Intelligent analysis method and system for equipment maintenance failure based on internet of things

Abstract

The application relates to the technical field of Internet of Things and artificial intelligence, and discloses a device maintenance fault intelligent analysis method and system based on the Internet of Things, wherein the device maintenance fault intelligent analysis method based on the Internet of Things comprises the following steps: constructing a time sequence knowledge graph by representing entities and relations as time functions; monitoring vector changes in a knowledge embedding space, detecting concept drift phenomena, and mining device performance evolution modes; constructing a device state transition model by using a recurrent neural network, predicting future states of the device, and evaluating fault risks; collecting feedback, processing knowledge conflicts, and updating the model; and generating multi-time scale fault risk early warning and maintenance suggestions; the application improves knowledge acquisition efficiency, provides dynamic knowledge support for the whole life cycle of a device, realizes early identification of gradual faults, has self-adaptive knowledge updating capability, provides multi-dimensional fault analysis, optimizes maintenance resource allocation, and improves device management efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet of Things and artificial intelligence, more particularly, it relates to an intelligent analysis method and system for equipment maintenance failure based on Internet of Things. BACKGROUND

[0002] With the rapid development of industry and intelligent manufacturing, Internet of Things technology is widely used in industrial equipment monitoring and maintenance. The stable operation of industrial equipment is a key factor to ensure production continuity and enterprise efficiency, so an efficient equipment maintenance and failure warning system is of great significance. Traditional equipment maintenance mainly relies on periodic inspection and passive response maintenance strategy, that is, equipment is repaired only after failure, which not only causes production interruption, but also causes great economic loss.

[0003] In recent years, predictive maintenance technology based on data-driven has gradually emerged, which collects equipment operation data through sensors and uses statistical analysis or machine learning methods to evaluate the health status of equipment and predict possible failure risks. However, these methods mainly focus on data pattern recognition, lack of deep utilization of equipment maintenance knowledge and modeling ability of long-term evolution trend. Especially for industrial equipment with long service life and complex failure modes, it is difficult to accurately grasp the complete evolution law of equipment state change with time by relying solely on short-term data analysis.

[0004] As an effective tool for knowledge representation and management, knowledge graph has shown strong knowledge organization and reasoning ability in many fields. Traditional knowledge graph technology mainly focuses on the representation of static knowledge, and constructs a domain knowledge network through entities and relationships. However, in the field of equipment maintenance, the knowledge of equipment performance characteristics, failure modes and maintenance methods evolves with the aging of equipment, changes in environment and updates of technology, which is difficult to be effectively expressed by static knowledge graph, resulting in the disconnection between the information in the knowledge base and the actual application.

[0005] Therefore, it is necessary to develop a comprehensive solution that can effectively integrate Internet of Things data collection, knowledge graph technology, time series feature modeling and intelligent failure analysis, in order to improve the efficiency of equipment maintenance, reduce maintenance cost and prolong the service life of equipment. SUMMARY

[0006] The present application provides an intelligent analysis method and system for equipment maintenance failure based on Internet of Things, which solves the technical problem of lack of deep utilization of equipment maintenance knowledge and modeling ability of long-term evolution trend in related technologies.

[0007] The present application provides an intelligent analysis method for equipment maintenance failure based on Internet of Things, which comprises the following steps:

[0008] automatically extracting entities and relations from equipment maintenance related documents, and constructing an initial knowledge graph;

[0009] Based on the constructed initial knowledge graph, the entities and relations are represented as a function of time, and a time-series knowledge graph is constructed.

[0010] Using the constructed time-series knowledge graph, by monitoring the changes of entity and relation vectors in the knowledge embedding space over time, the concept drift phenomenon is detected and the equipment performance evolution pattern is mined.

[0011] Based on the constructed time-series knowledge graph and current observation data, a recursive neural network is used to construct a device state transition model to predict the future state of the device and assess the risk of failure.

[0012] Collecting feedback from maintenance personnel on system prediction results, processing knowledge conflicts and updating the time-series knowledge graph and prediction model.

[0013] Combining real-time state data of the device and the updated time-series knowledge graph, generating multiple time scale fault risk warning and maintenance suggestions.

[0014] In a preferred embodiment, the step of automatically extracting entities and relations from equipment maintenance related documents comprises:

[0015] Pretreatment of maintenance manual, historical failure record and maintenance report and unstructured text;

[0016] A named entity recognition method based on conditional random field and bidirectional long short-term memory network is used to identify and extract three types of core entities: equipment components, failure phenomena and maintenance methods.

[0017] The relationship extraction algorithm using remote supervision and attention mechanism is used to identify the semantic relationship between entities.

[0018] The extracted entities and relations are integrated into a unified knowledge framework and consistency check is performed.

[0019] In a preferred embodiment, the step of representing entities and relations as a function of time comprises:

[0020] Introducing time attributes to entities and relations in the knowledge graph, and extending the traditional triple knowledge representation to quadruple.

[0021] Using tensor decomposition method, the entity vector and relation vector are combined with the time vector to get the time-aware embedding representation.

