Embedded multi-mode health management system and control method thereof

Abstract

The invention discloses an embedded multi-mode health management system and a control method. The embedded multi-mode health management system comprises a tested health management unit and a control end. The detected end comprises a data preprocessing module, a fault diagnosis module, a service life prediction module and a data storage module, and is used for receiving and processing multi-modal original data in real time so as to realize fault detection and service life prediction; the control end comprises a basic management module, a comprehensive data management module and the like, and cooperates with the tested end through a 1553B bus and the Ethernet. According to the method, multi-source data are fused, algorithm embedded deployment is optimized, a diagnosis model of local information increment and slow feature analysis and a prediction model of random process and deep learning fusion are adopted, closed-loop management is formed, the problems of insufficient precision, response lag and the like of an existing system are solved, and the precision and efficiency of equipment health management are improved.

Description

Technical Field

[0001] This invention belongs to the field of health technology for electronic devices, and in particular relates to an embedded multimodal health management system and its control method. Background Technology

[0002] With the rapid development of industries such as industrial automation, rail transportation, and new energy power generation, the operational reliability and safety of complex and critical equipment such as industrial robots, CNC machine tools, and wind turbine generators are becoming increasingly important. As a core technology ensuring stable equipment operation, health management systems achieve full lifecycle status monitoring and risk control of equipment through data collection, fault diagnosis, lifespan prediction, and maintenance decision-making, serving as a key support for the intelligent upgrading of industry.

[0003] Current core technologies for equipment health management fall into two categories, both with significant shortcomings: The first is an embedded single-modal health management system, deployed locally on the device. This system consists of data acquisition, simple data processing, single fault diagnosis, and data storage units. It collects single-type operational data through sensors, performs simple filtering, and then uses preset rules or basic algorithms to determine the status and store it locally, aiming to avoid response lag caused by data transmission delays. However, this solution lacks multimodal data processing capabilities, only analyzing single data points and failing to integrate multi-dimensional information, resulting in insufficient diagnostic and prediction accuracy. Furthermore, the core algorithms suffer from low embedded portability: they are not adapted to embedded hardware, leading to frequent lag and high resource consumption after porting. Finally, the operating mode lacks flexibility, only adapting to specific device data and making it difficult to adjust processing strategies.

[0004] Secondly, there is the ground-based centralized health management system. Its core processing and decision-making functions are deployed on a ground server, including data receiving, comprehensive analysis, maintenance decision-making, and visualization units. It receives data from distributed devices via bus or network, performs preprocessing, diagnosis, prediction, and decision-making on the ground, and then visualizes the results. The aim is to leverage the powerful computing and storage capabilities of the ground to achieve centralized management of multiple devices. However, this solution suffers from several drawbacks, including insufficient collaboration with the tested units (devices only collect and upload data, lacking local real-time preprocessing capabilities, resulting in large data transmission volumes and delayed responses), an incomplete health management loop (lack of efficient two-way interaction between ground decision-making and device execution, leading to high command delays), and insufficient data processing targeting (raw data is not accurately preprocessed, containing a large amount of outlier noise, affecting analysis accuracy and decision reliability). Currently, there is an urgent need for a health management solution that combines accurate multimodal data processing, efficient porting of core algorithms, deep collaboration between embedded and ground systems, flexible adaptation to working modes, and full-process closed-loop management. Summary of the Invention

[0005] The purpose of this invention is to provide an embedded multimodal health management system and its control method. By integrating multi-source heterogeneous data, optimizing the embedded deployment efficiency of algorithms, enhancing collaborative capabilities, and supporting dynamic switching of multiple working modes, it achieves high-precision perception, intelligent decision-making, and closed-loop control throughout the entire health management process.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is an embedded multimodal device health management microsystem, including a health management unit under test, used to predict the lifespan and diagnose faults of the target device for health management, and send the prediction or diagnosis results to the control terminal. The health management unit under test includes a data preprocessing module, a fault diagnosis module, a lifespan prediction module, and a data storage module; the health management unit under test establishes a communication connection with the control terminal.

[0007] The data preprocessing module is used to receive multimodal raw data and perform real-time preprocessing, and output the processed data to the fault diagnosis module and the life prediction module.

[0008] The fault diagnosis module is used to perform real-time detection and diagnosis of equipment faults from the preprocessed data and output the fault diagnosis results.

[0009] The life prediction module is used to perform equipment performance degradation modeling and remaining life prediction on the preprocessed data, and output the life prediction results.

