Fault prediction and health management systems and related methods for system management

Abstract

A fault prediction and health management system and related methods for system management. The fault prediction and health management system based on a machine learning model includes a machine sensor, a command receiver, a processor, and an alarm. The machine sensor dynamically receives data from the machine under test related to the operating status of the machine under test. The command receiver dynamically receives model-specified commands. The processor dynamically uses the damage warning machine learning model corresponding to the model-specified commands to process the data from the machine under test to predict the probability of an abnormal condition occurring in the machine under test. The processor also dynamically generates a damage possibility warning on the machine under test based on the abnormal condition probability and determines whether the machine under test should continue to operate. The alarm prompts the abnormal condition probability based on the damage possibility warning and prompts the machine under test whether to continue to operate.

Description

Technical Field

[0001] The present invention relates to a prognostic and health management (PHM) system and related methods for system management, and more particularly to a PHM system and related methods that dynamically employs multiple damage warning machine learning models to respond to different conditions currently faced by the system. Background Art

[0002] The core of fault prediction and health management mechanisms is to utilize various combinations of advanced sensors, along with various algorithms and artificial intelligence models, to predict, monitor, and manage the health status of various systems, ensuring the smooth operation of their autonomous diagnostic and self-repair mechanisms. Furthermore, fault prediction and health management mechanisms constantly monitor environmental parameters detected by various systems in real time to monitor the health status of the system or equipment, or even the scope and frequency of frequent failures. Furthermore, through data monitoring and analysis, they predict possible future failures, significantly improving the operational efficiency and stability of the system or equipment. Due to these characteristics, fault prediction and health management mechanisms generally include autonomous fault detection and isolation, fault diagnosis, fault prediction, health management, and lifecycle tracking and management of individual components within the system or equipment.

[0003] However, when companies attempt to introduce fault prediction and health management mechanisms to manage their systems and equipment, it can be quite time-consuming and labor-intensive to implement them from scratch. Specifically, the implementation process involves several stages: installing sensors to collect data to build a database, data pre-processing and image conversion, defining and classifying normal and abnormal image features, processing convolutional neural networks (CNNs) with convolution and pooling architectures, implementing deep neural networks (DNNs) with fully connected layer architectures and deep learning, testing and diagnosing new data, completing model building, and conducting on-line testing of deployed equipment.

[0004] During the establishment of these implementation mechanisms, engineers need to be trained to handle both programming and maintenance. Given typical recruitment and training requirements, the estimated time required for personnel training and system setup can include the following, which are not easily pipelined: three months of Python language training plus more than three years of practical experience in the field, eight hours of sensor setup, one week of sensor data acquisition, one month of data pre-processing and conversion, one month of feature extraction, one month of model building, two months of model training, one month of model prediction, one month of data interpretation, one month of convolutional neural network architecture planning, three months of VC++ language training, three months of programmable logic controller setup and operation, and three months of AI-related training. Among the aforementioned time-consuming and outcome-control measures, the areas requiring engineer-specific training are the most unstable and time-consuming. This is due to the rapid evolution and turnover of AI technology, which rapidly increases the cost of acquiring proficiency.

[0005] See also Figure 1 This is a schematic flow chart of a conventional fault prediction and health management mechanism. The principle is to diagnose and correct the system by confirming the normal operation relationship between current data and the current system / device (hereinafter referred to as "system", but the term "device" can be substituted). This assumes that the system being monitored by the fault prediction and health management mechanism continuously receives new input data and is continuously verified to be operating normally.

[0006] First, in step 102, a starting point determination is performed to determine whether the currently received input data will trigger any system activation state. If the currently received input data will not trigger any system activation state, step 104 is executed to delete the currently received input data, abandoning monitoring of that data, and then step 102 is executed again to determine whether the next received input data will trigger any system activation state. If, in step 102, it is determined that the currently received input data will trigger any system activation state, step 106 is executed.

