Predictive maintenance method based on deep data mining and domain knowledge fusion

Abstract

The present invention discloses a predictive maintenance method based on the fusion of deep data mining and domain knowledge, relates to the field of data processing technology, and solves the technical problem of insufficient recognition and prediction accuracy in the prior art for complex fault modes or application scenarios with high real-time requirements; the present invention comprises determining a data source to collect data, then processing the collected data, and then evaluating the quality of the data. If the evaluation is unqualified, the data is reprocessed until the evaluation is qualified and the data is divided into a training set, a verification set and a test set; a DD-DC model is selected as a framework, and then a time dimension is introduced, and then system dynamics and external factors are considered, and finally an optimization feedback mechanism is introduced to build an intelligent predictive maintenance model; domain knowledge is deeply integrated into the intelligent predictive maintenance model; the present invention improves the comprehensive perception capability of equipment status, so that the accuracy and reliability of fault prediction are significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a predictive maintenance method based on the fusion of deep data mining and domain knowledge. Background Art

[0002] In the current industrial field, predictive maintenance technology is increasingly valued. It aims to discover potential equipment failures in advance through data analysis, thereby avoiding unplanned downtime, reducing maintenance costs, and improving production efficiency. At present, there are many data-driven predictive maintenance methods on the market. Among them, the more common one is to use machine learning or deep learning algorithms to model equipment operation data to predict the future health status of the equipment. However, most of these methods rely on a single data source or a single algorithm model, ignoring the correlation and complementarity between multiple heterogeneous data sources during equipment operation, resulting in limited prediction accuracy and difficulty in adapting to complex and changing industrial environments. In addition, the existing models have low computational efficiency when processing large-scale, high-dimensional data, and it is difficult to meet industrial application scenarios with high real-time requirements.

[0003] In addition, traditional methods often fail to fully consider the dynamic changes of data in the time dimension and the dependencies between data at different levels, which limits the model's ability to identify complex failure modes and the accuracy of prediction. Summary of the invention

[0004] In order to solve the problems existing in the above-mentioned prior art, the present invention provides a predictive maintenance method based on deep data mining and domain knowledge fusion, which solves the technical problem of insufficient recognition and prediction accuracy in the prior art for complex failure modes or application scenarios with high real-time requirements.

[0005] Predictive maintenance methods based on deep data mining and domain knowledge integration include:

[0006] Step 1: Determine the data source and collect data, then process the collected data and evaluate the quality of the data. If the evaluation fails, reprocess the data until it passes the evaluation and then divide the data into training set, validation set and test set;

[0007] Step 2: Select the DD-DC model as the framework, then introduce the time dimension, consider system dynamics and external factors, and finally introduce an optimization feedback mechanism to build an intelligent predictive maintenance model;

[0008] Step 3: Deeply integrate domain knowledge into the intelligent predictive maintenance model, including converting the working principle, common failure modes and maintenance experience of the equipment into technical features and integrating them into the knowledge base of the optimized prediction model; using domain knowledge to correct the prediction results of the optimized prediction model; interpreting the output of the optimized prediction model with domain knowledge; and updating the domain knowledge base in real time as the operating environment and maintenance experience change;

[0009] Step 4: Integrate the final prediction model into the equipment monitoring system to monitor the equipment operating status in real time, predict potential failure risks, and provide feedback to maintenance personnel when abnormalities are found. The feedback information includes prediction results and maintenance decision recommendations.

[0010] Furthermore, based on the DD-DC model, the time dimension is introduced to consider the dynamic and predictive nature of the system, and the observation signal and control signal are represented as time series:

[0011] The observed signal is expressed as:

[0012] o(t)=[o1(t),o2(t),…,on(t)]

[0013] The control signal is expressed as:

[0014] u(t)=[u1(t),u2(t),…,um(t)]

[0015] Wherein, t represents time.

[0016] Furthermore, we consider the system dynamics and external factors: the system dynamics is described by a set of nonlinear differential equations and is affected by external factors e(t), where x(t) is the system state vector:

[0017]

[0018] In the DD-DC framework, the prediction model P can be expressed as a nonlinear function P, which accepts the observation signal o(t) as input and outputs the future control signal u(t+1); the calculation formula of the control signal can be expressed as:

[0019] u(t+1)=P(o(t))

[0020] Through the prediction model P, the future control signal u(t+1) can be predicted according to the current observation signal o(t), thereby realizing predictive control of the system. For predictive systems, the time dimension and prediction model can be introduced into the DD-DC framework, and the future control signal can be predicted through the prediction model, thereby realizing predictive control of the system.

[0021] Furthermore, the step 2 further comprises:

[0022] Multi-step prediction and feedback: The prediction model P not only predicts the control signal at the next moment, but also performs multi-step prediction and uses the feedback mechanism to adjust the prediction model, which is expressed as follows:

[0023] u(t+k)=P(o(t),u(t-1),…,u(tl))

[0024] Where k is the future time step and l is the length of the historical control signal considered.

[0025] Optimize control policy: Use reinforcement learning to optimize the control policy and optimize the control signal by maximizing the long-term reward, expressed as follows:

[0026]

[0027] Among them, γ is the discount factor and R is the reward coefficient;

[0028] Closed-loop control with feedback: Apply the predicted control signal to the system and observe its effect on the system state. Use feedback mechanisms to adjust the prediction model. Optimize the model parameters by minimizing the prediction error. Use gradient descent or other optimization algorithms to update the parameters of the prediction model to minimize the loss function:

[0029]

[0030] Where L represents the loss function, which is the mean square error here, and u(t+1) is the actual control signal. is the predicted control signal, m represents the dimension of the control signal vector, that is, the number of variables contained in the control signal, and i is the index used to traverse each element in the control signal vector.

