WO2021042687A1

WO2021042687A1 - Method and apparatus for improving adaptability of predictive maintenance model

Info

Publication number: WO2021042687A1
Application number: PCT/CN2020/077873
Authority: WO
Inventors: 鲍亭文; 金超
Original assignee: 北京天泽智云科技有限公司
Priority date: 2019-09-03
Filing date: 2020-03-05
Publication date: 2021-03-11
Also published as: CN110533251A; CN110533251B

Abstract

Disclosed are a method and apparatus for improving the adaptability of a predictive maintenance model. The method comprises: establishing an initial predictive maintenance model, and performing system maintenance by using the initial predictive maintenance model; acquiring system data in real time after the system is started, marking the data, and recording abnormal data and tag information; after the amount of the acquired data reaches a first set value, determining, according to the type of the initial predictive maintenance model, whether model conversion is required; if yes, performing training by using the acquired data to obtain a stable predictive maintenance model, and replacing the initial predictive maintenance model with the stable predictive maintenance model to perform system maintenance; after a model update trigger condition is met, updating the stable predictive maintenance model. According to the present invention, the predictive maintenance model can have better adaptability, and the accuracy and stability of the predictive maintenance model can be improved.

Description

Method and device for improving adaptability of predictive maintenance model

Technical field

The invention relates to the field of system maintenance, in particular to a method and a device for improving the adaptive capability of a predictive maintenance model.

Background technique

PHM (Prognostic and Health Management, fault prediction and health management) is a brand-new health management solution proposed by comprehensively using the latest research results of modern information technology and artificial intelligence technology. It is widely used in various fields. At present, in the maintenance of industrial systems, predictive maintenance models based on non-high-frequency system data such as SCADA (Supervisory Control And Data Acquisition) have gradually become a hot spot for research and development in recent years.

Existing technologies all focus on the design and optimization of a certain fault early warning model structure. These technical solutions can meet the needs of fault warning under their specific data and technical conditions, but generally lack the design of the applicability of the model under different data and technical conditions and the design of the growth path of the model under different development stages, making these technologies It cannot be applied in the full life cycle of the model.

In addition, most of the existing technical solutions are based on the assumption that sufficient data and labels are available. However, in the actual development process, there are often many influencing factors: such as insufficient data, lack of labels, brand-new models, environments, and working conditions, which make these technical solutions unable to be implemented or used in the full life cycle. For example, in the wind power industry, when an existing technical solution is applied to a brand-new wind farm, it often takes a certain amount of time (months or even a year) to reach the amount of training data required for the stable operation of the model. Predictive models driven by only a small amount of training data will lead to unstable model results or even completely unusable. In actual scenarios, it is often impossible to wait until the data is sufficient before being used online. For another example, when some data collection conditions are changed, some technical solutions are no longer the optimal solutions under the changed data conditions. At this time, it is often necessary to adjust the model architecture to break through the bottleneck of the existing model effect and continue to improve the performance of the model. Choosing a model architecture that is applicable under new conditions can achieve the effect of improving the performance of the model at this stage, but often the domain knowledge accumulated in previous model development cannot be precipitated into the new model architecture, or the model structure changes greatly, making a single development work While the amount is increasing, it may not be used in future scenarios, and the reusability is insufficient.

Summary of the invention

The embodiment of the present invention provides a method and device for improving the adaptability of a predictive maintenance model, so that the predictive maintenance model has better adaptability, and the accuracy and stability of the predictive maintenance model are improved.

To this end, the present invention provides the following technical solutions:

A method for improving the adaptability of a predictive maintenance model, the method comprising:

Collect system data in real time after the system is started, mark the data, and record abnormal data and label information;

Establish an initial predictive maintenance model, and use the initial predictive maintenance model to perform system maintenance;

After the amount of collected data reaches the first set value, determine whether a model conversion is required according to the type of the initial predictive maintenance model;

If so, use the collected data to train to obtain a stable predictive maintenance model, and replace the initial predictive maintenance model with the stable predictive maintenance model for system maintenance;

After the model update trigger condition is satisfied, the stable predictive maintenance model is updated, and the updated predictive maintenance model is used for system maintenance.

Optionally, the establishment of an initial predictive maintenance model includes:

If the mechanism parameters corresponding to the system can be obtained, a mechanism-based residual model is established, and the mechanism-based residual model is used as an initial predictive maintenance model;

Otherwise, judge whether the abnormal performance of different devices in the system is similar;

If yes, establish a cluster benchmarking model, and use the cluster benchmarking model as an initial predictive maintenance model;

Otherwise, judge whether there is a trained predictive maintenance model of equipment of the same model as the equipment in this system in other systems;

If yes, perform migration learning on the predictive maintenance model of the equipment of the same model to obtain the migration learning model, and use the migration learning model as the initial predictive maintenance model of the equipment in the system;

Otherwise, establish a rule-based model and use the rule-based model as the initial predictive maintenance model.

Optionally, the determining whether a model conversion is required according to the type of the initial predictive maintenance model includes:

If the initial predictive maintenance model is a mechanism-based residual model, or a cluster benchmarking model, or a migration learning model, it is determined that no model conversion is required;

If the initial predictive maintenance model is a rule-based model, it is determined that a model conversion is required.

Optionally, the training of using the collected data to obtain a stable predictive maintenance model includes:

If the number of recorded tag information does not reach the second set value, use the collected data to train to obtain a data-driven self-aligned residual model, and use the data-driven self-aligned residual model as a stable predictive Maintenance model

If the number of recorded label information reaches the second set value, use the collected data and the label information to train to obtain a supervised machine learning model, and use the supervised machine learning model as a stable predictive Maintain the model.

Optionally, the updating the stable predictive maintenance model after the model update trigger condition is met includes:

If the stable predictive maintenance model is a data-driven self-aligned residual model, after reaching the update cycle, or the equipment operating condition changes, or the model accuracy rate drops to a set level, the stable predictability Maintain the model for updates;

If the stable predictive maintenance model is a supervised machine learning model, then after the number of newly added tag information reaches the third set value, or after the newly added abnormal data reaches the set threshold, use the newly collected Data to update the stable predictive maintenance model;

If the stable predictive maintenance model is a migration learning model, after the amount of newly added data reaches the fourth set value, the stable predictive maintenance model is updated with the newly collected data.