[0007] In step 106, the current input data is subjected to a detailed action classification, and in step 108, it is determined whether the determined action classification corresponds to a complete system action. If it is determined in step 108 that it does not correspond to a complete system action, the current input data is temporarily stored in step 110, and then the next input data related to triggering the system is waited for in step 106, completing the first cycle. If it is determined in step 108 that the next data also does not correspond to a complete system action, the next data is merged with the temporarily stored input data from the first cycle in step 110, and the next data is further waited for in step 106, completing the second cycle. Finally, the data continues to merge with new data until it is determined in step 108 that it corresponds to a complete system action, forming the final merged action data, and then the process proceeds to step 112.

[0008] In step 112, the fault prediction and health management mechanism uses a convolutional neural network to determine whether the final combined actuation data will cause a complete abnormal system actuation. If it is determined that no abnormal actuation has occurred, diagnostic data representing normal operation is output. If it is determined that an abnormal system actuation has occurred, diagnostic data representing abnormal operation is output, prompting the system to perform further diagnostic or repair operations. Regardless of the diagnostic result in step 112, the fault prediction and health management mechanism temporarily stores the accumulated non-actuation data and combines it with any new input data and its derived diagnostic data to form a historical analysis and diagnostic record.

[0009] However, the fault prediction and health management mechanism of the above-mentioned prior art also has its shortcomings. Specifically, in practice, the operation of the fault prediction and health management mechanism will quickly accumulate a large amount of input data and the massive diagnostic data derived therefrom. Therefore, it is usually impossible to fully merge the input data and the diagnostic data in real time. Instead, a large amount of data is merged once a day on a daily basis, and is temporarily stored in a short-term storage database before merging. However, if the amount of data accumulated per unit time is too large, the complexity of the data merging will increase rapidly to a level that cannot be processed in real time. Ultimately, the fault prediction and health management mechanism of the prior art will not be able to correctly perform the function of abnormal warning, that is, because it does not have a sufficiently timely historical record. Summary of the Invention

[0010] In order to address the shortcomings of the aforementioned prior art, the present invention discloses a fault prediction and health management system and related methods based on a machine learning model.

[0011] In one example, the fault prediction and health management method includes dynamically receiving data related to the operating status of a machine under test; dynamically receiving a model-specifying instruction; dynamically using a damage warning machine learning model corresponding to the model-specifying instruction to process the machine under test data to predict a probability of an abnormal condition occurring in the machine under test; and dynamically generating a damage possibility warning for the machine under test based on the abnormal condition probability and determining whether to continue operating the machine under test. The damage warning machine learning model includes a full life cycle machine training model, a fault-free machine training model, and a numerical conversion image machine training model. The full life cycle machine training model is trained based on the full life cycle operation records of at least one machine, the fault-free machine training model is trained based on the operation records of at least one machine that has never experienced a fault, and the numerical conversion image machine training model is trained by converting numerical values ​​stored in the operation records of at least one machine into images for analysis.

[0012] In one example, the deep neural network model is a deep neural network model including a low-rank decomposition.

[0013] In one example, the method includes using logistic regression and a logistic model to perform low-rank decomposition to classify the tested machine data using a regression curve; using a deep neural network to establish a corresponding deep network model; and using the deep network model to diagnose the current tested machine status and the corresponding abnormal condition probability.

[0014] In one example, the fault-free machine training model includes a support vector data description model.

[0015] In one example, the method includes using frequency features and time features on the tested machine data to perform frequency domain and time domain operations on the tested machine data; and using support vector data to describe the tested machine data after frequency domain and time domain operations to establish an optimization model to classify the probability of abnormal conditions of different data points in the tested machine data.

[0016] In one example, the numerical image conversion machine training model includes a convolutional neural network model.

[0017] In one example, the method includes performing image data filtering and segmentation and abnormal data generation on the tested machine data to convert the tested machine data into image data; extracting feature values ​​of the image data based on image features on the image data to optimize parameters of the image data; and using a convolutional neural network model to analyze the parameter-optimized image data to diagnose the current condition of the tested machine and analyze the overall abnormal condition probability corresponding to the tested machine data.

[0018] In one example, the method includes dynamically switching to use another damage warning machine learning model that is different from the currently used damage warning machine learning model to process the tested machine data to update the prediction of the probability of the abnormal condition when the model designation instruction dynamically changes to another damage warning machine learning model.