[0022] Define time-series relationship reasoning rules to support time-constrained reasoning on the time-series knowledge graph.

[0023] In a preferred embodiment, the step of detecting concept drift and mining device performance evolution patterns includes:

[0024] Calculate the semantic displacement of an entity at two time points. When the semantic displacement exceeds a preset threshold, it is determined that the calculated entity has concept drift.

[0025] Utilize cyclic variational autoencoders to automatically extract periodic and trend patterns of knowledge evolution from temporal knowledge graphs;

[0026] Construct a knowledge evolution visualization module to present the detected concept drift and evolution patterns in an intuitive way.

[0027] In a preferred embodiment, the step of predicting the future state of the equipment and assessing the risk of failure includes:

[0028] Based on the time series knowledge graph, the knowledge state of the device at a specific historical time point can be reconstructed;

[0029] A device state transition model is constructed using a recurrent neural network. The model input is the device state vector and related knowledge graph information at the current time point, and the output is the state prediction at the future time point.

[0030] Calculate the similarity between the predicted state and known failure modes and generate probability distributions for different types of failures.

[0031] In a preferred embodiment, the steps of processing knowledge conflicts and updating the time series knowledge graph and prediction model include:

[0032] Calculate the credibility of conflicting knowledge based on the amount of supporting evidence and source reliability;

[0033] Demote or eliminate low-credibility knowledge;

[0034] Add new knowledge with high credibility;

[0035] For knowledge with timeliness differences, update the time attribute instead of directly replacing it;

[0036] An incremental learning method is used to update the parameters of the state transition model and the fault risk assessment model.

[0037] In a preferred embodiment, the step of generating fault risk warnings and maintenance suggestions at multiple time scales includes:

[0038] Generate equipment status predictions at different time scales based on the state transition model;

[0039] Determine the warning level based on the matching degree between the predicted status and the failure mode and the time urgency;

[0040] Based on the fault risk early warning result and the maintenance knowledge in the time sequence knowledge graph, a targeted maintenance suggestion is automatically generated;

[0041] Based on the device historical maintenance record, the time and resources required for maintenance are estimated to assist in the development of a maintenance plan.

[0042] In a preferred embodiment, the time attribute includes a time point, a time interval or a periodic feature, the time point represents a specific moment when the knowledge is valid, the time interval represents the start and end time when the knowledge is valid, and the periodic feature represents a seasonal failure mode.

[0043] In a preferred embodiment, the recurrent variational autoencoder includes an encoder and a decoder, the encoder encodes the time sequence entity representation sequence into the distribution parameter of the latent variable, and the decoder reconstructs the sequence based on the sampled latent variable and historical information.

[0044] In a preferred embodiment, the Internet of Things-based device maintenance fault intelligent analysis system is used to execute the Internet of Things-based device maintenance fault intelligent analysis method, and includes:

[0045] The multi-source text knowledge extraction and graph construction module is used to automatically extract entities and relationships from device maintenance related documents, and construct an initial knowledge graph.

[0046] The time sequence knowledge graph construction module is used to represent the entities and relationships in the constructed initial knowledge graph as a time function based on the constructed initial knowledge graph, and construct a time sequence knowledge graph.

[0047] The concept drift detection module is used to detect concept drift phenomena and mine device performance evolution patterns by monitoring the changes of entity and relationship vectors in the knowledge embedding space over time using the constructed time sequence knowledge graph.

[0048] The device state prediction module is used to construct a device state transition model using a recurrent neural network based on the constructed time sequence knowledge graph and current observation data, predict future device states and evaluate fault risks.

[0049] The knowledge update module is used to collect feedback from maintenance personnel on system prediction results, handle knowledge conflicts and update the time sequence knowledge graph and prediction model.

[0050] The fault early warning module is used to generate fault risk early warning and maintenance suggestions of multiple time scales in combination with real-time device state data and updated time sequence knowledge graph.

[0051] The beneficial effects of the present application are:

[0052] Significant improvement in knowledge acquisition efficiency: Through weakly supervised learning and distant supervision technology, the system can automatically extract device failure-related entities and relationships from a large amount of unstructured text. Compared with traditional manual construction methods, the knowledge acquisition efficiency is improved by several times, the accuracy is significantly improved, and the labor cost of knowledge graph construction and maintenance is greatly reduced.

[0053] Dynamic knowledge support for the whole life cycle of the device: Through the time series knowledge graph technology, the system can represent the evolution of knowledge over time and capture the state characteristics and failure modes of the device in different use stages, providing targeted knowledge support for the whole life cycle of the device maintenance, avoiding the timeliness problem of knowledge.

[0054] Early identification of gradual failure: Combined with the evolution pattern mining technology of the recurrent variational autoencoder, the system can identify the trend of small changes in device performance and discover early signs of gradual failure in advance. Compared with traditional methods, the early warning time of failure is significantly increased, providing sufficient time window for preventive maintenance.

[0055] Adaptive knowledge update: Through maintenance feedback collection and knowledge conflict processing mechanism, the system realizes the self-evolution of knowledge, continuously learns and optimizes the knowledge structure from actual maintenance cases, so that the system continuously improves the accuracy and practicality during use.