Optionally, the method further includes:

After the model upgrade conditions are met, the current predictive maintenance model used for system maintenance is upgraded, and the upgraded predictive maintenance model is used for system maintenance.

Optionally, the method further includes: recording the types of collected data;

After the model upgrade conditions are met, upgrading the predictive maintenance model currently used for system maintenance includes:

If the predictive maintenance model currently used for system maintenance is the data-driven self-aligned residual model, when there are new types of newly collected data, the input parameters of the data-driven self-aligned residual model Adding the new type of data to training, or upgrading the data-driven self-standard residual model to a supervised model after the recorded label information has grown from scratch;

If the predictive maintenance model currently used for system maintenance is the supervised machine learning model, when the newly collected data has a new type, the new type is added to the input parameters of the supervised machine learning model Data for training.

A device for improving the adaptability of a predictive maintenance model, the device comprising: a data acquisition module, a data processing module, an initial model establishment module, a model conversion judgment module, a stable model establishment module, a model update module, and a system maintenance module;

The data collection module is used to collect system data in real time after the system is started;

The data processing module is used to mark the data and record abnormal data and label information;

The initial model establishment module is used to establish an initial predictive maintenance model;

The system maintenance module is configured to use the initial predictive maintenance model to perform system maintenance;

The model conversion judgment module is configured to determine whether a model conversion needs to be performed according to the type of the initial predictive maintenance model after the amount of data collected by the data collection module reaches a first set value; if so, notify the office The stable model establishment module establishes a stable predictive maintenance model;

The stable model establishment module is used to train to obtain a stable predictive maintenance model by using the collected data;

Correspondingly, the system maintenance module is also used to replace the initial predictive maintenance model with the stable predictive maintenance model for system maintenance;

The model update module is used to update the stable predictive maintenance model after a model update trigger condition is met;

Correspondingly, the system maintenance module is further configured to use the updated predictive maintenance model to perform system maintenance after the model update module updates the stable predictive maintenance model.

Optionally, the initial model establishment module includes:

The mechanism-based residual model establishment unit is used to establish a mechanism-based residual model when the mechanism parameters corresponding to the system can be obtained, and use the mechanism-based residual model as an initial predictive maintenance model;

The cluster benchmarking model establishment unit is used to establish a cluster benchmarking model when the mechanism parameters corresponding to the system cannot be obtained and the abnormal conditions of different devices in the system are similar, and the cluster benchmarking model is used as the initial prediction Sexual maintenance model;

The transfer learning model establishment unit is used when the mechanism parameters corresponding to the system cannot be obtained, and the performance of abnormal conditions of different devices is not similar, and there are trained ones in other systems that belong to the same model as the devices in this system. In the predictive maintenance model of the equipment, the migration learning model is obtained by performing migration learning on the predictive maintenance model of the equipment of the same model, and the migration learning model is used as the initial predictive maintenance model of the equipment in the system;

The rule-based model establishment unit is used when the mechanism parameters corresponding to the system cannot be obtained, and the abnormal performance of different equipment is not similar, and there is no trained equipment in other systems that belongs to the same machine as the equipment in this system. In the case of a predictive maintenance model of a type of equipment, a rule-based model is established, and the rule-based model is used as the initial predictive maintenance model.

Optionally, the model conversion judgment module is specifically configured to determine that no model conversion is required when the initial predictive maintenance model is a mechanism-based residual model, a cluster benchmarking model, or a migration learning model; When the initial predictive maintenance model is a rule-based model, it is determined that a model conversion is required.

Optionally, the stable model establishment module includes:

The first model establishment unit is used to retrain the ruled model with the collected data when the number of recorded tag information does not reach the second set value, to obtain a data-driven self-standard residual model, and Use the data-driven self-aligned residual model as a stable predictive maintenance model;

The second model establishment unit is configured to use the collected data and the label information to train to obtain a supervised machine learning model when the number of recorded label information reaches a second set value, and to combine the supervised machine learning model with the The machine learning model serves as a stable predictive maintenance model.

Optionally, the model update module includes:

The first update unit is used for when the stable predictive maintenance model is a data-driven self-standard residual model, after the update period is reached, or the equipment operating condition changes, or the accuracy of the model drops to a set level, Update the stable predictive maintenance model;

The second update unit is used for when the stable predictive maintenance model is a supervised machine learning model, when the number of newly added tag information reaches the third set value, or when the newly added abnormal data reaches the set value After the threshold, use the newly collected data to update the stable predictive maintenance model;

The third update unit is used to update the stable predictive maintenance model with the newly collected data when the stable predictive maintenance model is a migration learning model, after the amount of newly added data reaches the fourth set value Update.

Optionally, the device further includes:

The model upgrade module is used to upgrade the predictive maintenance model currently used for system maintenance after meeting the model upgrade conditions;

Correspondingly, the system maintenance module is also used to perform system maintenance using the upgraded predictive maintenance model after the model upgrade module upgrades the predictive maintenance model currently used for system maintenance.

Optionally, the data processing module is also used to record the types of collected data;

The model upgrade module includes:

The first upgrade unit is used for the data-driven self-aligned residual model when the predictive maintenance model currently used for system maintenance is the data-driven self-aligned residual model, if the newly collected data has a new type, the data-driven self-aligned Add the new type of data to the input parameters of the standard residual model for training; if the recorded label information grows from scratch, upgrade the data-driven self-standard residual model to a supervised model;

The second upgrade unit is used for the current predictive maintenance model used for system maintenance is the supervised machine learning model, and when the newly collected data has a new type, the input parameters of the supervised machine learning model Add the new type of data for training.

An electronic device, including: one or more processors and memories;

The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the aforementioned method.