[0019] This disclosure discloses a machine learning model-based fault prediction and health management system comprising a machine sensor, a command receiver, a processor, and an alarm. The machine sensor is configured to dynamically receive data from a machine under test related to its operating status. The command receiver is configured to dynamically receive a model-specified command. The processor is configured to dynamically use a damage warning machine learning model corresponding to the model-specified command to process the machine under test data to predict the probability of an abnormal condition occurring in the machine under test. Based on the abnormal condition probability, the processor dynamically generates a damage potential warning for the machine under test and determines whether to continue operating the machine under test. The alarm is configured to provide a notification of the abnormal condition probability and a recommendation on whether to continue operating the machine under test based on the damage potential warning. The damage warning machine learning model comprises a full lifecycle machine training model, a fault-free machine training model, and a numerical conversion image machine training model. The full lifecycle machine training model is trained based on the full lifecycle operating records of at least one machine. The fault-free machine training model is trained based on the operating records of at least one machine that has never experienced a failure. The numerical value conversion image machine training model is trained by converting numerical values ​​stored in the operation records of at least one machine into images and analyzing them.

[0020] In one example, the processor includes a full lifecycle machine learning module, a fault-free machine learning module, and a numerical conversion image machine learning module. The full lifecycle machine learning module uses the full lifecycle machine training model. The fault-free machine learning module uses the fault-free machine training model. The numerical conversion image machine learning module uses the numerical conversion image machine training model. The processor dynamically designates one of the full lifecycle machine learning module, the fault-free machine learning module, and the numerical conversion image machine learning module to dynamically use the corresponding damage warning machine learning model to process the tested machine data.

[0021] In one example, the deep neural network model is a deep neural network model including a low-rank decomposition.

[0022] In one example, the full lifecycle machine learning module includes a logistic regression module, a logistic model module, and a deep neural network module. The logistic model module and the logistic regression module jointly perform low-rank decomposition to classify the test machine data using a regression curve. The deep neural network module builds a corresponding deep network model based on the classified test machine data and uses this deep network model to diagnose the current test machine status and the corresponding probability of an abnormal condition.

[0023] In one example, the fault-free machine training model includes a support vector data description model.

[0024] In one example, the fault-free machine learning module includes a frequency feature module, a time feature module, and a support vector data description module. The time feature module and the frequency feature module jointly utilize the frequency and time features of the tested machine data to perform frequency-domain and time-domain operations on the tested machine data. The support vector data description module uses the support vector data description on the tested machine data processed in the frequency-domain and time-domain operations to establish an optimization model to classify the probability of abnormal conditions at different data points in the tested machine data.

[0025] In one example, the numerical image conversion machine training model includes a convolutional neural network model.

[0026] In one example, the numerical conversion image machine learning module includes an image data filtering and cutting module, a virtual abnormal data generation module, and a convolutional neural network model module. The virtual abnormal data generation module and the image data filtering and cutting module jointly perform image data filtering and cutting and abnormal data generation on the tested machine data to convert the tested machine data into image data, and are used to extract feature values ​​of the image data based on image features on the image data to optimize the parameters of the image data. The convolutional neural network model module uses the convolutional neural network model to analyze the image data that has been parameter-optimized to diagnose the current condition of the tested machine and analyze the overall abnormal condition probability corresponding to the tested machine data.

[0027] In one example, when the model designation instruction dynamically changes the designation of another damage warning machine learning model that is different from the currently used damage warning machine learning model, the processor dynamically switches to use the other damage warning machine learning model to process the tested machine data to update the prediction of the probability of the abnormal condition. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The figure is a schematic flow chart of the fault prediction and health management mechanism implemented in the prior art.

[0029] Figure 2 Schematic diagram of a fault prediction and health management system based on a machine learning model according to an embodiment of the present disclosure.

[0030] Figure 3 According to one embodiment of the present disclosure, Figure 2 A high-level diagram of a processor launching a full-lifecycle machine learning module to train a model using a full-lifecycle machine learning model is shown.

[0031] Figure 4 According to one embodiment of the present disclosure, Figure 2 A schematic diagram of a processor shown initiating a fault-free machine learning module to employ a fault-free machine training model.