[0031] Furthermore, the step 1 comprises:

[0032] Step 1.1 Determine data sources: Identify all potential data sources related to predictive maintenance, including equipment operation data, sensor data, maintenance records, failure history, etc.; determine the accessibility, completeness and accuracy of the data;

[0033] Step 1.2 Data acquisition: Extract data from identified data sources using automated data collection tools;

[0034] Step 1.3 Data integration: Integrate the data collected from different data sources to form a unified data set;

[0035] Step 1.4 Data cleaning: Remove noise, duplicates, and invalid values ​​from the data; handle missing values, such as by interpolating, estimating, or deleting records containing missing values;

[0036] Step 1.5: Data standardization: Standardize or normalize the data to ensure that data from different sources and scales can be used in the same model;

[0037] Step 1.6 Feature extraction: extract features from the raw data;

[0038] Step 1.7 Data division: Divide the data set into training set, validation set and test set;

[0039] Step 1.8 Data quality assessment: Perform quality assessment on the preprocessed data. If the data quality does not meet the requirements, return to the data cleaning and standardization steps for iterative optimization.

[0040] Step 1.9 Feedback and iteration: During the data collection and preprocessing process, establish a self-supervised feedback mechanism to collect feedback; iteratively optimize the data collection and preprocessing process based on the feedback results to improve data quality and model performance.

[0041] Further, step 1.9 includes:

[0042] Step 1.9.1 Performance monitoring and visualization: Provide real-time feedback on model performance and discover problems in the data or model by monitoring the performance of the model;

[0043] Step 1.9.2 Data quality assessment: Based on the results of performance monitoring and visualization, perform quality assessment on the preprocessed data. If the data quality does not meet the requirements, iterative optimization is required.

[0044] Step 1.9.3 Return to the data cleaning and standardization step for iterative optimization: When the data quality assessment results show that the data does not meet the requirements, return to the data cleaning and standardization step and adjust the cleaning and standardization strategies based on the assessment results;

[0045] Step 1.9.4 Self-supervised feedback mechanism: Design the entire data collection and preprocessing process into a self-supervised feedback mechanism; through continuous monitoring, evaluation and optimization, gradually improve the quality of data and the performance of the model.

[0046] Furthermore, the step 2 comprises:

[0047] Step 2.1 Determine the model framework: Select Direct Data-Driven Control, or DD-DC, neural network model as the framework of the predictive maintenance model;

[0048] Step 2.2: training a data-driven prediction model;

[0049] Step 2.3 Model adaptability considerations: Using the direct control characteristics of a single neuron in the DD-DC model over the environment, design the model to adapt to the dynamic changes and flexibility requirements of the device state, ensuring that the model can capture subtle changes in the device state;

[0050] Step 2.4 Model architecture design details: Design the architecture of the DD-DC neural network model based on domain knowledge, including determining the depth, width, and connection method between neurons of the model to build a model structure that can effectively extract data features and adapt to predictive maintenance tasks;

[0051] Step 2.5 Training process and optimization strategy: Taking advantage of the characteristics of the DD-DC model, the back propagation algorithm and gradient descent method are used to optimize the model parameters during the training process to ensure that the model can accurately capture the dynamic changes and potential patterns in the data and improve the ability to predict equipment failures;

[0052] Step 2.6 Parameter adjustment and verification: Use the validation set data and the hyperparameter optimization library to adjust the model parameters in detail to ensure model performance and verify the effectiveness of the model architecture and training strategy.

[0053] Step 2.7 Consider computing resource configuration: Given the computing requirements of the DD-DC model, configure computing resources to support the model training process to ensure that the model can be trained within the limited time and has practical application value.

[0054] Furthermore, the step 3 comprises:

[0055] Step 3.1 Knowledge integration: Based on the DD-DC neural network model, deeply integrate domain knowledge, including converting professional knowledge into rules or features that the model can understand, and further enrich the model's knowledge base;

[0056] Step 3.2 Dynamic Adjustment: Use domain knowledge to dynamically adjust the prediction results of the DD-DC neural network model. When the model predicts that a certain device is about to fail, the prediction results can be corrected based on the maintenance history and usage of the device to improve the accuracy and reliability of the prediction.

[0057] Step 3.3 Interpretability enhancement: Integrate domain knowledge to enhance the interpretability of the model. By interpreting the output of the DD-DC neural network model with domain knowledge, specific troubleshooting and repair suggestions are provided to maintenance personnel.

[0058] Step 3.4 Continuous learning and optimization: With the changes in the equipment operating environment and the accumulation of maintenance experience, it is necessary to continuously update and enrich the domain knowledge base and deeply integrate it with the DD-DC neural network model to achieve continuous optimization and upgrading of the model;

[0059] Step 3.5 Consider uncertainty: Introduce a Bayesian neural network for prediction, evaluate the uncertainty of the prediction results, and give the corresponding variance or confidence interval.