A readable storage medium having instructions stored thereon, and the instructions are executed to implement the aforementioned method.

The method and device for improving the adaptability of the predictive maintenance model provided by the embodiments of the present invention use the predictive maintenance model to maintain the system, and adjust the predictive maintenance model adaptively according to the different stages of system operation. Specifically, in At the initial stage of the system startup, due to the lack of collected data and recorded tag information, an initial predictive maintenance model was established by means other than data-driven. With the operation of the system, the amount of collected data reached a certain amount. The collected data is trained to obtain a stable predictive maintenance model, which can then replace some specific types of initial predictive maintenance models to improve the accuracy of prediction and make the system better maintained. After the model update trigger condition is met, the stable predictive maintenance model is updated, so that the predictive maintenance model has better adaptability and meets various requirements of the system equipment operating conditions.

Further, after meeting the model upgrade conditions, the current predictive maintenance model used for system maintenance is upgraded to better improve the accuracy and stability of the predictive maintenance model.

The method and device for improving the adaptability of the predictive maintenance model provided by the embodiment of the present invention adopts a suitable strategy design for each stage of the predictive maintenance model life cycle, so that the predictive maintenance model at each stage can be better. Predictive maintenance effectiveness.

Description of the drawings

In order to explain the embodiments of the present application or the technical solutions in the prior art more clearly, the following will briefly introduce the drawings that need to be used in the embodiments. Obviously, the drawings in the following description are only recorded in the present invention. For some of the embodiments of, for those of ordinary skill in the art, other drawings may be obtained based on these drawings.

FIG. 1 is a flowchart of a method for improving the adaptability of a predictive maintenance model according to an embodiment of the present invention;

2 is a schematic diagram of the entire life cycle of a predictive maintenance model used for system maintenance in an embodiment of the present invention;

FIG. 3 is a structural block diagram of a device for improving the adaptability of a predictive maintenance model according to an embodiment of the present invention;

Fig. 4 is another structural block diagram of a device for improving the adaptability of a predictive maintenance model according to an embodiment of the present invention.

detailed description

In order to enable those skilled in the art to better understand the solutions of the embodiments of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and implementation manners.

The method and device for improving the adaptability of the predictive maintenance model provided by the embodiments of the present invention use the predictive maintenance model to maintain the system, and the predictive maintenance model is adaptively adjusted according to the different stages of system operation, so that each stage All of the predictive maintenance models can get better predictive maintenance effects.

As shown in FIG. 1, it is a flowchart of a method for improving the adaptability of a predictive maintenance model according to an embodiment of the present invention, which includes the following steps:

Step 101: Collect system data in real time after the system is started, mark the data, and record abnormal data and tag information.

Step 102: Establish an initial predictive maintenance model, and use the initial predictive maintenance model to perform system maintenance.

Since in the initial stage, within a certain period of time after the system is started, the amount of data collected, as well as the abnormal data and label information in it is small, it is difficult to meet the offline training requirements of some commonly used data-driven models. Therefore, at this stage, in consideration of the feasibility of the balance model and the accuracy of the result, the embodiment of the present invention can adopt the following two technical paths to construct the initial predictive maintenance model:

1) Models that require offline training: commonly used predictive maintenance models that require offline training include mechanism-based residual models, that is, models built based on the mechanical physical structure of components, usually regression models, because their coefficients reflect specific mechanical physics Nature, only a small amount of normal running data can be trained. In addition, it is also possible to directly use the trained model with generalization characteristics of the same model under similar operating conditions through transfer learning.

2) Models that do not require offline training: For example, they can include but are not limited to cluster benchmarking models and rule-based models. The cluster benchmarking model refers to a model established by comparing the current status of similar equipment under the same working condition and judging the failure probability according to whether the equipment parameters deviate from the cluster status. The rule-based model refers to a model established by a method of comprehensively judging fault-related parameters and their trends.

The specific process of establishing an initial predictive maintenance model is as follows:

First, consider whether a mechanism model can be established; if so, establish a mechanism-based residual model, and use the mechanism-based residual model as an initial predictive maintenance model; specifically, if the mechanism corresponding to the system can be obtained Parameter, the mechanism model can be established; otherwise, the mechanism model cannot be established, and only other models can be considered. The mechanism-based residual model may specifically include, but is not limited to, any one or more of the following models: multiple linear regression model, multiple nonlinear regression model, in addition, in the process of model establishment, it can also be combined with Kalman filtering, etc. Dynamically adapt to the system residuals.

Secondly, it is judged whether the abnormal performance of different equipment in the system is similar; if so, a cluster benchmarking model is established, and the cluster benchmarking model is used as an initial predictive maintenance model. Of course, in this case, there needs to be a device cluster in the system, that is, the target to be monitored is a cluster, and the deviation of the relative cluster under the same working condition is the characterization of the failure that needs to be predicted or the failure is related to the degree of deviation . The principle of cluster benchmarking is to compare the difference between the state of a single device and the average state of the device cluster to determine abnormal trends.

Thirdly, if there is a trained predictive maintenance model of equipment of the same model as the equipment in this system in other systems, then the predictive maintenance model of the same model of equipment can be migrated and learned to obtain Transfer learning model, and use the transfer learning model as the initial predictive maintenance model of the equipment in the system. That is to say, when the equipment in the system is similar to the scene and model of a certain equipment in other scenes, the model that has been trained in that scene can be used as a startup model in a short time to make up for insufficient data training. For example, the wind turbine of Plain Wind Farm A requires a cold start model with abnormal generator temperature, while the machine learning models with abnormal generator temperature in other projects are based on the same wind turbine model and are located in Plain Wind Farm B with a similar climate. It can be considered to directly use this model of wind field B as the model of plain wind field A. The model used in transfer learning can be an unsupervised model, a regression model, or a supervised model.

Finally, if none of the above models are feasible, you can also consider establishing a rule-based model, and use the rule-based model as an initial predictive maintenance model. Specifically, some industry expert knowledge or control logic can be used to form corresponding judgment rules.