[0056] Multi-dimensional fault analysis capability: The system can not only analyze the current fault, but also trace the development history and predict the future evolution trend through time series knowledge reasoning, providing all-round reference for maintenance decision-making, improving the accuracy and comprehensiveness of fault diagnosis.

[0057] Optimal allocation of maintenance resources: Based on the functions of multi-time scale fault warning and maintenance suggestion generation, the system can support the advance planning and optimal allocation of maintenance resources. Compared with passive response maintenance, the average maintenance cost is significantly reduced, the device availability is significantly improved, and the efficiency of enterprise equipment management is greatly improved. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 is a flowchart of the device maintenance fault intelligent analysis method based on Internet of Things of the present application;

[0059] Figure 2 is a line graph of the knowledge acquisition efficiency improvement of the present application;

[0060] Figure 3 is a column chart of the time series knowledge graph concept drift detection accuracy comparison of the present application;

[0061] Figure 4 is a column chart of the fault prediction accuracy performance on different device types of the present application;

[0062] Figure 5is a radar chart of the fault risk assessment performance of the present application;

[0063] Figure 6 is a bubble chart of the maintenance resource optimization effect of the present application. DETAILED DESCRIPTION

[0064] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that discussions of these implementations are merely provided to enable better understanding of the subject matter described herein, and can include various process or components according to one embodiment that can be omitted, substituted, or added according to other embodiments without deviating from the spirit and scope of the present disclosure. Various examples can omit, substitute, or add various processes or components according to one embodiment based on the discussion herein. Also, some embodiments described can be combined.

[0065] An Internet of Things-based equipment maintenance fault intelligent analysis method is disclosed in at least one embodiment of the present application, as shown in the accompanying drawings, comprising the following steps: Figure 1 as shown, comprising the following steps:

[0066] Step 1, automatically extracting entities and relationships from equipment maintenance related documents to build an initial knowledge graph;

[0067] Specifically, the following sub-steps are included:

[0068] Step 1.1, text preprocessing;

[0069] The unstructured text such as maintenance manual, historical fault record and maintenance report is preprocessed, including text segmentation, removal of stop words, part-of-speech tagging and syntax analysis.

[0070] A specific dictionary is constructed for professional terms in the field of equipment maintenance during the preprocessing process to improve the accuracy of term recognition.

[0071] Step 1.2, entity recognition and extraction;

[0072] A named entity recognition method based on Conditional Random Field (CRF) and Bidirectional Long Short-Term Memory Network (BiLSTM) is used to identify and extract three core entities of equipment components, fault phenomena and maintenance methods from the preprocessed text. The entity recognition model is initially trained by a small amount of manually annotated seed entities, and then further optimized the model performance by using a larger scale of unlabeled text combined with remote supervision technology.

[0073] In the embodiments of the present application, the BiLSTM-CRF model calculates the conditional probability of the label sequence, where the transition from one label to another has a transition score, and each position label output by the network has an emission score.

[0074] This model can learn the context features of words and the transition constraints between labels, improving the recognition accuracy of device-specific terminology.

[0075] When processing complex device maintenance documents, a domain pre-trained language model such as BERT or its variants can be introduced to further improve the accuracy of entity recognition. The pre-trained model can be fine-tuned for domain adaptability through device maintenance manuals and technical documents, better understanding the professional terminology and expression methods in the device maintenance field.

[0076] Step 1.3, relationship extraction;

[0077] Based on the extracted entities, a relationship extraction algorithm using remote supervision and attention mechanism is used to identify the semantic relationships between entities.

[0078] First, a small number of predefined relationship templates are used to extract sentences containing known entity pairs from the text as training samples, and then a convolutional neural network with attention mechanism is used to automatically learn the relationship representation between entities.

[0079] Finally, relationships such as "component has fault phenomenon", "fault phenomenon needs maintenance method", and "component contains sub-component" are extracted.

[0080] Step 1.4, knowledge graph integration and consistency check;

[0081] Integrate the extracted entities and relationships into a unified knowledge framework to form an initial knowledge graph.

[0082] Use rule-based and statistical consistency checking methods to identify and handle knowledge conflicts, including entity disambiguation, relationship redundancy elimination, and logical consistency verification.

[0083] Finally, the integrated knowledge data is stored in a graph database to form a queryable and inferable graph structure.

[0084] Step 2, based on the constructed initial knowledge graph, represent entities and relationships as a function of time to construct a time-series knowledge graph;

[0085] Specifically, the following sub-steps are included:

[0086] Step 2.1, time attribute modeling;

[0087] Introduce time attributes to entities and relationships in the knowledge graph, and extend the traditional triple knowledge representation <Entity 1, Relationship, Entity 2> to quadruple <Entity 1, Relationship, Entity 2, Time Attribute>.

[0088] The temporal attribute can be a time point (representing a specific moment when the knowledge is true) or a time interval (representing the start and end time when the knowledge is valid), or can contain periodic characteristics (such as seasonal failure modes).

[0089] In some embodiments, the temporal attribute can be further subdivided into creation time, update time, valid period start time, and valid period end time to support more fine-grained temporal knowledge management.