[0032] Figure 5 According to one embodiment of the present disclosure, Figure 2 The processor shown is a schematic diagram of starting the numerical conversion image machine learning module to adopt the numerical conversion image machine training model.

[0033] Figure 6 The figure shows a schematic flow chart of the fault prediction and health management method based on machine learning model in this disclosure.

[0034] Description of reference numerals:

[0035] 102, 104, 106, 108, 110, 112, steps 114, 602, 604, 606, 608

[0037] 200 Fault Prediction and Health Management System

[0038] 210 Machine Sensor

[0039] 220 Command Receiver

[0040] 230 processors

[0041] 240 alarm

[0042] 300 Full Lifecycle Machine Learning Module

[0043] 310 Logis Regression Module

[0044] 320 Logical Model Module

[0045] 330 Deep Neural Network Module

[0046] 340, 540 other test data

[0047] 350, 440, 550 test results

[0048] 400 Trouble-Free Machine Learning Module

[0049] 410 Frequency Characteristics Module

[0050] 420 Time Feature Module

[0051] 430 Support vector data description module

[0052] 500 Numerical Conversion Image Machine Learning Module

[0053] 510 Image Data Filtering and Cutting Module

[0054] 520 Virtual abnormal data generation module

[0055] 530 Convolutional Neural Network Model Module DETAILED DESCRIPTION

[0056] To address data processing issues associated with prior art implementations of fault prediction and health management mechanisms, the present disclosure proposes a fault prediction and health management method based on machine learning models to improve the problem of decreased fault prediction accuracy caused by inadequate data consolidation and processing during the implementation of the fault prediction and health management mechanism. Furthermore, the present disclosure selectively and dynamically utilizes three different damage warning machine learning models, each with its own trade-offs, based on varying levels of data completeness (i.e., varying degrees of information asymmetry and / or anomaly risk) or data processing methods, to process a continuous stream of input data. This model estimates the probability of an anomaly occurring in the current system or device under test (hereinafter referred to as the "tested device") and, based on the predicted probability, prepares for the potential anomaly. This allows the tested device to prioritize and categorize the large amount of analytical data generated during continuous operation, taking into account differences in data completeness. This allows for customized and simplified data processing methods to be optimized, reducing the burden of subsequent data processing, consolidation, and historical record creation. This also addresses the challenges associated with the rapid consolidation of large amounts of data in prior art.

[0057] Furthermore, the various damage warning machine learning models employed in this disclosure may include a full-lifecycle machine training model for monitoring the entire lifecycle of the equipment under test, a fault-free machine training model that assumes ideal operation without considering fault conditions, and a value-to-image machine training model that converts values ​​into images for comparison. As previously mentioned, these three machine learning models can each respond to varying degrees of information asymmetry and / or correspondingly varying degrees of abnormal condition risk.

[0058] The characteristic of the full lifecycle machine training model is that, because it is based on observing the complete lifecycles of other machines, it has the most complete training data and can capture the most error conditions. In other words, the full lifecycle machine training model is ideal for monitoring environments where damage is more likely to occur. However, conversely, for monitoring environments where damage is less likely to occur, using only the full lifecycle machine training model is less efficient and tends to generate redundant analysis data. Therefore, this disclosure selectively favors the full lifecycle machine training model in situations where there is a high risk of information asymmetry and / or anomalies in the input data.

[0059] A characteristic of the fault-free machine training model is that it is built by observing the conditions of other machines that have never experienced failures. This results in minimal data and improved processing efficiency. However, in contrast to the full lifecycle machine training model, using only the fault-free machine training model to monitor machines will fail to consider many real-world failure scenarios, resulting in a relatively less rigorous monitoring process. Therefore, this disclosure selectively favors the use of the fault-free machine training model in less complete input data, that is, in less idealized scenarios with lower information asymmetry and / or anomaly risk.

[0060] The numerical conversion image machine training model is characterized by its data processing method, which uses images after numerical conversion as units for processing, including comparison. Therefore, it is suitable for processing large amounts of regional data in a short period of time and is the fastest of the three machine learning models mentioned above. However, the numerical conversion image machine training model still lacks the monitoring rigor of the full life cycle machine training model. Therefore, this disclosure selectively prefers to use the numerical conversion image machine training model when massive input data is present and the data needs to be simplified as quickly as possible for subsequent processing.