The difference between the rule-based model and the cluster benchmarking model is that the rules are based on expert knowledge and may not be able to cover all abnormalities under all working conditions.

Step 103: It is judged whether the amount of collected data reaches the first set value; if it is, step 104 is executed; otherwise, step 103 is returned.

Step 104: Determine whether a model conversion is required according to the type of the initial predictive maintenance model; if so, perform step 105; otherwise, return to step 104. At this time, continue to use the initial predictive maintenance model for system maintenance.

Specifically, if the initial predictive maintenance model is a mechanism-based residual model, a cluster benchmarking model, or a migration learning model, no model conversion is required; if the initial predictive maintenance model is a rule-based model, It is determined that a model conversion is required.

Step 105: Use the collected data to train to obtain a stable predictive maintenance model, and replace the initial predictive maintenance model with the stable predictive maintenance model for system maintenance.

At this stage, the amount of data collected by the system has met the training requirements of the model, and the data-driven model can be run stably. Therefore, a stable predictive maintenance model is established using a data-driven approach, that is, the initial predictive maintenance model is transformed into stable predictive maintenance model.

Specifically, depending on whether there is sufficient label information, there are mainly the following two types of models:

1) Unsupervised learning: It is mostly used when the data is sufficient but the label information is insufficient. The unsupervised models of predictive maintenance mainly include mechanism models, self-standard residual models, and cluster-standard models. Among them, the self-standardized residual model is a model established by using its own historical health status to predict the current equipment operating parameters in a healthy state and compare them with the measured values to determine the probability of failure.

2) Supervised learning: It is used when the data and label information are sufficient. There are mainly classification models or neural network models.

In the embodiment of the present invention, when the amount of collected data reaches the first set value, and the number of recorded tag information does not reach the second set value, the collected data can be used for training to obtain a data-driven self-calibration residual Model, and use the data-driven self-aligned residual model as a stable predictive maintenance model. The data-driven self-standard residual model can be, but is not limited to, the following various models: multiple linear/non-linear regression model, random forest, XGBoost, LightGBM, AutoEncoder, SVM, anoGAN. When the amount of collected data reaches the first set value, and the number of recorded tag information reaches the second set value, the collected data and the tag information can be used to train to obtain a supervised machine learning model, and The supervised machine learning model serves as a stable predictive maintenance model. The supervised machine learning model can be, but is not limited to, the following various models: neural network models (such as ANN, CNN, RNN, LSTM), random forests, and logistic regression models.

Step 106: After the model update trigger condition is met, update the stable predictive maintenance model, and use the updated predictive maintenance model to perform system maintenance.

Specifically, different types of the stable predictive maintenance model can be used to adopt different update triggering conditions, such as:

Of course, the data-driven self-standard residual model can also be updated manually after the system mechanism parameters change. For example, when equipment parts are replaced, lubricants are added, control parameter settings are changed, etc., the data-driven self-calibration residual model is retrained by manual triggering.

In the same way, for a supervised machine learning model, it is also possible to manually trigger the retraining of the model when abnormal data increases.

It should be noted that the update process of the above model is actually the retraining process of the model. The original model architecture is used, and the model parameters are renewed with brand new data and label information without adjusting the input data type and model structure. Training, or use the original data and label information, as well as the new data and label information to retrain the model parameters. In addition, the adjustment of the threshold may be involved in the retraining process, so that the model can better adapt to more diverse working conditions. For example, when new label information corresponding to a small amount of abnormal data is added, the threshold of the model can be adjusted. Specifically, normal operation data and abnormal data can be tested as offline test data at the same time, and corresponding thresholds can be adjusted according to the test results, thereby reducing false positives and false negatives of the model, and improving the accuracy of model prediction results.

The method for improving the adaptability of the predictive maintenance model provided by the embodiment of the present invention uses the predictive maintenance model to maintain the system, and the predictive maintenance model is adaptively adjusted according to the different stages of system operation, specifically, when the system is started In the initial stage, due to the lack of collected data and recorded tag information, an initial predictive maintenance model was established by means other than data-driven. With the operation of the system, after the amount of collected data reaches a certain amount, use the collected data Data training obtains a stable predictive maintenance model, and then replaces some specific types of initial predictive maintenance models to improve the accuracy of predictions and make the system better maintained. After the model update trigger condition is met, the stable predictive maintenance model is updated, so that the predictive maintenance model has better adaptability and meets various requirements of the system equipment operating conditions.

Further, in another embodiment of the method for improving the adaptability of a predictive maintenance model of the present invention, the types of collected data can also be recorded. In addition, after meeting the model upgrade conditions, the current predictive maintenance model used for system maintenance is upgraded, and the upgraded predictive maintenance model is used for system maintenance, so as to better improve the accuracy and stability of the predictive maintenance model Sex.

There are several situations in which the predictive maintenance model used in the current system maintenance can be upgraded:

If the predictive maintenance model currently used for system maintenance is the data-driven self-aligned residual model, when there are new types of newly collected data, the input parameters of the data-driven self-aligned residual model Adding the new type of data to training, or upgrading the data-driven self-calibrated residual model to a supervised model after the recorded label information has grown from scratch.

In the above model upgrade process, the adjustment methods of input parameters include but are not limited to: manual selection and addition, selection by correlation coefficient (such as Pearson, Spearman, etc.), selection by ANOVA analysis of variance, and feature importance selection based on tree model, etc.

In addition, when there are multiple models for predictive maintenance of the system, these models can be upgraded to integrated models, and the results of multiple models can be used for comprehensive early warning. The integrated model is for multiple separate models to judge the abnormalities separately, and then comprehensively judge based on the results of these separate models to get the final warning model. It is difficult for the integrated model to make all available models have the conditions for stable operation when it is started. Therefore, it is generally suitable to use the integrated model after the data and label information have been accumulated.

The method for improving the adaptability of the predictive maintenance model provided by the embodiment of the present invention adopts a suitable strategy design for each stage of the predictive maintenance model life cycle, so that the predictive maintenance model at each stage can be better predicted Maintenance effect.