[0090] For example, for some seasonal failure modes, their historical occurrence time, expected valid period, and periodic characteristic parameters can be recorded.

[0091] Step 2.2, time-aware embedding representation;

[0092] Representing entities and relationships in the knowledge graph as time functions And Instead of static vectors.

[0093] In specific implementation, using tensor decomposition method, the entity vector And the relationship vector And the time vector Are combined to obtain the time-aware embedding representation:

[0094] ;

[0095] ;

[0096] Where, Represents the time-aware embedding representation of the entity at time It reflects the semantic characteristics of the entity changing over time; Represents the time-aware embedding representation of the relationship at time It reflects the semantic characteristics of the relationship changing over time; Is the basic embedding vector of the entity, representing the semantic representation of the entity in the initial state; Is the basic embedding vector of the relationship, representing the semantic representation of the relationship in the initial state; Is a time parameter, which can be a timestamp or an encoded representation of a time interval; Is the time mapping matrix of the entity, used to map the time parameter to the entity semantic space, controlling how time affects the semantic change of the entity; Is the time mapping matrix of the relationship, used to map the time parameter to the relationship semantic space, controlling how time affects the semantic change of the relationship; Is the entity time influence strength parameter, used to adjust the degree of influence of time on the semantic change of the entity; is the time influence intensity parameter, used to adjust the degree of influence of time on the semantic change of relations.

[0097] This representation method allows the same entity or relation to have different semantic representations at different time points, capturing the evolution of concepts over time.

[0098] The system can also use a nonlinear mapping function instead of the above linear combination method to build more complex time mapping relationships through a multi-layer perception network:

[0099] ;

[0100] where, represents the time-aware entity representation obtained using a multi-layer perception network; is the basic embedding vector of the entity, representing the static semantic features of the entity in the initial state; is a time parameter, which can be a timestamp or an encoded representation of a specific time point; represents a multi-layer perception network composed of multiple neural networks, capable of capturing the nonlinear interaction between entity vectors and time; represents the entity vector and the time parameter as input into the multi-layer perception network to obtain the output vector, representing the nonlinear influence of time on entity semantics. This nonlinear mapping method can express more complex time evolution patterns than simple linear combinations, and is particularly suitable for modeling the nonlinear changes of device component performance over time.

[0101] Step 2.3, definition of time sequence reasoning rules;

[0102] A set of time sequence reasoning rules are defined to support time-constrained reasoning on the time sequence knowledge graph.

[0103] These rules include time point inclusion relationship, time interval overlap relationship, and time sequence constraints of causal relationship, etc.

[0104] For example, if relation R1 is true at time t1, relation R2 is true at time t2, and t1 precedes t2, there may be a causal relationship between R1 and R2.

[0105] Step 3, using the constructed time sequence knowledge graph, by monitoring the changes of entity and relation vectors in the knowledge embedding space over time, detecting concept drift phenomenon and mining device performance evolution patterns;

[0106] Specifically, the following sub-steps are included:

[0107] Step 3.1, concept drift detection;

[0108] By monitoring the changes of entity and relation vectors in the knowledge embedding space over time, we identify the drift of concept definitions and relation strengths.

[0109] For an entity , we compute its semantic shift between two time points and :

[0110] ;

[0111] where denotes the semantic shift of entity between two time points; is the base embedding vector of the entity; denotes the first time point; denotes the second time point; denotes the vector representation of entity at time point ; denotes the vector representation of entity at time point ; denotes the Euclidean distance between two vectors (other distance measures can also be used) to quantify the magnitude of semantic change of the entity.

[0112] When the semantic shift exceeds a pre-set threshold, we determine that there is a concept drift for this entity. Similarly, we can compute the semantic shift of a relation to detect changes in relation strength.

[0113] By clustering analysis, we categorize the detected concept drifts into two patterns: gradual (slow and continuous change) and abrupt (rapid and discrete change).

[0114] Step 3.2, Evolution pattern mining;

[0115] We utilize a Recurrent Variational Autoencoder (RVAE) to automatically extract periodic and trend patterns of knowledge evolution from the temporal knowledge graph. RVAE combines the generative ability of Variational Autoencoder and the sequence modeling ability of Recurrent Neural Network, which can effectively capture the temporal dependencies in sequence data.

[0116] The encoder part of RVAE model encodes the sequence of temporal entity representations into the distribution parameters of latent variables, while the decoder part reconstructs the sequence based on the sampled latent variables and historical information. The model is trained by minimizing the combined loss function of reconstruction error and KL divergence.

[0117] By analyzing the learned latent space structure, the system can identify the evolution law of the device performance parameters, such as linear degradation trend, periodic fluctuation, and stage change, etc. mode, providing the basis for fault prediction.

[0118] In addition, the present application also provides another implementation manner, which can replace RVAE with a time sequence model adopting a self-attention mechanism to directly extract long and short term dependencies from time sequence data. The method calculates the correlation of different time points in the sequence through a self-attention layer, without explicit sequence modeling, and is suitable for capturing the dependencies between distant time points.