[0010] The data storage module is used to store preprocessed data, fault diagnosis results, and life prediction results, and to interact bidirectionally with the control terminal.

[0011] Furthermore, the multimodal raw data includes logical data, timing data, and state data; the logical data includes device switching status and command response signals; the timing data includes current-voltage change curves and rotational speed timing sequences; and the state data includes steady-state parameters such as temperature and vibration amplitude.

[0012] Furthermore, the real-time preprocessing operations of the data preprocessing module include outlier identification and removal, data normalization and feature extraction, and data denoising; the outlier identification and removal is used to remove outliers in the data, the data normalization and feature extraction is used to extract key features and perform normalization processing, and the data denoising is used to reduce the impact of environmental interference on data quality.

[0013] Furthermore, the diagnostic scope of the fault diagnosis module includes minor fault identification and multivariate statistical process monitoring, and the fault diagnosis module deploys a fault detection model based on local information increment and slow feature analysis; the fault diagnosis results include fault type, occurrence time and severity level.

[0014] Furthermore, the lifetime prediction module supports multi-indicator fusion evaluation and deploys a lifetime prediction model based on stochastic processes and a lifetime prediction model based on deep learning; the lifetime prediction model based on stochastic processes includes a linear Wiener process model, a nonlinear Wiener process model, and an exponential stochastic degradation model.

[0015] Furthermore, it includes the tested health management unit and control terminal as described in any one of claims 1-5; the control terminal includes a basic management module, a comprehensive data management module, a maintenance decision module, a visualization module, a human-computer interaction module, and a health management center unit; the basic management module is used for equipment information management, user permission management, and model parameter configuration;

[0016] The integrated data management module is used to receive and integrate multi-source data uploaded by the tested health management unit, and provides data query, structured storage and backup functions;

[0017] The maintenance decision module generates maintenance plans and recommendations based on fault diagnosis results and life prediction results.

[0018] The visualization module displays data from the entire health management process in a graphical format;

[0019] The human-computer interaction module provides an operation interface and supports interactive functions such as command issuance and parameter setting.

[0020] The health management center unit also includes a data preprocessing module, a fault diagnosis module, a life prediction module, a data storage module, and a maintenance decision module, which are used to achieve global optimization and model training; the tested health management unit and the control terminal transmit control commands through a 1553B bus and transmit data through Ethernet.

[0021] A control method for an embedded multimodal device health management microsystem includes the following steps:

[0022] S1. The tested health management unit collects multimodal raw data through sensors or a bus;

[0023] S2, the data preprocessing module performs real-time preprocessing operations on the multimodal raw data and outputs high-quality data;

[0024] S3. The fault diagnosis module, based on the preprocessed data, realizes real-time fault detection and diagnosis through the fault detection model, outputs the fault diagnosis results and stores them in the data storage module, and uploads them to the control terminal at the same time.

[0025] S4. The lifetime prediction module, based on the preprocessed data, performs performance degradation modeling and lifetime prediction through the remaining lifetime prediction model, outputs lifetime prediction results and stores them in the data storage module, and uploads them to the control terminal at the same time.

[0026] S5. The control terminal receives and integrates the data uploaded by the tested health management unit, generates a maintenance strategy through the maintenance decision module, displays it through the visualization module, and supports user operation and command issuance through the human-computer interaction module.

[0027] S6. The tested health management unit receives control commands from the control terminal, executes the corresponding operations, and feeds back the execution results, forming a closed-loop management system.

[0028] Furthermore, in step S2, the real-time preprocessing operation includes outlier removal, data normalization and feature extraction, and data denoising.

[0029] The outlier removal method employs the standard deviation algorithm, box plot method, or Hampel function method.

[0030] The data normalization and feature extraction employ empirical mode decomposition and principal component analysis, combined with normalization processing;

[0031] The data denoising employs a moving average algorithm or singular value decomposition filtering.

[0032] Furthermore, in step S3, the specific process of fault diagnosis includes:

[0033] S31. Perform dynamic characteristic discrimination on the preprocessed monitoring data;

[0034] S32. Fault detection is performed using a fault detection model based on local information increment and slow feature analysis. The model is trained offline using training data and then used for online detection using test data.

[0035] S33. Calculate the mean of local information increments as a statistic, and construct a dynamic threshold based on the S-shaped membership function;

[0036] S34. Determine whether a sample is a faulty sample based on the comparison results of the statistics and the dynamic threshold, and output the fault diagnosis results.