Figure 2 shows the entire life cycle of a predictive maintenance model for system maintenance. The growth of the entire life cycle of the predictive maintenance model is divided into the following stages. The predictive maintenance models that can be used in each stage are shown in the figure Shown in 2.

The following briefly describes the stages in the entire life cycle of the predictive maintenance model.

1) Model startup stage: The characteristic of this stage is that the amount of collected data is small, the data does not have corresponding label information or the label information is less, and it is difficult to meet the offline training requirements of some commonly used data-driven models. At this stage, in consideration of the feasibility of the balance model and the accuracy of the results, there are two types of models to choose from, namely: models that require offline training and models that do not require offline training. Among them, the models that need offline training mainly include: mechanism-based residual models and models obtained through transfer learning, that is, directly use trained models with generalization characteristics of the same model and similar operating conditions; models that do not require offline training mainly There are: cluster benchmarking model and rule-based model. The selection of different types of models has been described in detail above, so I will not repeat them here.

Models that require offline training: commonly used predictive maintenance cold-start models that require offline training include mechanism-based residual models, which are models constructed based on the mechanical physical structure of components, usually regression models, because their coefficients reflect specific mechanical physics Nature, only a small amount of normal operating data can be trained; transfer learning, that is, directly use the trained model of the same model and similar working conditions with generalization characteristics.

Models that do not require offline training: directly use some models that do not require offline training, including but not limited to cluster benchmarking models and rule-based models. Cluster benchmarking is a method of comparing the current status of similar equipment under the same working condition, and judging the failure probability according to whether the equipment parameters deviate from the cluster status. The rule-based model is a method of comprehensively judging the fault-related parameters and their trends.

2) Model stable operation stage: The characteristic of this stage is that the amount of data collected has met the training requirements of the model, and the data-driven model can be stably run online. At this stage, depending on whether there is sufficient and effective label information, there are mainly two applicable optional model types, as follows:

Unsupervised learning model: It is mostly used when the data is sufficient but there is still a lack of effective label information. Mainly include: mechanism model, data-based self-benchmarking residual model, and cluster benchmarking model. Among them, the data-based self-calibration residual model is a method that uses its own historical health status to establish a model to predict the current equipment operating parameters in a healthy state, and compare with the actual measured values to determine the probability of failure.

Supervised learning model: used when the data and label information are sufficient, mainly include: classification model and neural network model that provide probability distribution.

3) Model update phase: During the stable operation phase of the model, as time goes by, as the label information increases, the accuracy of some models will gradually decrease. The main reason is that the model is affected by seasonality, or the original training data does not fully include all possible working conditions. At this time, retraining of the model is required to increase the accuracy of the model. According to different types of operating models, models affected by seasonality can be triggered regularly, models that need to cover various working conditions as much as possible can trigger training based on certain conditions, and models that need to be adjusted by false alarms can be manually triggered from time to time.

4) Model upgrade stage: After having abundant data and label information, due to the limitation of the model structure itself, retraining cannot further improve the accuracy and stability of the model. At this time, the model can be upgraded. The model upgrade mainly considers the following two changes:

Data volume-driven model changes: When the access parameters of the model increase and the amount of data increases, consider upgrading from a simpler fitting model such as a regression model to a model that is not easy to overfit and can better learn between variables and failures. Models of non-linear relationships, such as integrated tree models. . When the fault label starts from scratch, consider upgrading from an unsupervised model to a supervised model that can self-learn multiple failure development modes.

Single model to multiple models or model fusion: When there are multiple feasible models, you can use an integrated model or a model architecture that integrates a multi-model structure to further improve the accuracy and stability of the model.

It should be noted that, in actual application, in the full life cycle management of the model, the self-growth path of the model can be artificially performed technical iterations according to the above description, or can be automatically completed by the model operating system, which is not limited in the embodiment of the present invention.

The solutions of the embodiments of the present invention can be applied to a variety of systems that require predictive maintenance, such as wind power systems, machining manufacturing systems, chemical systems, tobacco production systems, and so on.

The following takes the application to a wind power system as an example to further describe the scheme of the present invention in detail.

In the model start-up phase, the start model selection judgment is first performed on the abnormal anemometer in the system. Anemometer abnormalities mainly include anemometer stuck and anemometer loose. The anemometer stuck shows that the measured wind speed is continuously lower than the true wind speed or even continues to be 0, and the anemometer loose shows that the measured wind speed jumps. These two types of faults are not necessarily slow-changing faults and do not have relevant mechanism models, and the performance judgments of different fault individuals on the data are not necessarily similar. Therefore, in the initial stage, the model for predictive maintenance of the system chooses to use a rule-based model. The design of the rules refers to the main control logic of the wind turbine, the judgment logic of maintenance and repair, etc., to comprehensively judge the wind speed measurement value, the wind speed and wind direction measurement value changes over time and other indicators within a specific power range.

When the collected operating data is accumulated for 2-3 months, based on the logical judgment of the above model growth mechanism, the model for predictive maintenance of the system is selected as the data-driven self-standard residual model. The model is based on the assumption that wind speed, power, blade angle, wind direction, etc. meet certain corresponding relationships under normal operation of the wind turbine. Use the relevant data points to predict the current wind speed, then compare the difference between the wind speed measured by the anemometer and the predicted wind speed, and use the residual distribution to make early warning and judgment of the fault. When constructing a data-driven model, the selection of variables is through a combination of mechanism and data-driven methods. Variables whose mechanisms are known to be relevant are selected, and the collected data is screened based on correlations.

Based on the characteristics of the wind power system, two retraining modes of conditional triggering retraining and manual triggering retraining are set for the data-driven self-aligned residual model. Because the tested wind farm is in a mountainous region, environmental factors such as wind speed and direction are significantly affected by the season, and the model training data at the initial stage of stable operation cannot cover the annual seasonal working conditions, so a retraining mode triggered by timing is set. At regular intervals, the system automatically uses data from normal operating conditions in recent months as input to retrain the model. In addition, a logic to compare the prediction time period and the working conditions of the training data is set in the operating system at the same time to detect changes in the working conditions in advance. When the environmental operating conditions in recent days are obviously inconsistent with the environmental operating conditions during training, the system prompts the user to manually trigger the retraining of the model.