[0119] Step 3.3, knowledge evolution visualization;

[0120] The knowledge evolution visualization module is constructed to present the detected concept drift and evolution mode in an intuitive way.

[0121] The visualization content includes entity semantic drift trajectory graph, relationship strength change curve, evolution mode clustering view, etc., which helps maintenance personnel understand the change law of device knowledge over time.

[0122] Step 4, based on the constructed time sequence knowledge graph and current observation data, a device state transition model is constructed by using a recurrent neural network to predict the future state of the device and evaluate the fault risk;

[0123] Specifically, the following sub-steps are included:

[0124] Step 4.1, history state reconstruction;

[0125] Based on the time sequence knowledge graph, the knowledge state of a specific historical time point of the device is reconstructed. Given a time point t, the system retrieves the entity and relationship representation corresponding to the time point and , reconstructs the knowledge graph snapshot of the time point, and supports query and analysis of historical fault cases and maintenance records.

[0126] The knowledge graph snapshot refers to the knowledge graph state at a specific time point , which contains all entities and their relationship information at the time point;

[0127] The reconstruction process refers to the process of extracting and restoring the complete knowledge structure of the time point from the time sequence knowledge graph according to the time parameter .

[0128] Step 4.2, state transition model construction;

[0129] A recurrent neural network (RNN) is used to construct a device state transition model to capture the evolution of the device state over time. The model input is the device state vector at the current time point t and the relevant knowledge graph information, and the output is the state prediction at the future time point.

[0130] According to one embodiment of the present application, the device state vector includes key parameters such as temperature, vibration, pressure, etc. collected from Internet of Things sensors, as well as advanced features derived from these parameters; the knowledge graph information includes time-aware representation of entities and relationships related to the current device.

[0131] In some embodiments, a long short-term memory network (LSTM) or a gated recurrent unit (GRU) can be used instead of a basic RNN to better handle long-term dependencies. The LSTM model can effectively learn long-term dependencies in long sequences by introducing memory cells and multiple gating mechanisms, and is particularly suitable for modeling the long-term running state of a device.

[0132] In addition, for different time scale prediction tasks, a multi-scale state transition model can be constructed. For example, for short-term prediction, fine-grained time steps and more sensor features are used; for long-term prediction, coarse-grained time steps and more historical trend features are used.

[0133] Step 4.3, failure risk assessment;

[0134] Based on the prediction results of the state transition model and the failure mode information in the time series knowledge graph, the risk of future failure of the device is assessed. The similarity between the predicted state and the known failure mode is calculated to generate a probability distribution of different types of failure.

[0135] The similarity calculation method used in the present application is based on distance metric function, and various distance metric methods such as Euclidean distance, cosine similarity or Mahalanobis distance can be selected.

[0136] The system calculates the distance between the predicted state and each type of failure mode, and converts it to a probability distribution through a soft-max function, to realize the risk quantification of multiple possible failure types.

[0137] The system can also combine the historical failure frequency of the device as prior knowledge, use the Bayesian framework to calculate the posterior failure probability, and improve the accuracy of risk assessment.

[0138] In addition, the system can also generate a confidence interval of the failure risk to provide an uncertainty measure of the risk assessment, helping maintenance personnel to better understand the reliability of the prediction results.

[0139] Step 5, Collecting maintenance personnel's feedback on system prediction results, handling knowledge conflicts and updating the time sequence knowledge graph and prediction model;

[0140] Specifically including the following sub-steps:

[0141] Step 5.1, Maintenance feedback collection;

[0142] Design maintenance feedback collection interface to collect maintenance personnel's evaluation on system prediction results and suggestions, as well as new information found in actual maintenance process.

[0143] Feedback content includes fault diagnosis accuracy score, maintenance method effectiveness evaluation, newly discovered fault phenomena and causes, etc.

[0144] Step 5.2, Knowledge conflict detection and handling;

[0145] Compare maintenance feedback with information in existing knowledge graph to detect potential knowledge conflicts.

[0146] When new feedback conflicts with existing knowledge, the system handles it through the following strategies:

[0147] Calculate the credibility of conflicting knowledge based on the number of supporting evidence and source reliability;

[0148] Down-weight or eliminate low-credibility knowledge;

[0149] Add high-credibility new knowledge;

[0150] For knowledge with time difference, update its time attribute instead of directly replacing it.

[0151] The system can use fuzzy logic or Bayesian network method to handle knowledge uncertainty.

[0152] By introducing confidence or probability values, the system can represent and reason about knowledge with uncertainty, rather than simple binary true or false judgments. For example, the association between a fault and a component can be represented as a relationship with 0.8 confidence, and this confidence can be dynamically adjusted as more evidence accumulates.

[0153] Step 5.3, Model parameter update;

[0154] Based on accumulated maintenance feedback data, regularly update the parameters of state transition model and fault risk assessment model.

[0155] Use incremental learning method to integrate new knowledge on the basis of preserving existing knowledge, avoiding catastrophic forgetting problem.