[0060] Furthermore, the step 4 comprises:

[0061] Step 4.1 System integration: Deploy the trained DD-DC neural network model that has been integrated with domain knowledge into the actual equipment monitoring system;

[0062] Step 4.2 Real-time monitoring and early warning: The deployed model will monitor the operating status of the equipment in real time and use its predictive capabilities to detect potential failure risks in advance. Once an abnormality is found, the model will immediately trigger the early warning mechanism to notify maintenance personnel to intervene in time;

[0063] Step 4.3 Maintenance decision support: The model not only provides fault warnings, but also provides maintenance personnel with specific maintenance decision suggestions based on prediction results and domain knowledge;

[0064] Step 4.4 Effect evaluation and feedback: In actual application, the prediction effect and maintenance decision of the model are continuously evaluated. By collecting the actual operation data and maintenance feedback of the equipment, the performance of the model is quantitatively evaluated, and the model is adjusted and optimized as necessary according to the evaluation results.

[0065] Step 4.5 User training and support: Provide relevant training and support to maintenance personnel to ensure that the model can play the expected role in actual maintenance work.

[0066] Furthermore, the step 4.3 of maintaining decision support includes:

[0067] Step 4.3.1 Fault prediction and analysis: The model first uses deep data mining technology to perform real-time analysis and processing of multi-source heterogeneous data during the operation of the equipment to predict the future health status of the equipment. When the model predicts that the equipment has a potential failure risk, it will immediately trigger an early warning mechanism to notify maintenance personnel to pay attention to the operating status of the equipment.

[0068] Step 4.3.2 Domain knowledge integration and interpretation: After predicting a failure, the model will further integrate domain knowledge to explain and correct the prediction results; domain knowledge includes the working principle of the equipment, common failure modes, and maintenance experience;

[0069] Step 4.3.3 Maintenance decision suggestion generation: Based on the fusion of prediction results and domain knowledge, the model generates specific maintenance decision suggestions; the suggestions include recommended maintenance time, maintenance method and required maintenance resources;

[0070] Step 4.3.4 Decision suggestion optimization and feedback: After receiving the decision suggestion, the maintenance personnel will make selections and adjustments based on the actual situation. At the same time, the model will continuously optimize and adjust the decision suggestion generation strategy based on the maintenance personnel’s feedback and the actual operating status of the equipment to improve the accuracy and practicality of the suggestion.

[0071] Step 4.3.5 Real-time monitoring and dynamic adjustment: During the maintenance decision execution process, the model will continue to monitor the operating status of the equipment in real time; if new failure risks or abnormal conditions are found, the model will promptly adjust the maintenance decision recommendations to ensure the safe and stable operation of the equipment

[0072] The beneficial effects of the present invention include:

[0073] First, by fusing multi-source heterogeneous data, the present invention fully utilizes the complementarity between different data sources, thereby improving the comprehensive perception capability of equipment status and significantly improving the accuracy and reliability of fault prediction.

[0074] Secondly, the present invention introduces domain knowledge, which enables the model to better understand and adapt to specific failure modes in different industrial scenarios, thereby enhancing the generalization ability of the model in complex and changing environments.

[0075] Finally, by implementing a dynamic feedback and optimization mechanism, the present invention enables the model to dynamically adjust its internal parameters and strategies according to the prediction results to optimize subsequent prediction and maintenance decisions, further improving the accuracy and efficiency of predictive maintenance.

[0076] In summary, the present invention solves the problems of limited prediction accuracy, poor adaptability and low computational efficiency in the prior art, and provides a more efficient, accurate and adaptable intelligent solution for predictive maintenance in the industrial field. The algorithm performs well in multiple experimental scenarios, with a prediction accuracy of more than 90% and a maximum of 97%, showing the ability to accurately predict control signals. The computational efficiency is high, and the computational time of all experiments is less than 0.08 seconds, with an average of about 0.06 seconds, which is suitable for real-time computing and control needs. The resource consumption is low, with an average score of about 3 points and a minimum of 1 point, which reduces the overall operating cost of the system. The uncertainty estimation ability is strong, and most experiments are rated high or medium, which helps to deal with uncertainty factors in practical applications. In summary, the algorithm performs well in prediction, control, computational efficiency, resource consumption and uncertainty estimation, and has high practical value and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 Flowchart of the predictive maintenance method based on deep data mining and domain knowledge fusion.

[0078] Figure 2It is a schematic diagram of the structure of the DC-DC model involved in the embodiment of the present application. DETAILED DESCRIPTION

[0079] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0080] Predictive maintenance methods based on deep data mining and domain knowledge fusion, such as Figure 1 As shown, including:

[0081] Step 1: Determine the data source and collect data, then process the collected data and evaluate the quality of the data. If the evaluation fails, reprocess the data until it passes the evaluation and then divide the data into training set, validation set and test set;

[0082] Step 2: Select the DD-DC model as the framework to build an intelligent predictive maintenance model;

[0083] Step 3: Deeply integrate domain knowledge into the intelligent predictive maintenance model, including converting the working principle, common failure modes and maintenance experience of the equipment into technical features and integrating them into the knowledge base of the optimized prediction model; using domain knowledge to correct the prediction results of the optimized prediction model; interpreting the output of the optimized prediction model with domain knowledge; and updating the domain knowledge base in real time as the operating environment and maintenance experience change;

[0084] Step 4: Integrate the final prediction model into the equipment monitoring system to monitor the equipment operating status in real time, predict potential failure risks, and provide feedback to maintenance personnel when abnormalities are found. The feedback information includes prediction results and maintenance decision recommendations.