In the above process, it is found through the verification of early warning results that if the working conditions of the wind turbines are screened more carefully, the accuracy of the data-based wind speed prediction model can be improved. Therefore, the collected data adds the turbine status indicators such as pitching and shutting down, and these newly added data types are used as newly added input parameters to upgrade the model. At the same time, after adding new input parameters, the threshold parameters of the new model are adjusted accordingly in the model training.

Correspondingly, the embodiment of the present invention also provides a device for improving the adaptability of the predictive maintenance model, as shown in FIG. 3, which is a structural block diagram of the device.

In this embodiment, the device includes the following modules:

The data acquisition module 301, the data processing module 302, the initial model establishment module 303, the model conversion judgment module 304, the stable model establishment module 305, the model update module 306, and the system maintenance module 300. among them:

The data collection module 301 is used to collect system data in real time after the system is started;

The data processing module 302 is used to mark the data, record abnormal data and label information;

The initial model establishment module 303 is used to establish an initial predictive maintenance model;

The system maintenance module 300 is configured to use the initial predictive maintenance model to perform system maintenance;

The model conversion judgment module 304 is configured to determine whether a model conversion is required according to the type of the initial predictive maintenance model after the amount of data collected by the data collection module 301 reaches a first set value; if so, notify The stable model establishing module 305 establishes a stable predictive maintenance model;

The stable model establishing module 305 is configured to use the collected data to train to obtain a stable predictive maintenance model;

Correspondingly, the system maintenance module 300 is also used to replace the initial predictive maintenance model with the stable predictive maintenance model for system maintenance;

The model update module 306 is configured to update the stable predictive maintenance model after meeting the model update trigger condition;

Correspondingly, the system maintenance module 300 is further configured to use the updated predictive maintenance model to perform system maintenance after the model update module 306 updates the stable predictive maintenance model.

The aforementioned initial model establishment module 303 may specifically adopt different technical paths to construct an initial predictive maintenance model. For example, a specific structure of the initial model establishment module 303 may include the following units:

The transfer learning model establishment unit is used when the mechanism parameters corresponding to the system cannot be obtained, and the performance of abnormal conditions of different devices is not similar, and there are other systems that have been trained and belong to the same model as the devices in this system. In the predictive maintenance model of the equipment, the migration learning model is obtained by performing migration learning on the predictive maintenance model of the equipment of the same model, and the migration learning model is used as the initial predictive maintenance model of the equipment in the system;

The aforementioned model conversion judgment module 304 determines that no model conversion is required when the initial predictive maintenance model is a mechanism-based residual model, or a cluster benchmarking model, or a migration learning model; when the initial predictive maintenance model is When using a rule-based model, it is determined that a model conversion is required.

The above-mentioned stable model establishment module 305 may specifically select different types of models according to whether there is sufficient label information, for example, there may be the following two types of models: unsupervised learning model and supervised learning model. Correspondingly, a specific structure of the stable model establishing module 305 may include the following units:

The second model establishment unit is configured to use the collected data and the label information to train to obtain a supervised machine learning model when the number of recorded label information reaches a second set value, and to combine the supervised machine learning model with the Machine learning models are used as stable predictive maintenance models, such as classification models or neural network models.

The aforementioned model update module 306 may adopt different update trigger conditions for different types of the stable predictive maintenance model. For example, a specific structure of the model update module 306 may include the following units:

Of course, in practical applications, for a data-driven self-aligned residual model, the model update module 306 may be manually triggered to retrain the model after the system mechanism parameters change, so that the model is updated. For example, when equipment parts are replaced, lubricants are added, control parameter settings are changed, etc., the data-driven self-calibration residual model is retrained by manual triggering. Similarly, for a supervised machine learning model, when abnormal data increases, the model update module 306 can be manually triggered to retrain the model.

It should be noted that the process of updating the model by the model update module 306 is actually a retraining process of the model. The original model architecture is used, and brand new data and labels are used without adjusting the input data type and model structure. Information retrains the model parameters, or uses the original data and label information, as well as the newly added data and label information to retrain the model parameters. In addition, the adjustment of the threshold may be involved in the retraining process, so that the model can better adapt to more diverse working conditions. For example, when new label information corresponding to a small amount of abnormal data is added, the threshold of the model can be adjusted. Specifically, normal operation data and abnormal data can be tested as offline test data at the same time, and corresponding thresholds can be adjusted according to the test results, thereby reducing false positives and false negatives of the model, and improving the accuracy of model prediction results.

The device for improving the adaptability of the predictive maintenance model provided by the embodiment of the present invention uses the predictive maintenance model to maintain the system, and adjusts the predictive maintenance model adaptively according to the different stages of system operation, specifically, when the system is started In the initial stage, due to the lack of collected data and recorded tag information, an initial predictive maintenance model was established by means other than data-driven. With the operation of the system, after the amount of collected data reaches a certain amount, use the collected data Data training obtains a stable predictive maintenance model, and then replaces some specific types of initial predictive maintenance models to improve the accuracy of predictions and make the system better maintained. After the model update trigger condition is met, the stable predictive maintenance model is updated, so that the predictive maintenance model has better adaptability and meets various requirements of the system equipment operating conditions.

As shown in FIG. 4, it is another structural block diagram of the device for improving the adaptability of the predictive maintenance model in the embodiment of the present invention.

Compared with the embodiment shown in FIG. 3, in this embodiment, the device further includes:

The model upgrade module 307 is used to upgrade the predictive maintenance model currently used for system maintenance after the model upgrade conditions are met.

Correspondingly, in this embodiment, the system maintenance module 300 is further configured to use the upgraded predictive maintenance model to perform system maintenance after the model upgrade module 307 upgrades the predictive maintenance model currently used for system maintenance. maintain.