[0061] Since the above three machine learning models each have their own advantages and disadvantages, the fault prediction and health management system and method of the present disclosure will dynamically switch between the above three machine learning models according to the current input data situation to balance the impact of the advantages and disadvantages of the above three machine learning models, and ultimately achieve better data processing rate and abnormality judgment accuracy than the existing technology.

[0062] See also Figure 2 , which is a schematic diagram of a fault prediction and health management system 200 based on a machine learning model according to an embodiment of the present disclosure. The fault prediction and health management system 200 includes a machine sensor 210, a command receiver 220, a processor 230, and an alarm 240.

[0063] The machine sensor 210 is used to sense and generate machine data formed from multiple monitored parameters of the machine under test, which serves as dynamic input data for the fault prediction and health management system 200. The instruction receiver 220 is used to dynamically receive model-specific instructions generated in response to different input data conditions (e.g., the degree of information asymmetry and / or the input data occurrence rate) to determine the damage warning machine model used by the processor 230. The processor 230 is equipped with the fault prediction and health management method described in this disclosure. Specifically, the processor 230 dynamically switches, in real time, based on the model-specific instructions received by the instruction receiver 220, to use one of the three aforementioned machine learning models to process the current machine under test data. This predicts the probability of an abnormal condition occurring in the machine under test, thereby facilitating subsequent data classification before data merging, such as by classifying the required data based on the abnormal condition probability. Furthermore, the processor 230 also dynamically generates a damage potential warning for the machine under test based on the predicted abnormal condition probability to determine whether the machine under test can continue to operate. For example, if the processor 230 predicts an abnormality probability higher than a critical abnormality probability, the aforementioned damage possibility warning will be issued. This critical abnormality probability may vary under different machine learning models and / or different tested machine data, and is determined dynamically. If the processor 230 predicts an abnormality probability lower than the critical abnormality probability, no damage possibility warning will be issued on the tested machine, or a damage possibility warning within a safe probability range will be issued. Finally, the alarm 240 will issue a warning based on the abnormality probability predicted by the processor 230 and, based on its built-in response mechanism, provide a recommendation on whether the tested machine can continue to operate.

[0064] The fault prediction and health management method implemented by processor 230 includes the aforementioned damage warning machine learning models, each described in detail below. Processor 230 may be individually equipped with a complete lifecycle machine learning module 300, a fault-free machine learning module 400, and a numerical image conversion machine learning module 500.

[0065] First, as mentioned above, the damage warning machine learning model used in this disclosure includes a full life cycle machine training model. In one embodiment of this disclosure, the model may include a low-rank factorization deep neural network (LRF DNN) model. Figure 3FIG2 is a schematic diagram illustrating the process of the processor 230 activating the full lifecycle machine learning module 300 to employ a full lifecycle machine training model according to an embodiment of the present disclosure. The full lifecycle machine learning module 300 includes a logistic regression module 310, a logit model module 320, and a deep neural network module 330.

[0066] When the model-specifying instruction received by the instruction receiver 220 specifies a full lifecycle machine training model, the processor 230 selects the full lifecycle machine learning module 300 to process the machine data under test received by the machine sensor 210. The logistic regression module 310 and the logical model module 320 cooperate to perform low-rank decomposition to pre-classify the machine data under test using a regression curve. Subsequently, the deep neural network module 330 establishes a corresponding deep network model based on other test data 340 (such as machine learning training data) or previously accumulated non-actuated data) and the classified data. The deep network model is used to diagnose the current condition of the machine under test and the corresponding probability of abnormal conditions, ultimately generating a test result 350 as a basis for generating a potential damage warning.

[0067] As mentioned above, the advantage of selecting the full life cycle machine learning module 300 is that, because the data for training the deep neural network is based on the full life cycle, the prediction results are usually closest to the actual damage that will occur to the tested machine and have a certain degree of prediction accuracy.