[0085] In another embodiment, the DC-DC model is as follows Figure 2 As shown in the figure, the neuron environment is modeled as a linear dynamic system represented in discrete time state space, that is, a flowchart that combines the control system model and the sensor actuator model. It contains the following main parts:

[0086] x t+1 =Ax t +bu t

[0087] yt =Cx t

[0088] In the above formula, represents the potential state of the environment at time t, u t represents the control signal generated by the neuron, y t represents the observation value received by the neuron, and A, b and C are predefined dynamic parameters.

[0089]

[0090] In the above formula, Represents the optimal control signal and the estimated latent state variables.

[0091] If the controller does not know the parameters A, b, C, DD-DC can be used to construct the optimal control signal. In this case, the control signal is generated directly from the observations. The DD-DC controller learns to map the observations to the control signal from the history of this pairing.

[0092]

[0093] Assume that the linearly independent matrix consisting of the x and u rows on the right has full row rank, a situation called continuous excitation, and each observation-control pair can be expressed as a linear combination of k historical pairs. DD-DC solves u in the above equation by t , the control signal is computed directly from the history pairing, thus eliminating the need for explicit representation of the dynamic system (A, b, and C) and the latent state x in the controller.

[0094] In another embodiment, based on the DD-DC model, the time dimension is introduced, the dynamic and predictive properties of the system are considered, and the observation signal and the control signal are represented as a time series:

[0095] The observed signal is expressed as:

[0096] o(t)=[o1(t),o2(t),…,on(t)]

[0097] The control signal is expressed as:

[0098] u(t)=[u1(t),u2(t),…,um(t)]

[0099] Wherein, t represents time.

[0100] In another embodiment, the system dynamics and external factors are considered: the system dynamics is described by a set of nonlinear differential equations and is affected by external factors e(t), where x(t) is the system state vector:

[0101]

[0102] In the DD-DC framework, the prediction model P can be expressed as a nonlinear function P, which accepts the observation signal o(t) as input and outputs the future control signal u(t+1); the calculation formula of the control signal can be expressed as:

[0103] u(t+1)=P(o(t))

[0104] Through the prediction model P, the future control signal u(t+1) can be predicted according to the current observation signal o(t), thereby realizing predictive control of the system. For predictive systems, the time dimension and prediction model can be introduced into the DD-DC framework, and the future control signal can be predicted through the prediction model, thereby realizing predictive control of the system.

[0105] In another embodiment, the step 2 further comprises:

[0106] Multi-step prediction and feedback: The prediction model P not only predicts the control signal at the next moment, but also performs multi-step prediction and uses the feedback mechanism to adjust the prediction model, which is expressed as follows:

[0107] u(t+k)=P(o(t),u(t-1),…,u(tl))

[0108] Where k is the future time step and l is the length of the historical control signal considered.

[0109] Optimize control policy: Use reinforcement learning to optimize the control policy and optimize the control signal by maximizing the long-term reward, expressed as follows:

[0110]

[0111] Among them, γ is the discount factor and R is the reward coefficient;

[0112] Closed-loop control with feedback: Apply the predicted control signal to the system and observe its effect on the system state. Use feedback mechanisms to adjust the prediction model. Optimize the model parameters by minimizing the prediction error. Use gradient descent or other optimization algorithms to update the parameters of the prediction model to minimize the loss function:

[0113]

[0114] Where L represents the loss function, which is the mean square error here, and u(t+1) is the actual control signal. is the predicted control signal, m represents the dimension of the control signal vector, that is, the number of variables contained in the control signal, and i is the index used to traverse each element in the control signal vector.

[0115] In another embodiment, the step 1 comprises:

[0116] Step 1.1 Identify data sources: Identify all potential data sources related to predictive maintenance, including equipment operation data, sensor data, maintenance records, failure history, etc.; determine the accessibility, completeness and accuracy of the data.

[0117] Step 1.2 Data Acquisition: Use automated data collection tools (such as web crawlers, API interfaces, database queries, etc.) to extract data from identified data sources; and ensure that the data collection process complies with data privacy and security regulations.

[0118] Step 1.3 Data integration: Integrate the data collected from different data sources to form a unified data set; resolve inconsistencies and conflicts between data.

[0119] Step 1.4 Data cleaning: Remove noise, duplicates, and invalid values ​​from the data; handle missing values, such as by interpolation, estimation, or deletion of records containing missing values.

[0120] Step 1.5 Data standardization: Standardize or normalize the data to ensure that data from different sources and scales can be used in the same model; convert the data format, such as converting time series data into a format suitable for model input.

[0121] Step 1.6 Feature extraction: Extract meaningful features from the raw data, which should be relevant to the predictive maintenance task; use ViT for feature extraction.

[0122] Step 1.7 Data division: Divide the dataset into training set, validation set, and test set to ensure the generalization ability of the model; ensure the randomness and representativeness of the data during the division process.

[0123] Step 1.8 Data quality assessment: Perform quality assessment on the preprocessed data to ensure the accuracy, completeness and consistency of the data; if the data quality does not meet the requirements, return to the data cleaning and standardization steps for iterative optimization.

[0124] Step 1.9 Feedback and iteration: During the data collection and preprocessing process, establish a self-supervisory feedback mechanism, such as collecting feedback through data quality reports, model performance monitoring, etc.; iteratively optimize the data collection and preprocessing process based on the feedback results to improve data quality and model performance. Specifically:

[0125] In another embodiment, step 1.9 comprises:

[0126] Step 1.9.1 Performance monitoring and visualization: Provides real-time feedback on model performance. By monitoring the performance of the model, problems in the data or model can be discovered, thereby guiding subsequent data quality assessment.