[0156] Step 6, Combine real-time state data of equipment with updated time sequence knowledge graph to generate fault risk early warning and maintenance suggestions on multiple time scales;

[0157] Specifically, the following sub-steps are included:

[0158] Step 6.1, Scale State Prediction;

[0159] Based on the state transition model, generate device state predictions at different time scales (short-term, medium-term, and long-term).

[0160] Short-term prediction focuses on acute failure risk in the next few hours to days;

[0161] Medium-term prediction focuses on the trend of failure development in the next few weeks;

[0162] Long-term prediction focuses on the performance degradation law of the device on a monthly to quarterly basis.

[0163] Step 6.2, Warning Level Determination;

[0164] According to the matching degree of predicted state and failure mode and the time urgency, determine the warning level.

[0165] The warning level is divided into four levels from low to high: normal, attention, warning, and danger, each corresponding to different handling strategies and response time limits.

[0166] Step 6.3, Maintenance Suggestion Generation;

[0167] Based on the failure risk warning results and maintenance knowledge in the time sequence knowledge graph, automatically generate targeted maintenance suggestions.

[0168] The maintenance suggestion content includes possible failure causes, suggested parts for inspection, recommended maintenance methods, and required tools and materials.

[0169] The system can also estimate the time and resources required for maintenance based on the device's historical maintenance records, assisting in maintenance plan development.

[0170] The system can generate multiple maintenance plans and sort them based on factors such as device importance, spare parts supply, and maintenance personnel availability, providing decision-makers with multiple options.

[0171] In addition, the system can also recommend the best maintenance time window based on the device's usage plan, minimizing the impact of maintenance activities on production.

[0172] For complex device systems, the system can also provide collaborative maintenance suggestions, combining the maintenance tasks of multiple related devices to reduce maintenance downtime and improve overall maintenance efficiency.

[0173] Application examples of this embodiment:

[0174] To better illustrate the practical application effect of the device maintenance fault intelligent analysis method based on Internet of Things provided in the present application, the actual application case of a cold rolling mill of a certain steel plant will be described in detail below.

[0175] Application scenario:

[0176] A certain steel plant has multiple cold rolling mills. These devices are complex in structure, containing hydraulic systems, transmission systems, electrical control systems and other subsystems, and there are complex mutual influence relationships between the systems.

[0177] These devices operate in harsh environments, work under high load all year round, and have frequent equipment failures and high maintenance costs.

[0178] Traditional periodic maintenance methods cannot cope with sudden failures caused by equipment aging and performance degradation, and after-maintenance often leads to long-term shutdown of production lines, causing huge economic losses.

[0179] The plant decided to adopt the method provided in the present application to build an Internet of Things-based device maintenance fault intelligent analysis system to realize intelligent maintenance of the cold rolling mill throughout its life cycle.

[0180] The system collects device operation data in real time through Internet of Things sensors, and constructs a time sequence knowledge graph in combination with historical maintenance records to realize early warning and predictive maintenance of faults.

[0181] Implementation process instance:

[0182] Multi-source text knowledge extraction and graph construction:

[0183] The system first collected the documents such as maintenance manuals, fault records and maintenance reports of the cold rolling mill of the plant in the past five years, totaling about 8,500 text documents.

[0184] In the text preprocessing stage, the system performed word segmentation, stop word removal and syntax analysis on these documents, and improved the accuracy of term recognition in combination with a cold rolling device-specific dictionary.

[0185] Through the BiLSTM-CRF model for entity recognition, the system identified 2,364 device component entities (such as “roller bearing”, “hydraulic cylinder”, “sensor”, etc.), 976 fault phenomenon entities (such as “abnormal vibration”, “pressure fluctuation”, “excessive temperature”, etc.) and 843 maintenance method entities (such as “replace the sealing ring”, “adjust the gap”, “clean the filter”, etc.) from the preprocessed text.

[0186] Based on the entity recognition results, the system applies a relationship extraction algorithm with remote supervision and attention mechanism to extract about 12,000 entity relationships, including "the roller bearing has abnormal vibration", "the abnormal vibration needs to replace the bearing" and other relationship types. After consistency check and redundancy elimination, an initial knowledge graph containing more than 4,000 entities and more than 10,000 relationships is finally constructed. The efficiency of knowledge acquisition is improved as Figure 2 As shown in FIG. 6, the performance change of the knowledge extraction method based on weakly supervised learning with the increase of the number of labeled samples is shown.

[0187] Temporal property extension and concept drift detection:

[0188] The system adds time attributes to entities and relationships in the knowledge graph, converting them into a four-tuple structure. For example, "the roller bearing has abnormal vibration on March 15, 2021" indicates that the relationship is recorded at a specific time point. Through a time-aware embedding representation method, the system represents entities and relationships as a function of time rather than a static vector.

[0189] By monitoring the semantic displacement in the embedding space, the system detects multiple concept drift phenomena. For example, the definition of the "roller bearing abnormal vibration" fault phenomenon has changed over time. Early (before 2018), it was mainly related to improper installation, while recently (after 2020), it is more related to insufficient lubrication. In addition, the system also found that the effectiveness of some maintenance methods decreases with the increase of equipment service life, such as "adjusting the hydraulic valve" has a significant decrease in effect for equipment used for more than 8 years.