[0037] Furthermore, the specific process for lifetime prediction includes:

[0038] S41. Degradation trend determination is performed on the preprocessed monitoring data;

[0039] S42. If the monitoring data shows no significant degradation trend, a deep learning-based remaining life prediction model is used, and the prediction results are output after offline training and online testing.

[0040] S43. If the monitoring data shows a significant degradation trend, construct four types of degradation models: linear, power, logarithmic, and exponential. Use a residual lifetime prediction model based on stochastic processes for prediction and combine the results of the deep learning model to obtain a fused prediction result.

[0041] S44. Select the optimal model using the AIC criterion or mean square error, and output the final lifetime prediction result.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] This invention achieves deep fusion of multi-source heterogeneous data, including logical data, time-series data, and state data, through an innovative data preprocessing mechanism. Specifically, it employs targeted operations such as outlier identification and removal, data normalization and feature extraction, and data denoising to provide high-quality input for subsequent analysis, fundamentally improving the accuracy of fault feature extraction and equipment status identification. Secondly, addressing the issue of limited embedded hardware resources, the core diagnostic and prediction algorithms have been deeply optimized and efficiently ported. The fault diagnosis module uses a fault detection model based on local information increment and slow feature analysis to adaptively construct dynamic thresholds, effectively identifying minor faults. The lifespan prediction module integrates a stochastic process-based model and a deep learning-based model, intelligently selecting the optimal prediction path using the AIC criterion, ensuring both computational accuracy and real-time performance and smoothness on the embedded microsystem. Thirdly, a deep collaborative architecture between the tested health management unit and the control terminal is constructed, with the embedded terminal responsible for real-time acquisition of local data. Preprocessing and edge computing enable rapid response, while the control end aggregates multi-source data, performs comprehensive analysis through the maintenance decision module to generate maintenance strategies, and then displays and issues commands through the visualization and human-machine interaction modules. Both transmit control commands via 1553B bus and data via Ethernet, forming a complete intelligent decision-making closed loop of "perception-diagnosis-prediction-decision-control-feedback", which reduces the data transmission burden and ensures the timeliness of decision-making and the effectiveness of execution. Fourth, the system supports dynamic switching of multiple working modes. The life prediction module first judges the degradation trend of the monitoring data. For data without significant degradation trend, deep learning models are used for feature mining and life prediction. For data with significant degradation trend, it automatically switches to stochastic process models or performs model fusion prediction, and selects the optimal model through the AIC criterion. This data-driven mode switching strategy significantly enhances the system's flexibility, adaptability, universality and practicality under different equipment and operating conditions. In summary, through the above four technological innovations, this invention effectively achieves high-precision perception, intelligent decision-making, and closed-loop control throughout the entire process of industrial equipment health management, significantly improving the operational reliability, maintenance efficiency, and economy of the equipment. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic diagram of the embedded multimodal health management system of this embodiment;

[0046] Figure 2 This is a schematic diagram of the local information increment structure in this embodiment;

[0047] Figure 3 This is a diagram of the deep belief network structure in this embodiment;

[0048] Figure 4 This is a schematic diagram of the Boltzmann machine in this embodiment;

[0049] Figure 5 This is a diagram of the long short-term memory network structure in this embodiment;

[0050] Figure 6 This is a structural diagram of the gated loop unit in this embodiment;

[0051] Figure 7 This is a diagram of the bidirectional long short-term memory network structure in this implementation. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] This embodiment provides an embedded multimodal health management system and control method; wherein, as Figure 1 The health management target system includes a test end and a control end; wherein the test end includes a test health management unit (embedded end) for predicting the lifespan and diagnosing faults of the health management target equipment, and sending the prediction or diagnosis results to the control end; the control end includes a basic management module, a comprehensive data management module, a maintenance decision module, a visualization module; and a health management center unit.

[0054] This embodiment uses a prototype as a device to realize the core functions and behaviors of real equipment. In actual application scenarios, this prototype can be replaced by key components of various complex equipment that require health management, such as automobiles, industrial robots, CNC machine tools, wind turbine generators, and rail transit traction systems, for example, the electronic control unit (ECU) of automobiles.

[0055] In this embodiment, the health management unit under test includes a data preprocessing module, a fault diagnosis module, a lifespan prediction module, and a data storage module.