[0127] Step 1.9.2 Data quality assessment: Based on the results of performance monitoring and visualization, perform quality assessment on the preprocessed data. If the data quality does not meet the requirements, iterative optimization is required.

[0128] Step 1.9.3 Return to the data cleaning and standardization step for iterative optimization: When the data quality assessment results show that the data does not meet the requirements, return to the data cleaning and standardization step and adjust the cleaning and standardization strategies based on the assessment results to improve data quality.

[0129] Step 1.9.4 Self-supervised feedback mechanism: The entire data collection and preprocessing process forms a self-supervised feedback mechanism. Through continuous monitoring, evaluation and optimization, the quality of data and the performance of the model are gradually improved.

[0130] Furthermore, the step 2 comprises:

[0131] Step 2.1 Core technology selection and understanding: Select the Direct Data-Driven Control (DD-DC) neural network model as the core technology of predictive maintenance, deeply understand its environmental perception ability and human brain imitation mechanism, and ensure that the model is suitable for handling equipment fault prediction tasks.

[0132] Step 2.2 Data-driven prediction model:

[0133] Step 2.2.1 Collect historical observation data: In order to train an efficient and accurate prediction model, it is necessary to first collect a large amount of historical observation data. These data cover multiple dimensions such as the operating status of the equipment, sensor readings, maintenance records, and fault history, providing a comprehensive information foundation for the model. The data collection process strictly complies with data privacy and security regulations to ensure the legality and compliance of the data.

[0134] Step 2.2.2 Select a deep learning model such as LSTM: After the data is prepared, select the long short-term memory network (LSTM) as the deep learning model. LSTM performs well in processing time series data and can capture long-term dependencies in the data, making it very suitable for equipment failure prediction tasks. By leveraging the powerful learning ability of the LSTM model, implicit regularities and abnormal patterns can be extracted from the data.

[0135] Step 2.2.3: Train the model to predict control signals: The LSTM model is trained using the collected historical observation data. During the training process, the model gradually learns the features in the data through an iterative optimization algorithm and establishes a mapping relationship between input and output. Ultimately, the trained model can accurately predict future control signals, providing strong support for predictive maintenance of equipment.

[0136] Step 2.3 Model adaptability considerations: Make full use of the direct control characteristics of a single neuron in the DD-DC model over the environment, design the model to adapt to the dynamic changes in device status and high flexibility requirements, and ensure that the model can capture subtle changes in device status.

[0137] Step 2.4 Model architecture design details: Combine domain knowledge and carefully design the architecture of the DD-DC neural network model, including determining the depth, width, and connection mode between neurons of the model, in order to build a model structure that can effectively extract data features and adapt to predictive maintenance tasks. Specifically, according to the working principle and common failure modes of the equipment, determine the core data features and variables in the model, so that the model architecture is more in line with the needs of actual application scenarios and improve the prediction accuracy and adaptability of the model.

[0138] Step 2.5 Training process and optimization strategy: Taking advantage of the characteristics of the DD-DC model, the back propagation algorithm and gradient descent method are used to optimize the model parameters during the training process to ensure that the model can accurately capture the dynamic changes and potential patterns in the data and improve the ability to predict equipment failures.

[0139] Step 2.6 Parameter adjustment and verification: Use the validation set data and the hyperparameter optimization library to adjust the model parameters in detail to ensure that the model can show stable performance on unseen data and verify the effectiveness of the model architecture and training strategy.

[0140] Step 2.7 Considerations for computing resource configuration: Given the computing requirements of the DD-DC model, high-performance computing resources should be reasonably configured to support an efficient model training process, ensuring that the model can be trained within a reasonable time and has practical application value.

[0141] Furthermore, the step 3 comprises:

[0142] Step 3.1 Knowledge integration: Based on the DD-DC neural network model, deeply integrate domain knowledge. This includes converting professional knowledge such as the working principle of the equipment, common failure modes, and maintenance experience into rules or features that the model can understand, further enriching the model's "knowledge base".

[0143] Step 3.2 Dynamic Adjustment: Use domain knowledge to dynamically adjust the prediction results of the DD-DC neural network model. For example, when the model predicts that a certain device is about to fail, the prediction results can be corrected based on the maintenance history and usage of the device to improve the accuracy and reliability of the prediction.

[0144] Step 3.3 Enhanced interpretability: Integrating domain knowledge can also enhance the interpretability of the model. By interpreting the output of the DD-DC neural network model with domain knowledge, we can more intuitively understand why the model makes such predictions, thereby providing maintenance personnel with more specific troubleshooting and repair suggestions.

[0145] Step 3.4 Continuous learning and optimization: The integration of domain knowledge is not a one-time task, but a continuous process. With the changes in the equipment operating environment and the accumulation of maintenance experience, it is necessary to continuously update and enrich the domain knowledge base and deeply integrate it with the DD-DC neural network model to achieve continuous optimization and upgrading of the model.

[0146] Step 3.5 Consider uncertainty: Bayesian neural network prediction: Introduce Bayesian neural network for prediction. This method can evaluate the uncertainty of the prediction results and give the corresponding variance or confidence interval.

[0147] In another embodiment, step 4 comprises:

[0148] Step 4.1 System Integration: Deploy the trained DD-DC neural network model that has been integrated with domain knowledge into the actual equipment monitoring system. This involves interfacing with the existing system, setting up data transmission protocols, and real-time feedback of model prediction results.