[0068] Furthermore, the damage warning machine learning model used in this disclosure includes a fault-free machine training model. In one embodiment of this disclosure, the model may include a Support Vector Data Description (SVDD) model. Figure 4 , which is a schematic diagram illustrating the process of the processor 230 activating the fault-free machine learning module 400 to employ a fault-free machine training model according to one embodiment of the present disclosure. The fault-free machine learning module 400 includes a frequency feature module 410, a temporal feature module 420, and a support vector data description module 430.

[0069] When the model-specifying instruction received by the instruction receiver 220 specifies a fault-free machine training model, the processor 230 selects the fault-free machine learning module 400 to process the test machine data received by the machine sensor 210. The frequency feature module 410 and the time feature module 420 perform certain frequency-domain and time-domain operations on the test machine data, followed by final processing by the support vector data description module 430. In one embodiment, the support vector data description module 430 projects the training data into a high-dimensional space to construct an optimized model shaped like a hypersphere that encompasses most of the training data and has a minimum volume. Based on this model, the test machine data is then divided into regions representing normal, warning, and highly abnormal conditions, corresponding to different probabilities of abnormal conditions. Furthermore, during the process of building this optimized model, the support vector data description module 430 learns a decision boundary using normal data (i.e., data with a low probability of abnormality). Using this learned decision boundary, the module determines whether new input data points exceed the current hypersphere boundary, thereby classifying data points that exceed the hypersphere boundary as having a higher probability of abnormality. Finally, the support vector data description module 430 outputs test results 440 based on the data classified by abnormal condition probability, which serves as the basis for generating damage possibility warnings. The advantage of this is that the output data is relatively simple and easy to process, thereby speeding up the data merging rate.

[0070] Finally, the damage warning machine learning model used in this disclosure includes a numerical conversion image machine training model. In one embodiment of this disclosure, the model may include a convolutional neural network (CNN) model. Figure 5 , which is a schematic diagram illustrating the process of processor 230 activating the numerical image conversion machine learning module 500 to employ a numerical image conversion machine training model according to one embodiment of the present disclosure. The numerical image conversion machine learning module 500 includes an image data filtering and segmentation module 510, a virtual abnormal data generation module 520, and a convolutional neural network model module 530.

[0071] When the model-specifying instruction received by the instruction receiver 220 specifies a numerical image conversion machine training model, the processor 230 selects the numerical image conversion machine learning module 500 to process the machine data under test received by the machine sensor 210. The image data filtering and segmentation module 510, in conjunction with the virtual abnormality data generation module 520, first converts the machine data under test into image data and extracts feature values ​​based on the image features present in the image data, which serve as the basis for determining whether the machine data under test is abnormal (e.g., determining the probability of abnormality at each data point). Because the machine data under test has been parameter-optimized during this process, it can better mitigate past errors in image data and make more accurate judgments. The convolutional neural network model module 530 then analyzes the optimized image data with other test data 540 (e.g., machine learning training data) to diagnose the current machine condition under test and the corresponding overall data abnormality probability, ultimately generating a test result 550 that serves as the basis for subsequent damage potential warnings.

[0072] As mentioned above, after the data of the tested machine is converted into image data, the feature values ​​in the image can be extracted to determine whether an abnormality has occurred. Therefore, the amount of data to be processed is reduced, and the data that has previously been image-recognized can be used as an aid to avoid feature misidentification.

[0073] Figure 6 The figure shows a schematic flow chart of the machine learning model-based fault prediction and health management method of this disclosure. The main steps in the flow chart are the various execution steps of the fault prediction and health management system 200 mentioned above, so they are not repeated here. The flow chart includes the following steps.

[0074] Step 602: Dynamically receive data of a tested machine related to an operating status of the tested machine;

[0075] Step 604: Dynamically receive a model specification instruction;

[0076] Step 606: Dynamically use a damage warning machine learning model corresponding to the model-specified instruction to process the tested machine data to predict an abnormality probability of the tested machine experiencing an abnormality; and

[0077] Step 608: Dynamically generate a damage possibility warning for the tested machine based on the abnormal condition probability, and decide whether to continue to operate the tested machine.