[0190] Using a recurrent variational autoencoder, the system mines multiple evolution patterns from historical data, including seasonal fluctuation patterns of roller bearing temperature, linear degradation trends of hydraulic system pressure, and phased change patterns of transmission system vibration. The accuracy comparison of concept drift detection of the temporal knowledge graph is shown in FIG. 7. Figure 3 As shown in FIG. 6, the performance change of the knowledge extraction method based on weakly supervised learning with the increase of the number of labeled samples is shown.

[0191] State prediction and fault warning:

[0192] Based on the constructed temporal knowledge graph and state transition model, the system can predict the future state of the equipment. The performance of fault prediction accuracy on different types of equipment is shown in FIG. 8, which shows the comparison of fault prediction accuracy of the method of the present application and the traditional method on different types of equipment, as well as the percentage of accuracy improvement. Figure 4

[0193] ​In a practical application, the system monitored the pressure fluctuation frequency of the hydraulic system of No. 1 cold rolling mill, which gradually increased and the temperature slowly rose. By comparing with the fault patterns in historical data, it predicted that there was a 78% probability of hydraulic system failure in the next 14 days, mainly due to the aging of the hydraulic pump seal ring.

[0194] The system generated a "warning" level alert and recommended checking the hydraulic pump seal ring during the planned downtime, with an estimated repair time of 4 hours and required materials including a specific model seal ring and hydraulic oil.

[0195] The maintenance personnel checked the hydraulic pump seal ring during the planned downtime according to the system's suggestion, and indeed found signs of aging, which was timely replaced to avoid possible sudden failure. The fault risk assessment performance is shown in Figure 5 , which compares the performance of different fault risk assessment methods in multiple key performance indicators.

[0196] Knowledge update and system optimization:

[0197] After the repair, the maintenance personnel submitted the actual findings and handling through the feedback interface, confirming the accuracy of the system's diagnosis, but also supplementing a new finding: the seal ring of this type of hydraulic pump ages faster in high temperature environments. The system added this new knowledge to the time series knowledge graph and updated the parameters of the related fault models. The maintenance resource optimization effect is shown in Figure 6 , which shows the maintenance resource optimization effect of different size equipment groups.

[0198] In the next three months of operation, the system's warning accuracy rate for similar hydraulic systems improved from 78% to 86%, and the warning lead time increased from an average of 14 days to 19 days, providing a more sufficient time window for maintenance planning. Technical effect verification:

[0199] After one year of application of the system in this cold rolling mill, significant technical effects have been achieved:

[0200] Fault prediction accuracy rate improved: the system's prediction accuracy rate for major fault types is shown in Table 1:

[0201] Table 1: System's prediction accuracy rate for major fault types

[0202]

[0203] Equipment availability and maintenance cost improved: the comparison of equipment operation indicators before and after the application of the system is shown in Table 2:

[0204] Table 2: Comparison of equipment operation indicators before and after the application of the system

[0205]

[0206] It can be seen from the above data that the device maintenance fault intelligent analysis method based on the Internet of Things provided in the application has achieved remarkable technical effects in actual application, not only greatly improving the accuracy and advance of fault prediction, but also significantly improving the device availability, reducing the maintenance cost, verifying the practicability and effectiveness of the method.

[0207] The above describes the embodiments of the application, but the embodiments are not limited to the specific implementation described above, and the specific implementation described above is only illustrative but not restrictive, and the ordinary skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which all belong to the protection of the embodiments.

Claims

1. An intelligent analysis method for equipment maintenance failure based on the Internet of Things, characterized in that: The following steps are involved: Automatically extract entities and relationships from equipment maintenance-related documents to build an initial knowledge graph; Based on the constructed initial knowledge graph, entities and relationships are represented as time functions to construct a temporal knowledge graph; The steps to represent entities and relations as functions of time include: Introducing time attributes to entities and relationships in the knowledge graph, expanding the traditional triple knowledge representation to quadruple; Using tensor decomposition, we combine entity and relationship vectors with the time vector to obtain a time-aware embedding representation. Specifically: in, Represents an entity at time The time-aware embedding representation of entities reflects the semantic characteristics of entities that change over time; Represents a relationship in time The time-aware embedding representation of the relationship reflects the semantic characteristics of the relationship changing over time; is the basic embedding vector of the entity, which represents the semantic representation of the entity in its initial state; is the basic embedding vector of the relationship, which represents the semantic representation of the relationship in the initial state; It is a time parameter, which is an encoded representation of a timestamp or time interval; It is the time mapping matrix of the entity, which is used to map the time parameters to the entity semantic space and control how time affects the semantic changes of the entity; It is the time mapping matrix of the relationship, which is used to map the time parameters to the relationship semantic space and control how time affects the semantic changes of the relationship; It is the entity time influence intensity parameter, which is used to adjust the degree of influence of time on the semantic change of the entity; is the relation time influence intensity parameter, which is used to adjust the degree of influence of time on relation semantic changes; By using the constructed temporal knowledge graph and monitoring the changes of entity and relationship vectors in the knowledge embedding space over time, we can detect concept drift and explore the evolutionary patterns of device performance. Based on the constructed time-series knowledge graph and current observation data, a recurrent neural network is used to build a device state transition model to predict the future state of the device and assess the risk of failure. Collect feedback from maintenance personnel on the system's prediction results, resolve knowledge conflicts, and update the time series knowledge graph and prediction model; Combining the real-time status data of the equipment and the updated time series knowledge graph, fault risk warnings and maintenance suggestions at multiple time scales are generated.