[0149] Step 4.2 Real-time monitoring and early warning: The deployed model will monitor the operating status of the equipment in real time and use its predictive capabilities to detect potential failure risks in advance. Once an abnormality is found, the model will immediately trigger the early warning mechanism to notify maintenance personnel to intervene in time.

[0150] Step 4.3 Maintenance decision support: The model not only provides fault warnings, but also provides maintenance personnel with specific maintenance decision suggestions based on prediction results and domain knowledge. This includes recommended maintenance time, maintenance methods, and required maintenance resources, etc., to help maintenance personnel perform equipment maintenance more scientifically and efficiently.

[0151] Step 4.4 Effect evaluation and feedback: In actual application, the prediction effect and maintenance decision of the model are continuously evaluated. By collecting the actual operation data and maintenance feedback of the equipment, the performance of the model is quantitatively evaluated, and the model is adjusted and optimized as necessary according to the evaluation results.

[0152] Step 4.5 User training and support: In order to ensure that the model can play the greatest role in actual maintenance work, maintenance personnel need to be trained and supported. This includes how to use the model, interpret warning information, and execute maintenance decisions, to ensure that maintenance personnel can fully utilize the prediction and decision support provided by the model.

[0153] In another embodiment, the maintenance decision support includes:

[0154] Step 4.3.1 Fault prediction and analysis:

[0155] (1) The model first uses deep data mining technology to perform real-time analysis and processing of multi-source heterogeneous data during equipment operation to predict the future health status of the equipment.

[0156] (2) When the model predicts that a device has a potential failure risk, it will immediately trigger an early warning mechanism to notify maintenance personnel to pay attention to the operating status of the device.

[0157] Step 4.3.2 Domain knowledge integration and interpretation:

[0158] (1) After predicting a failure, the model will further integrate domain knowledge to explain and correct the prediction results.

[0159] (2) Domain knowledge includes the working principle of the equipment, common failure modes, maintenance experience, etc. This information helps the model to more accurately understand the causes and possible impacts of equipment failures.

[0160] Step 4.3.3 Maintenance decision suggestion generation:

[0161] (1) Based on the fusion of prediction results and domain knowledge, the model generates specific maintenance decision recommendations.

[0162] (2) The recommendations include the recommended maintenance time (such as immediate maintenance, regular maintenance, etc.), maintenance method (such as replacing parts, adjusting parameters, etc.), and required maintenance resources (such as human, material, financial, etc.).

[0163] Step 4.3.4 Decision suggestion optimization and feedback:

[0164] (1) After receiving decision suggestions, maintenance personnel can make choices and adjustments based on actual conditions.

[0165] (2) At the same time, the model will continuously optimize and adjust the decision-making recommendation generation strategy based on the feedback from maintenance personnel and the actual operating status of the equipment to improve the accuracy and practicality of the recommendations.

[0166] Step 4.3.5 Real-time monitoring and dynamic adjustment:

[0167] (1) During the maintenance decision execution process, the model will continuously monitor the operating status of the equipment in real time.

[0168] (2) If new failure risks or abnormal conditions are discovered, the model will promptly adjust maintenance decision recommendations to ensure safe and stable operation of the equipment.

[0169] In summary, the model can accurately predict the health status of equipment through the integration of deep data mining and domain knowledge, and provide maintenance personnel with scientific and efficient maintenance decision support. This not only helps to reduce equipment maintenance costs and improve production efficiency, but also effectively prolongs the service life of equipment and reduces the occurrence of unexpected downtime.

[0170] Specifically, the simulation effect of this method is shown in Table 1:

[0171] Table 1 Simulation experiment data table

[0172]

[0173] Among them: Prediction accuracy: indicates the accuracy of the prediction model for the control signal.

[0174] Control effect score: Evaluate the actual impact of the control strategy on the system status. The higher the score, the better the control effect.

[0175] Computation time: The time required to perform prediction and control calculations.

[0176] Resource consumption score: evaluates the consumption of computing resources (such as CPU and memory) during the execution of the algorithm. The lower the score, the less resource consumption.

[0177] Uncertainty estimation accuracy: Evaluates the accuracy of the algorithm's estimation of prediction uncertainty, which is divided into three levels: high, medium, and low.

[0178] This method has shown good performance in multiple experimental scenarios. From the perspective of prediction accuracy, the algorithm has a high degree of prediction accuracy for control signals, and the prediction accuracy of most experimental numbers is above 90%. At the same time, the control effect score is generally high, indicating that the algorithm can effectively control the system state and achieve the expected control effect. In terms of computing time, the algorithm has shown high efficiency, and the computing time of all experiments has been kept at a low level. This is due to the optimization design of the algorithm and the implementation of real-time computing strategies; the algorithm has also shown good performance in resource consumption, and the resource consumption scores of most experiments are low, indicating that the algorithm consumes less computing resources during execution; from the perspective of uncertainty estimation accuracy, the algorithm has a strong ability to estimate prediction uncertainty, and most experiments have received high or medium ratings. This helps to better consider and deal with uncertainty factors in practical applications and improve the stability and reliability of control.

[0179] In summary, this method has shown good performance in prediction accuracy, control effect, calculation time, resource consumption and uncertainty estimation, and has high practical value and application prospects.

[0180] The above-mentioned embodiments only express the specific implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the protection scope of the present application. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the technical solution concept of the present application, and these all belong to the protection scope of the present application.