[0056] The data preprocessing module is used to perform real-time preprocessing on the multimodal raw data collected from the sensors. The multimodal raw data includes logical data, time-series data, and state data. Through the above real-time preprocessing operations, the data preprocessing module can provide high-quality data input for the subsequent fault diagnosis module and life prediction module. Its preprocessing process specifically covers outlier identification and removal, data normalization and feature extraction, data denoising, and other operations.

[0057] Specifically, the data preprocessing module uses methods such as standard deviation algorithm, box plot method, and Hampel function to remove outliers; in the feature extraction stage, it uses methods such as empirical mode decomposition (EMD) and principal component analysis (PCA) to normalize the data and extract features; and in the data denoising process, it uses methods such as moving average and singular value decomposition (SVD) for processing.

[0058] Furthermore, the data preprocessing module receives raw data from sensors or a bus, and after completing preprocessing, outputs the processed data to the fault diagnosis module and the life prediction module.

[0059] A fault diagnosis module is included. This module performs real-time detection and diagnosis of equipment faults based on preprocessed data. Its diagnostic scope includes minor fault identification and multivariate statistical process monitoring. Specifically, during fault detection, the module deploys a fault detection model based on local information increment and slow feature analysis to achieve accurate identification and judgment of equipment faults. The fault diagnosis module receives preprocessed data output from the data preprocessing module. After completing the fault diagnosis, it outputs the fault diagnosis results (including fault type, occurrence time, etc.) to the data storage module and uploads the results to the integrated data management module at the control end via a communication interface.

[0060] The lifespan prediction module is used to model equipment performance degradation and predict remaining lifespan based on preprocessed data, and supports multi-indicator fusion evaluation to improve the comprehensiveness and accuracy of the prediction. Specifically, the lifespan prediction module deploys a remaining lifespan prediction model based on stochastic processes and a remaining lifespan prediction model based on deep learning. By modeling and analyzing the patterns of equipment performance degradation, it achieves accurate prediction of remaining lifespan. The lifespan prediction module receives preprocessed data output from the data preprocessing module, outputs the lifespan prediction results to the data storage module after completing the lifespan prediction, and uploads the prediction results to the integrated data management module at the control end through a communication interface.

[0061] The data storage module stores preprocessed data, fault diagnosis results, lifespan prediction results, and other data information. It also supports local caching and historical data management to achieve orderly storage and control of various key data. The data storage module receives data from the data preprocessing module, fault diagnosis module, and lifespan prediction module, transmits the stored data to the control terminal via a communication interface, and receives data query and read commands from the control terminal, thereby completing bidirectional data interaction and response.

[0062] In some specific implementations, the control terminal includes a basic management module. This module manages the system's basic configuration, including device information management, user access control, and model parameter configuration. Effective control over these basic configurations provides fundamental support for the overall system operation. The basic management module also has data interaction capabilities: it provides necessary configuration information to other modules within the system and receives various configuration commands input by the user through the human-machine interaction module to respond to system configuration adjustment needs.

[0063] The integrated data management module receives, integrates, and manages multi-source data uploaded from the tested health management unit. This multi-source data includes real-time data, historical data, diagnostic results, and prediction results. The module also provides data query, structured storage, and backup functions, enabling systematic control and efficient retrieval of the multi-source data. Furthermore, the integrated data management module receives data transmitted from the data storage module of the tested health management unit, processes it, and provides the integrated dataset to the maintenance decision-making module and the visualization module, providing data support for their operation.

[0064] The visualization module is used to graphically display data from the entire health management process, including real-time equipment status data, fault diagnosis results, lifespan prediction results, and maintenance decision results. The visualization module supports various visualization formats, including curves, charts, and alarm information, to intuitively present various data information. The visualization module receives data from the integrated data management module and the maintenance decision module, processes it to form a visualization interface, and then provides this visualization interface to the human-computer interaction module, supporting users in intuitively obtaining health management-related information.

[0065] The human-computer interaction (HCI) module provides a user interface and supports various HCI functions such as command issuance, parameter setting, data querying, and mode switching. It facilitates convenient operation and information exchange between the user and the system, ensuring efficient user management of system operation. The HCI module receives various commands input by the user through the interface and forwards them to relevant modules within the system (such as the basic management module and maintenance decision module) to drive corresponding operations. Simultaneously, the HCI module retrieves various display content from the visualization module and presents it to the user on the interface, ensuring intuitive access to system feedback.

