝐹𝑁(2) F 1 = 2PR P + R(3)

Among them, n 𝑇𝑃 , n 𝐹𝑃 , n 𝐹𝑁 and n 𝑇𝑁 are the indicators of the positive and negative judgement of the actual labelled values of the entity categories and the labelled values of the model test, and the related explanations are shown in Table 3.

n 𝑇𝑃 Indicates that the actual labelled value of the entity in the text is the correct category, and the category labelled using the model is also the correct category; n 𝐹𝑃 indicates the category whose actual labelled value is wrong but the model labelled value is correct; n 𝐹𝑁 indicates the category whose actual labelled value is correct but the model labelled value is wrong; n 𝑇𝑁 indicates the category whose actual labelled value is wrong as well as the model labelled value.

The BERT-BiLSTM-CRF model proposed in this paper is trained and applied to the task of knowledge extraction of lighting system fault diagnosis data with BiLSTM-CRF model and BiLSTM model, and compared and analysed according to the experimental results. The performance comparison of each model is shown in Fig. 6, where Fig. 6(a) shows the loss curve of each model and Fig. 6(b) shows the accuracy curve of each model. Analysing Fig. 6, it can be seen that the BERT-BiLSTM-CRF model constructed in this paper has a loss value close to 0 when the epoch is 30, and the accuracy is close to 1 when the epoch is 20, so the model constructed in this paper can achieve a lower loss value with fewer epoch iterations, and it can obtain a higher accuracy rate quickly.

The specific experimental results are shown in Table 4, the model knowledge extraction method constructed in this paper improves the precision rate by 17.58%, the recall rate by 11.65%, and the F 1 value by 14.76% compared with the BiLSTM-CRF model, thus verifying the accuracy and effectiveness of the proposed method in this paper.

OPEN IN VIEWER

6.1 Troubleshooting rules module

By adopting the deep learning method to extract knowledge from the basic information of each device in the lighting system, national standard documents of the lighting industry, troubleshooting work orders and expert experience, the knowledge graph of lighting system troubleshooting is constructed, which can assist the fault analysts to carry out lighting system troubleshooting, analyse the causes of faults according to the fault names, provide solutions to the faults and realise the faulty devices. Rapid positioning of faulty equipment. In this paper, in order to further improve the intelligent level of lighting system fault diagnosis and reduce the involvement of personnel in the process, the algorithm for determining the cause of the fault is integrated into the lighting system fault diagnosis knowledge graph and combined with the deduction grid algorithm, a fault diagnosis rule module is established, which is used for real-time judgement of whether a fault occurs in the lighting system and generates a fault descriptive language, realizing the data-driven intelligent fault diagnosis application.

The fault causes that account for a relatively high percentage of faults in the lighting system are selected to set the judgement rules in this paper as shown in Table 5, including nine fault causes such as loss of voltage, phase breakage, three-phase load imbalance, and so on. According to the definition and judgement methods of these faults in the lighting field, combined with the knowledge graph and the deduction grid algorithm to improve, forming the rule module of lighting system fault diagnosis in this paper.

When judging the possible faults of the lighting system, there is a sequence between the steps of the judgement algorithm for each fault cause. In this paper, the sequential relationship between the steps of these judgement algorithms is organised into the form of ternary groups and integrated into the lighting system fault diagnosis knowledge graph, and then the rule reasoning process is improved by combining the deduction grid algorithm with the lighting system fault diagnosis knowledge graph, and the steps of the fault cause judgement algorithms are fused and optimised, and a complete fault diagnosis rule module is formed in the end.

OPEN IN VIEWER

The deduction lattice algorithm is an intelligent and efficient partial order inference model [27], and the inference rules include the following three points.

(1)

C 1 ∧ C 2 ∧ … ∧ C n → h , C 1 , C 2 , … , C n are the conditions and h is the conclusion;

Before the use of deduction grid algorithm optimisation, usually each kind of fault needs to establish a judgement algorithm to carry out the diagnosis of the fault, however, due to the fact that the judgement algorithms of these fault causes are independent of each other, it is necessary to judge each algorithm one by one in the actual use of the algorithms, so it will produce a large amount of repetitive work, resulting in the judgement of the speed of the slow speed, which affects the speed of the fault diagnosis of the lighting system.

OPEN IN VIEWER

After optimising the fault judgement algorithm by combining the deduction lattice algorithm with the fault diagnosis knowledge graph, the algorithm flow of the established fault diagnosis rule module is shown in Fig. 7, and the optimised algorithm reduces the redundant judgement steps and improves the speed of the lighting system fault diagnosis. Combining the deduction grid algorithm with the knowledge graph, making full use of the powerful relational expression capability of the knowledge graph, the established fault diagnosis rule module can judge whether the lighting system is faulty or not in real time by obtaining data such as the equipment state information generated during the operation of the lighting system, realizing the conversion from data to the descriptive language of faults, thus reducing the involvement of personnel in the fault diagnosis process, and realizing the data-driven lighting system fault diagnosis application. The application of data-driven lighting system fault diagnosis.

The constructed lighting system fault diagnosis knowledge graph is stored and visualised in the graph database Neo4j. The process of applying the lighting system fault diagnosis knowledge graph for fault diagnosis is shown in Fig. 8. Firstly, obtain the state information and operation-related parameter data of the equipment in the running lighting system, which are mainly semi-structured data in JSON format, recording the parameters such as voltage, current and power of the equipment; then input these data into the constructed fault diagnostic rule module to determine whether the system is faulty and output the name of the fault; finally, use Neo4j’s Finally, using Neo4j’s Cypher query statement to search the fault name in the knowledge graph, when there is a corresponding fault in the lighting system fault diagnosis knowledge graph, not only can we query the exact location of the fault information, but also can directly get the cause of the fault and the solution. In addition, if the corresponding fault does not exist in the lighting system fault diagnosis knowledge graph, it is necessary to deposit the fault as new knowledge into the knowledge graph to update it, so that the knowledge graph provides a more accurate, comprehensive, real-time fault diagnosis basis. Finally, the updated knowledge graph is combined with the deduction lattice algorithm to achieve the updating and improvement of the fault diagnosis rule module.