Claims

1. A predictive maintenance method based on deep data mining and domain knowledge fusion, characterized by: include: Step 1: Determine the data source and collect data, then process the collected data and evaluate the quality of the data. If the evaluation fails, reprocess the data until it passes the evaluation and then divide the data into training set, validation set and test set; Step 2: Select the DD-DC model as the framework, then introduce the time dimension, consider system dynamics and external factors, and finally introduce an optimization feedback mechanism to build an intelligent predictive maintenance model; Step 3: Deeply integrate domain knowledge into the intelligent predictive maintenance model, including converting the working principle, common failure modes and maintenance experience of the equipment into technical features and integrating them into the knowledge base of the optimized prediction model; using domain knowledge to correct the prediction results of the optimized prediction model; interpreting the output of the optimized prediction model with domain knowledge; and updating the domain knowledge base in real time as the operating environment and maintenance experience change; Step 4: Integrate the final prediction model into the equipment monitoring system to monitor the equipment operating status in real time, predict potential failure risks, and provide feedback to maintenance personnel when abnormalities are found. The feedback information includes prediction results and maintenance decision recommendations.

2. The predictive maintenance method based on deep data mining and domain knowledge fusion according to claim 1 is characterized in that: Based on the DD-DC model, the time dimension is introduced, the dynamic and predictive nature of the system is considered, and the observation signal and control signal are represented as time series: The observed signal is expressed as: o(t)=[o1(t),o2(t),…,on(t)] The control signal is expressed as: u(t)=[u1(t),u2(t),…,um(t)] Wherein, t represents time.

3. The predictive maintenance method based on deep data mining and domain knowledge fusion according to claim 1 is characterized in that: Considering system dynamics and external factors: The system dynamics is described by a set of nonlinear differential equations and is affected by external factors e(t), where x(t) is the system state vector: In the DD-DC framework, the prediction model P can be expressed as a nonlinear function P, which accepts the observation signal o(t) as input and outputs the future control signal u(t+1); the calculation formula of the control signal can be expressed as: u(t+1)=P(o(t)) Through the prediction model P, the future control signal u(t+1) can be predicted according to the current observation signal o(t), thereby realizing predictive control of the system. For predictive systems, the time dimension and prediction model can be introduced into the DD-DC framework, and the future control signal can be predicted through the prediction model, thereby realizing predictive control of the system.

4. The predictive maintenance method based on deep data mining and domain knowledge fusion according to claim 1 is characterized in that: The step 2 also includes: Multi-step prediction and feedback: The prediction model P not only predicts the control signal at the next moment, but also performs multi-step prediction and uses the feedback mechanism to adjust the prediction model, which is expressed as follows: u(t+k)=P(o(t),u(t-1),…,u(tl)) Where k is the future time step and l is the length of the historical control signal considered; Optimize control policy: Use reinforcement learning to optimize the control policy and optimize the control signal by maximizing the long-term reward, expressed as follows: Among them, γ is the discount factor and R is the reward coefficient; Closed-loop control with feedback: Apply the predicted control signal to the system and observe its effect on the system state. Use feedback mechanisms to adjust the prediction model. Optimize the model parameters by minimizing the prediction error. Use gradient descent or other optimization algorithms to update the parameters of the prediction model to minimize the loss function: Where L represents the loss function, which is the mean square error here, and u(t+1) is the actual control signal. is the predicted control signal, m represents the dimension of the control signal vector, that is, the number of variables contained in the control signal, and i is the index used to traverse each element in the control signal vector.

5. The predictive maintenance method based on deep data mining and domain knowledge fusion according to claim 1 is characterized in that: The step 1 comprises: Step 1.1 Identify data sources: Identify all potential data sources related to predictive maintenance, including equipment operation data, sensor data, maintenance records, and failure history; determine the accessibility, completeness, and accuracy of the data; Step 1.2 Data acquisition: Extract data from identified data sources using automated data collection tools; Step 1.3 Data integration: Integrate the data collected from different data sources to form a unified data set; Step 1.4 Data cleaning: Remove noise, duplicates, and invalid values ​​from the data; handle missing values, such as by interpolating, estimating, or deleting records containing missing values; Step 1.5: Data standardization: Standardize or normalize the data to ensure that data from different sources and scales can be used in the same model; Step 1.6 Feature extraction: extract features from the raw data; Step 1.7 Data division: Divide the data set into training set, validation set and test set; Step 1.8 Data quality assessment: Perform quality assessment on the preprocessed data. If the data quality does not meet the requirements, return to the data cleaning and standardization steps for iterative optimization. Step 1.9 Feedback and iteration: During the data collection and preprocessing process, establish a self-supervised feedback mechanism to collect feedback; iteratively optimize the data collection and preprocessing process based on the feedback results to improve data quality and model performance.

6. The predictive maintenance method based on deep data mining and domain knowledge fusion according to claim 5 is characterized in that: Step 1.9 includes: Step 1.9.1 Performance monitoring and visualization: Provide real-time feedback on model performance and discover problems in the data or model by monitoring the performance of the model; Step 1.9.2 Data quality assessment: Based on the results of performance monitoring and visualization, perform quality assessment on the preprocessed data. If the data quality does not meet the requirements, iterative optimization is required. Step 1.9.3 Return to the data cleaning and standardization step for iterative optimization: When the data quality assessment results show that the data does not meet the requirements, return to the data cleaning and standardization step and adjust the cleaning and standardization strategies based on the assessment results; Step 1.9.4 Self-supervised feedback mechanism: Design the entire data collection and preprocessing process into a self-supervised feedback mechanism; through continuous monitoring, evaluation and optimization, gradually improve the quality of data and the performance of the model.