1) If the predictive maintenance model currently used for system maintenance is the data-driven self-aligned residual model, when there are new types of newly collected data, the data-driven self-aligned residual model The new type of data is added to the input parameters for training, or the data-driven self-standardized residual model is upgraded to a supervised model after the recorded label information has grown from scratch.

2) If the predictive maintenance model currently used for system maintenance is the supervised machine learning model, when there are new types of newly collected data, add the supervised machine learning model to the input parameters Training on new types of data.

To this end, in this embodiment, the data processing module 302 may also record the types of collected data.

A specific structure of the model upgrade module 307 may include the following units:

The first upgrade unit is used for the data-driven self-aligned residual model when the predictive maintenance model currently used for system maintenance is the data-driven self-aligned residual model. If the newly collected data has a new type, then the data-driven self-aligned residual model Add the new type of data to the input parameters of the standard residual model for training; if the recorded label information grows from scratch, upgrade the data-driven self-standard residual model to a supervised model;

In addition, when there are multiple models for predictive maintenance of the system, the model upgrade module 307 can also upgrade these models to integrated models, and use the results of multiple models to comprehensively perform early warning. The integrated model is for multiple separate models to judge the abnormalities separately, and then comprehensively judge based on the results of these separate models to get the final warning model.

The device for improving the adaptability of the predictive maintenance model provided by the embodiment of the present invention adopts an adaptive strategy design for each stage of the predictive maintenance model life cycle, so that the predictive maintenance model at each stage can be better predicted Maintenance effect.

It should be noted that for the foregoing embodiments of the device of the present invention, since the functional implementation of each module and unit is similar to that of the corresponding method, the description of each embodiment of the dialog generating device is relatively simple, and the relevant points may be See the description of the corresponding part of the method embodiment.

It should be noted that the terms “first” and “second” in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and not necessarily used to describe a specific sequence or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so that the embodiments of the present invention described herein can be implemented in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations of them are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those clearly listed. Those steps or units may include other steps or units that are not clearly listed or are inherent to these processes, methods, products, or equipment.

The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. Moreover, the system embodiments described above are only illustrative, and the modules and units described as separate components may or may not be physically separated, that is, they may be located on one network unit, or they may be distributed to multiple On the network unit. Some or all of the modules can be selected according to actual needs to achieve the objectives of the solutions of the embodiments. Those of ordinary skill in the art can understand and implement it without creative work.

A person of ordinary skill in the art can understand that all or part of the steps in the above method embodiments can be implemented by a program instructing relevant hardware. The program can be stored in a computer readable storage medium, which is referred to herein as storage. Media, such as: ROM/RAM, floppy disk, optical disk, etc.

Correspondingly, an embodiment of the present invention also provides a device for improving the adaptability of a predictive maintenance model. The device is an electronic device, such as a mobile terminal, a computer, a tablet device, a medical device, or a fitness device. , Personal Digital Assistant, etc. The electronic device may include one or more processors and memories; wherein, the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, so as to implement the foregoing method.

The embodiments of the present invention are described in detail above, and specific implementations are used to illustrate the present invention. The descriptions of the above embodiments are only used to help understand the methods and devices of the present invention, which are only part of the embodiments of the present invention. Not all examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work should fall within the protection scope of the present invention, and the content of this specification should not be construed as limiting the present invention. Therefore, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

A method for improving the adaptability of a predictive maintenance model, characterized in that the method includes:

Collect system data in real time after the system is started, mark the data, and record abnormal data and label information;

Establish an initial predictive maintenance model, and use the initial predictive maintenance model to perform system maintenance;

After the amount of collected data reaches the first set value, determine whether a model conversion is required according to the type of the initial predictive maintenance model;

If so, use the collected data to train to obtain a stable predictive maintenance model, and replace the initial predictive maintenance model with the stable predictive maintenance model for system maintenance;

After the model update trigger condition is satisfied, the stable predictive maintenance model is updated, and the updated predictive maintenance model is used for system maintenance.
The method according to claim 1, wherein said establishing an initial predictive maintenance model comprises:

If the mechanism parameters corresponding to the system can be obtained, a mechanism-based residual model is established, and the mechanism-based residual model is used as an initial predictive maintenance model;

Otherwise, judge whether the abnormal performance of different devices in the system is similar;

If yes, establish a cluster benchmarking model, and use the cluster benchmarking model as an initial predictive maintenance model;

Otherwise, judge whether there is a trained predictive maintenance model of equipment of the same model as the equipment in this system in other systems;

If yes, perform migration learning on the predictive maintenance model of the equipment of the same model to obtain the migration learning model, and use the migration learning model as the initial predictive maintenance model of the equipment in the system;

Otherwise, establish a rule-based model and use the rule-based model as the initial predictive maintenance model.
The method according to claim 2, wherein the determining whether a model conversion is required according to the type of the initial predictive maintenance model comprises:

If the initial predictive maintenance model is a mechanism-based residual model, or a cluster benchmarking model, or a migration learning model, it is determined that no model conversion is required;

If the initial predictive maintenance model is a rule-based model, it is determined that a model conversion is required.
The method according to claim 2, wherein the training to obtain a stable predictive maintenance model using the collected data comprises:

If the number of recorded tag information does not reach the second set value, use the collected data to train to obtain a data-driven self-aligned residual model, and use the data-driven self-aligned residual model as a stable predictive Maintenance model

If the number of recorded label information reaches the second set value, use the collected data and the label information to train to obtain a supervised machine learning model, and use the supervised machine learning model as a stable predictive Maintain the model.
The method according to claim 4, wherein the updating the stable predictive maintenance model after the model update trigger condition is satisfied comprises:

If the stable predictive maintenance model is a data-driven self-aligned residual model, after reaching the update cycle, or the equipment operating condition changes, or the model accuracy rate drops to a set level, the stable predictability Maintain the model for updates;

If the stable predictive maintenance model is a supervised machine learning model, then after the number of newly added tag information reaches the third set value, or after the newly added abnormal data reaches the set threshold, use the newly collected Data to update the stable predictive maintenance model;