[0078] In summary, the machine learning-based fault prediction and health management system and related methods disclosed herein primarily address existing issues such as the massive and rapidly generated data volume, which hinders processing and merging, and consequently affects the accuracy of early warnings. This approach utilizes a method that dynamically switches between damage warning machine learning models tailored to the probability of different abnormal conditions.

Claims

1. A fault prediction and health management method based on a machine learning model, comprising: Dynamically receiving a piece of tested machine data related to an operating condition of the tested machine; Dynamically receiving a model-specifying instruction; Dynamically using a damage warning machine learning model corresponding to the model-specified instruction to process the data of the tested machine to predict an abnormality probability of the tested machine experiencing an abnormality; and Dynamically generating a damage possibility warning for the tested machine based on the probability of the abnormal condition and determining whether the tested machine should continue to operate; The damage warning machine learning model includes a complete life cycle machine training model, a fault-free machine training model, and a numerical conversion image machine training model. The complete life cycle machine training model is trained based on the complete life cycle operation records of at least one machine. The fault-free machine training model is trained based on the operation records of at least one machine that has never experienced a failure. The numerical conversion image machine training model is trained by converting the numerical values ​​stored in the operation records of at least one machine into images for analysis. in, Based on the current input of the tested machine data, the full life cycle machine training model, the fault-free machine training model, and the numerical conversion image machine training model are dynamically replaced to balance the effects of the advantages and disadvantages of the full life cycle machine training model, the fault-free machine training model, and the numerical conversion image machine training model. Among them, the complete life cycle machine training model is selected when there is high information asymmetry and / or abnormality risk in the input data, the fault-free machine training model is used under ideal conditions where there is low information asymmetry and / or abnormality risk in the input data, and the numerical conversion image machine training model is used when massive input data appears and the data needs to be simplified as quickly as possible for subsequent processing.

2. The method of claim 1, wherein the full lifecycle machine training model comprises a low-rank decomposition deep neural network model.

3. The method of claim 2 , wherein dynamically using the damage warning machine learning model corresponding to the model-specified instruction to process the tested machine data to predict the probability of the tested machine experiencing an abnormal condition comprises: Use logistic regression and logistic model to perform low-rank decomposition to classify the tested machine data using regression curves; Using a deep neural network to establish a corresponding deep network model; and Use this deep network model to diagnose the current condition of the tested machine and the corresponding probability of abnormal conditions.

4. The method of claim 1, wherein the fault-free machine training model comprises a support vector data description model.

5. The method of claim 1 , wherein dynamically using the damage warning machine learning model corresponding to the model-specified instruction to process the tested machine data to predict the probability of the tested machine experiencing an abnormal condition comprises: Using frequency characteristics and time characteristics on the data of the tested machine to perform frequency domain and time domain operations on the data of the tested machine; and Support vector data is used to describe the tested machine data processed by frequency domain and time domain operations to establish an optimization model to classify the abnormality probability of different data points in the tested machine data.

6. The method of claim 1, wherein the numerical conversion image machine training model comprises a convolutional neural network model.

7. The method of claim 1 , wherein dynamically using the damage warning machine learning model corresponding to the model-specified instruction to process the tested machine data to predict the probability of the tested machine experiencing an abnormal condition comprises: Performing image data filtering and cutting and abnormal data generation on the tested machine data to convert the tested machine data into image data; extracting feature values ​​of the image data according to image features of the image data to optimize parameters of the image data; and A convolutional neural network model is used to analyze the image data after parameter optimization to diagnose the current condition of the tested machine and analyze the overall abnormal condition probability corresponding to the tested machine data.

8. The method of claim 1, wherein dynamically using the damage warning machine learning model corresponding to the model-specifying instruction to process the data of the machine under test to predict the probability of an abnormal condition occurring in the machine under test comprises: when the model-specifying instruction dynamically changes to specify another damage warning machine learning model that is different from the currently used damage warning machine learning model, dynamically switching to use the other damage warning machine learning model to process the data of the machine under test to update the prediction of the probability of an abnormal condition.