7. The predictive maintenance method based on deep data mining and domain knowledge fusion according to claim 1 is characterized in that: The step 2 comprises: Step 2.1 Determine the model framework: Select Direct Data-Driven Control, or DD-DC, neural network model as the framework of the predictive maintenance model; Step 2.2: training a data-driven prediction model; Step 2.3 Model adaptability considerations: Using the direct control characteristics of a single neuron in the DD-DC model over the environment, design the model to adapt to the dynamic changes and flexibility requirements of the device state, ensuring that the model can capture subtle changes in the device state; Step 2.4 Model architecture design details: Design the architecture of the DD-DC neural network model based on domain knowledge, including determining the depth, width, and connection method between neurons of the model to build a model structure that can effectively extract data features and adapt to predictive maintenance tasks; Step 2.5 Training process and optimization strategy: Taking advantage of the characteristics of the DD-DC model, the back propagation algorithm and gradient descent method are used to optimize the model parameters during the training process to ensure that the model can accurately capture the dynamic changes and potential patterns in the data and improve the ability to predict equipment failures; Step 2.6 Parameter adjustment and verification: Use the validation set data and the hyperparameter optimization library to adjust the model parameters in detail to ensure model performance and verify the effectiveness of the model architecture and training strategy. Step 2.7 Consider computing resource configuration: Given the computing requirements of the DD-DC model, configure computing resources to support the model training process to ensure that the model can be trained within the limited time and has practical application value.

8. The predictive maintenance method based on deep data mining and domain knowledge fusion according to claim 1, characterized in that: The step 3 comprises: Step 3.1 Knowledge integration: Based on the DD-DC neural network model, deeply integrate domain knowledge, including converting professional knowledge into rules or features that the model can understand, and further enrich the model's knowledge base; Step 3.2 Dynamic Adjustment: Use domain knowledge to dynamically adjust the prediction results of the DD-DC neural network model; when the model predicts that a certain device is about to fail, the prediction results are corrected based on the maintenance history and usage of the device; Step 3.3 Interpretability enhancement: Integrate domain knowledge to enhance the interpretability of the model. By interpreting the output of the DD-DC neural network model with domain knowledge, specific troubleshooting and repair suggestions are provided to maintenance personnel. Step 3.4 Continuous learning and optimization: With the changes in the equipment operating environment and the accumulation of maintenance experience, it is necessary to continuously update and enrich the domain knowledge base and deeply integrate it with the DD-DC neural network model to achieve continuous optimization and upgrading of the model; Step 3.5 Consider uncertainty: Introduce a Bayesian neural network for prediction, evaluate the uncertainty of the prediction results, and give the corresponding variance or confidence interval.

9. The predictive maintenance method based on deep data mining and domain knowledge fusion according to claim 1, characterized in that: The step 4 comprises: Step 4.1 System integration: Deploy the trained DD-DC neural network model that has been integrated with domain knowledge into the actual equipment monitoring system; Step 4.2 Real-time monitoring and early warning: The deployed model will monitor the operating status of the equipment in real time and use its predictive capabilities to detect potential failure risks in advance. Once an abnormality is found, the model will immediately trigger the early warning mechanism to notify maintenance personnel to intervene in time; Step 4.3 Maintenance decision support: The model not only provides fault warnings, but also provides maintenance personnel with specific maintenance decision suggestions based on prediction results and domain knowledge; Step 4.4 Effect evaluation and feedback: In actual application, the prediction effect and maintenance decision of the model are continuously evaluated. By collecting the actual operation data and maintenance feedback of the equipment, the performance of the model is quantitatively evaluated, and the model is adjusted and optimized as necessary according to the evaluation results. Step 4.5 User training and support: Provide relevant training and support to maintenance personnel to ensure that the model can play the expected role in actual maintenance work.

10. The predictive maintenance method based on deep data mining and domain knowledge fusion according to claim 9, characterized in that: The step 4.3 of maintaining decision support includes: Step 4.3.1 Fault prediction and analysis: The model first uses deep data mining technology to perform real-time analysis and processing of multi-source heterogeneous data during the operation of the equipment to predict the future health status of the equipment. When the model predicts that the equipment has a potential failure risk, it will immediately trigger an early warning mechanism to notify maintenance personnel to pay attention to the operating status of the equipment. Step 4.3.2 Domain knowledge integration and interpretation: After predicting a failure, the model will further integrate domain knowledge to explain and correct the prediction results; domain knowledge includes the working principle of the equipment, common failure modes, and maintenance experience; Step 4.3.3 Maintenance decision suggestion generation: Based on the fusion of prediction results and domain knowledge, the model generates specific maintenance decision suggestions; the suggestions include recommended maintenance time, maintenance method and required maintenance resources; Step 4.3.4 Decision suggestion optimization and feedback: After receiving the decision suggestion, the maintenance personnel will make selections and adjustments based on the actual situation. At the same time, the model will continuously optimize and adjust the decision suggestion generation strategy based on the maintenance personnel’s feedback and the actual operating status of the equipment to improve the accuracy and practicality of the suggestion. Step 4.3.5 Real-time monitoring and dynamic adjustment: During the maintenance decision execution process, the model will continue to monitor the operating status of the equipment in real time; if new failure risks or abnormal conditions are found, the model will promptly adjust the maintenance decision recommendations to ensure the safe and stable operation of the equipment.

Images