If the stable predictive maintenance model is a migration learning model, after the amount of newly added data reaches the fourth set value, the stable predictive maintenance model is updated with the newly collected data.
The method according to claim 4 or 5, wherein the method further comprises:

After the model upgrade conditions are met, the current predictive maintenance model used for system maintenance is upgraded, and the upgraded predictive maintenance model is used for system maintenance.
The method according to claim 6, characterized in that, the method further comprises: recording the type of the collected data;

After the model upgrade conditions are met, upgrading the predictive maintenance model currently used for system maintenance includes:

If the predictive maintenance model currently used for system maintenance is the data-driven self-aligned residual model, when there are new types of newly collected data, the input parameters of the data-driven self-aligned residual model Adding the new type of data to training, or upgrading the data-driven self-standard residual model to a supervised model after the recorded label information has grown from scratch;

If the predictive maintenance model currently used for system maintenance is the supervised machine learning model, when the newly collected data has a new type, the new type is added to the input parameters of the supervised machine learning model Data for training.
A device for improving the adaptability of a predictive maintenance model, characterized in that the device includes: a data collection module, a data processing module, an initial model establishment module, a model conversion judgment module, a stable model establishment module, a model update module, and system maintenance Module

The data collection module is used to collect system data in real time after the system is started;

The data processing module is used to mark the data and record abnormal data and label information;

The initial model establishment module is used to establish an initial predictive maintenance model;

The system maintenance module is configured to use the initial predictive maintenance model to perform system maintenance;

The model conversion judgment module is configured to determine whether a model conversion needs to be performed according to the type of the initial predictive maintenance model after the amount of data collected by the data collection module reaches a first set value; if so, notify the office The stable model establishment module establishes a stable predictive maintenance model;

The stable model establishment module is configured to use the collected data to train to obtain a stable predictive maintenance model;

Correspondingly, the system maintenance module is also used to replace the initial predictive maintenance model with the stable predictive maintenance model for system maintenance;

The model update module is used to update the stable predictive maintenance model after a model update trigger condition is met;

Correspondingly, the system maintenance module is also used to perform system maintenance using the updated predictive maintenance model after the model update module updates the stable predictive maintenance model.
The device according to claim 8, wherein the initial model establishment module comprises:

The mechanism-based residual model establishment unit is used to establish a mechanism-based residual model when the mechanism parameters corresponding to the system can be obtained, and use the mechanism-based residual model as an initial predictive maintenance model;

The cluster benchmarking model establishment unit is used to establish a cluster benchmarking model when the mechanism parameters corresponding to the system cannot be obtained and the abnormal conditions of different devices in the system are similar, and the cluster benchmarking model is used as the initial prediction Sexual maintenance model;

The transfer learning model establishment unit is used when the mechanism parameters corresponding to the system cannot be obtained, and the performance of abnormal conditions of different devices is not similar, and there are trained ones in other systems that belong to the same model as the devices in this system. In the predictive maintenance model of the equipment, the migration learning model is obtained by performing migration learning on the predictive maintenance model of the equipment of the same model, and the migration learning model is used as the initial predictive maintenance model of the equipment in the system;

The rule-based model establishment unit is used when the mechanism parameters corresponding to the system cannot be obtained, and the abnormal performance of different equipment is not similar, and there is no trained equipment in other systems that belongs to the same machine as the equipment in this system. In the case of a predictive maintenance model of a type of equipment, a rule-based model is established, and the rule-based model is used as the initial predictive maintenance model.
The device according to claim 9, wherein:

The model conversion judgment module is specifically configured to determine that no model conversion is required when the initial predictive maintenance model is a mechanism-based residual model, or a cluster benchmarking model, or a migration learning model; When the maintenance model is a rule-based model, it is determined that a model conversion is required.
The device according to claim 9, wherein the stable model establishment module comprises:

The first model establishment unit is used to retrain the ruled model using the collected data when the number of recorded tag information does not reach the second set value, to obtain a data-driven self-standardized residual model, and Using the data-driven self-aligned residual model as a stable predictive maintenance model;

The second model establishment unit is configured to use the collected data and the label information to train to obtain a supervised machine learning model when the number of recorded label information reaches a second set value, and to combine the supervised machine learning model with the The machine learning model serves as a stable predictive maintenance model.
The device according to claim 11, wherein the model update module comprises:

The first update unit is used for when the stable predictive maintenance model is a data-driven self-standard residual model, after the update period is reached, or the equipment operating condition changes, or the accuracy of the model drops to a set level, Update the stable predictive maintenance model;

The second update unit is used for when the stable predictive maintenance model is a supervised machine learning model, when the number of newly added tag information reaches the third set value, or when the newly added abnormal data reaches the set value After the threshold, use the newly collected data to update the stable predictive maintenance model;

The third update unit is used to update the stable predictive maintenance model with the newly collected data when the stable predictive maintenance model is a migration learning model, after the amount of newly added data reaches the fourth set value Update.
The device according to claim 11 or 12, wherein the device further comprises:

The model upgrade module is used to upgrade the predictive maintenance model currently used for system maintenance after meeting the model upgrade conditions;

Correspondingly, the system maintenance module is also used to perform system maintenance using the upgraded predictive maintenance model after the model upgrade module upgrades the predictive maintenance model currently used for system maintenance.
The device of claim 13, wherein:

The data processing module is also used to record the types of collected data;

The model upgrade module includes:

The first upgrade unit is used for the data-driven self-aligned residual model when the predictive maintenance model currently used for system maintenance is the data-driven self-aligned residual model, if the newly collected data has a new type, the data-driven self-aligned Add the new type of data to the input parameters of the standard residual model for training; if the recorded label information grows from scratch, upgrade the data-driven self-standard residual model to a supervised model;

The second upgrade unit is used for the current predictive maintenance model used for system maintenance is the supervised machine learning model, and when the newly collected data has a new type, the input parameters of the supervised machine learning model Add the new type of data for training.