9. A fault prediction and health management system based on a machine learning model, comprising: a machine sensor for dynamically receiving a machine under test data related to an operating condition of the machine under test; a command receiver for dynamically receiving a model-specified command; a processor configured to dynamically use a damage warning machine learning model corresponding to the model-specified instruction to process the data of the machine under test to predict a probability of an abnormal condition occurring in the machine under test. The processor also dynamically generates a damage possibility warning for the machine under test based on the abnormal condition probability and determines whether to continue operating the machine under test; and An alarm device is used to indicate the probability of abnormal condition based on the damage possibility warning and to provide a suggestion on whether the tested machine should continue to operate; The damage warning machine learning model includes a complete life cycle machine training model, a fault-free machine training model, and a numerical conversion image machine training model. The complete life cycle machine training model is trained based on the complete life cycle operation records of at least one machine. The fault-free machine training model is trained based on the operation records of at least one machine that has never experienced a failure. The numerical conversion image machine training model is trained by converting the numerical values ​​stored in the operation records of at least one machine into images for analysis. in, Based on the current input of the tested machine data, the full life cycle machine training model, the fault-free machine training model, and the numerical conversion image machine training model are dynamically replaced to balance the effects of the advantages and disadvantages of the full life cycle machine training model, the fault-free machine training model, and the numerical conversion image machine training model. Among them, the complete life cycle machine training model is selected when there is high information asymmetry and / or abnormality risk in the input data, the fault-free machine training model is used under ideal conditions where there is low information asymmetry and / or abnormality risk in the input data, and the numerical conversion image machine training model is used when massive input data appears and the data needs to be simplified as quickly as possible for subsequent processing.

10. The fault prediction and health management system according to claim 9, wherein the processor comprises: a full lifecycle machine learning module that employs the full lifecycle machine training model; a fault-free machine learning module that employs the fault-free machine training model; and a numerically converted image machine learning module, which uses the numerically converted image machine training model; The processor dynamically specifies to use one of the complete life cycle machine learning module, the fault-free machine learning module, and the numerical conversion image machine learning module to dynamically adopt the corresponding damage warning machine learning model to process the data of the tested machine.

11. The fault prediction and health management system as claimed in claim 10, wherein the full life cycle machine training model comprises a low-rank decomposition deep neural network model.

12. The fault prediction and health management system of claim 11, wherein the full lifecycle machine learning module comprises: A logistic regression module; a logistic model module, configured to perform low-rank decomposition together with the logistic regression module to classify the tested machine data using a regression curve; and A deep neural network module is used to establish a corresponding deep network model based on the classified data of the tested machine, and diagnose the current condition of the tested machine and the corresponding probability of an abnormal condition based on the deep network model. 13 . The fault prediction and health management system as claimed in claim 10 , wherein the fault-free machine training model comprises a support vector data description model.

14. The fault prediction and health management system of claim 13, wherein the fault-free machine learning module comprises: a frequency characteristic module; a time feature module, used to work with the frequency feature module to apply frequency features and time features to the data of the tested machine to perform frequency domain and time domain operations on the data of the tested machine; and A support vector data description module is used to use the support vector data description on the tested machine data processed by frequency domain and time domain operations to establish an optimization model to classify the abnormality probability of different data points in the tested machine data.

15. The fault prediction and health management system as claimed in claim 10, wherein the numerical conversion image machine training model comprises a convolutional neural network model.

16. The fault prediction and health management system according to claim 15, wherein the numerical value conversion image machine learning module comprises: an image data filtering and cutting module; a virtual abnormal data generation module, which, together with the image data filtering and cutting module, performs image data filtering and cutting and abnormal data generation on the tested machine data to convert the tested machine data into image data, and is used to extract feature values ​​of the image data based on image features on the image data to optimize parameters of the image data; and A convolutional neural network model module is used to use the convolutional neural network model to analyze the image data that has been parameter-optimized to diagnose the current condition of the tested machine and analyze the overall abnormal condition probability corresponding to the tested machine data.

17. The fault prediction and health management system as described in claim 10, wherein when the model designation instruction dynamically changes the designation of another damage warning machine learning model that is different from the currently used damage warning machine learning model, the processor is used to dynamically switch to using the other damage warning machine learning model to process the tested machine data to update the prediction of the abnormal condition probability.

Images