2. The method for intelligent analysis of equipment maintenance failures based on the Internet of Things according to claim 1 is characterized in that: The step of automatically extracting entities and relationships from equipment maintenance related documents includes: Preprocess maintenance manuals, historical fault records, maintenance reports, and unstructured text; A named entity recognition method based on a combination of conditional random fields and bidirectional long short-term memory networks is used to identify and extract three core entities: equipment components, fault phenomena, and maintenance methods. Relation extraction algorithms using distant supervision and attention mechanisms to identify semantic relationships between entities; The extracted entities and relations are integrated into a unified knowledge framework and consistency checked.

3. The intelligent analysis method for equipment maintenance failure based on the Internet of Things according to claim 1 is characterized in that: The steps of detecting concept drift and mining device performance evolution patterns include: Calculate the semantic displacement of an entity at two time points. When the semantic displacement exceeds a preset threshold, it is determined that the calculated entity has concept drift. Utilize cyclic variational autoencoders to automatically extract periodic and trend patterns of knowledge evolution from temporal knowledge graphs; Construct a knowledge evolution visualization module to present the detected concept drift and evolution patterns in an intuitive way.

4. The method for intelligent analysis of equipment maintenance failures based on the Internet of Things according to claim 1 is characterized in that: The steps of predicting the future state of the equipment and assessing the risk of failure include: Based on the time series knowledge graph, the knowledge state of the device at a specific historical time point can be reconstructed; A device state transition model is constructed using a recurrent neural network. The model input is the device state vector and related knowledge graph information at the current time point, and the output is the state prediction at the future time point. Calculate the similarity between the predicted state and known failure modes and generate probability distributions for different types of failures.

5. The method for intelligent analysis of equipment maintenance failures based on the Internet of Things according to claim 1 is characterized in that: The steps of processing knowledge conflicts and updating the time series knowledge graph and prediction model include: Calculate the credibility of conflicting knowledge based on the amount of supporting evidence and source reliability; Demote or eliminate low-credibility knowledge; Adding new knowledge with high confidence; For knowledge with timeliness differences, update the time attribute instead of directly replacing it; An incremental learning method is used to update the parameters of the state transition model and the fault risk assessment model.

6. The method for intelligent analysis of equipment maintenance failures based on the Internet of Things according to claim 1 is characterized in that: The steps of generating fault risk warnings and maintenance suggestions at multiple time scales include: Generate equipment status predictions at different time scales based on the state transition model; Determine the warning level based on the matching degree between the predicted status and the failure mode and the time urgency; Automatically generate targeted maintenance suggestions based on fault risk warning results and maintenance knowledge in the time series knowledge graph; Based on the equipment's historical maintenance records, the time and resources required for maintenance are estimated to assist in formulating maintenance plans.

7. The method for intelligent analysis of equipment maintenance failures based on the Internet of Things according to claim 1 is characterized in that: Time attributes include time points, time intervals, or periodic characteristics. Time points represent the specific moment when knowledge is established, time intervals represent the start and end time when knowledge is valid, and periodic characteristics represent seasonal failure patterns.

8. The method for intelligent analysis of equipment maintenance failures based on the Internet of Things according to claim 3 is characterized in that: The cyclic variational autoencoder consists of an encoder and a decoder. The encoder encodes the temporal entity representation sequence into the distribution parameters of latent variables, and the decoder reconstructs the sequence based on the sampled latent variables and historical information.

9. An intelligent analysis system for equipment maintenance faults based on the Internet of Things, configured to execute the intelligent analysis method for equipment maintenance faults based on the Internet of Things according to any one of claims 1 to 8, characterized in that: include: Multi-source text knowledge extraction and graph construction module, used to automatically extract entities and relationships from equipment maintenance-related documents and construct an initial knowledge graph; The temporal knowledge graph construction module represents the entities and relationships in the initial knowledge graph as time functions based on the constructed initial knowledge graph to construct a temporal knowledge graph; The concept drift detection module uses the constructed temporal knowledge graph to monitor the changes of entity and relationship vectors in the knowledge embedding space over time to detect concept drift and explore the evolutionary patterns of device performance. The device status prediction module uses a recursive neural network to build a device state transition model based on the constructed time series knowledge graph and current observation data, predict the future state of the device and assess the risk of failure; The knowledge update module is used to collect maintenance personnel's feedback on the system's prediction results, handle knowledge conflicts, and update the time series knowledge graph and prediction model; The fault warning module is used to combine the real-time status data of the equipment and the updated time series knowledge graph to generate fault risk warnings and maintenance suggestions at multiple time scales.

Images