[0066] The Health Management Center Unit, as the core processing unit of the control end, is used to realize health management functions in collaboration with the tested Health Management Unit (embedded end). It also includes a data preprocessing module, a fault diagnosis module, a lifespan prediction module, and a data storage module. These modules are used to implement equipment health management at the control end, and their functions and data interaction logic are the same as the corresponding modules on the tested end. Unlike the tested management unit, the Health Management Center Unit also includes a maintenance decision module. The Health Management Center Unit encompasses global optimization of fault diagnosis, lifespan prediction, and maintenance decisions, as well as the training of relevant health management models. Through collaborative processing, it improves the overall operational efficiency and decision accuracy of the health management system.

[0067] In some specific implementations, the maintenance decision module is used to analyze and decide on maintenance strategies based on the fault diagnosis results and life prediction results provided by the integrated data management module, thereby generating corresponding maintenance plans and suggestions. The maintenance decision module is equipped with a maintenance optimization model, which optionally considers factors such as cost, availability, and risk during the decision-making process to ensure the scientific and rational nature of the maintenance strategy. The maintenance decision module receives fault diagnosis data and life prediction data provided by the integrated data management module, and after completing the maintenance decision, outputs the maintenance decision results to the visualization module, providing support for the presentation and application of the maintenance plan.

[0068] In some specific implementations, the tested end and the control end transmit control commands such as test start and mode switching via a 1553B bus, and transmit real-time data, historical data, diagnostic and prediction results via Ethernet. Within the tested health management unit, the data preprocessing module transmits preprocessed data to the fault diagnosis module and life prediction module, and the fault diagnosis module and life prediction module transmit result data to the data storage module. Within the control end, the integrated data management module provides integrated data to the maintenance decision module and visualization module, the maintenance decision module transmits decision results to the visualization module, and the human-machine interaction module transmits user commands and parameters to each module.

[0069] Specifically, the health management unit at the tested end, as the core execution entity, first completes the acquisition of multimodal raw data through the sensor array and standardized bus (including RS-422 bus and AD analog interface). The raw data covers equipment operation logic data (such as switch status and command response signal), timing data (such as real-time current / voltage change curve and speed timing sequence), and status data (such as steady-state parameters such as temperature and vibration amplitude). After data acquisition, the built-in data preprocessing module performs real-time processing on the raw data: First, outlier removal, using standard deviation algorithm, box plot analysis, or Hampel filtering function to identify and remove outliers (such as jump values ​​caused by sudden interference) to ensure basic data accuracy; second, data normalization and feature extraction, employing Empirical Mode Decomposition (EMD) to adaptively decompose non-stationary time series data, combined with Principal Component Analysis (PCA) for dimensionality reduction and extraction of key feature parameters, while normalizing the extracted feature parameters to ensure compatibility with subsequent modules; third, data denoising, using a moving average algorithm to smooth high-frequency noise, or employing Singular Value Decomposition (SVD) to perform low-rank approximate filtering on the data matrix to further reduce the impact of environmental interference on data quality. After preprocessing, the generated high-quality data will be synchronously transmitted to the fault diagnosis module and life prediction module at the tested end, providing data support for subsequent analysis.

[0070] The fault diagnosis module of both the tested end and the control end (health management center unit) takes the preprocessed data from the data preprocessing module as input and employs a fault detection method based on local information increment and slow feature analysis to extract slowly varying feature components from the data, achieving precise fault location. When a fault is detected, the module automatically generates a fault diagnosis result, which includes key information such as fault type (e.g., mechanical fault, electrical fault), fault occurrence time, and fault severity level. This result is stored in the data storage module (e.g., embedded database) of the tested end. Simultaneously, the fault diagnosis result is uploaded in real time to the integrated data management module of the control end via Ethernet, ensuring that the control end promptly obtains the equipment fault status.

[0071] The integrated data management module at the control end acts as the data hub, receiving all data uploaded by the tested end via Ethernet. This data includes real-time preprocessed data, historical operating data (historical records synchronized from the tested end's data storage module), fault diagnosis results, and lifespan prediction results. The integrated data management module first integrates the received data, converting unstructured data (such as raw time-series curves) into structured formats (such as data tables and feature vectors), and establishing categorized storage indexes according to data type (such as fault data and lifespan data). Subsequently, it performs data backup operations (such as local disk array backup and off-site cloud backup) to ensure data security. Simultaneously, the integrated data management module provides standardized query interfaces (such as SQL query interfaces and API interfaces) to support data access by other functional modules or users. The integrated dataset is then sent to the maintenance decision-making module and visualization module at the control end for subsequent data applications.

[0072] The maintenance decision module takes the fault diagnosis and life prediction results output by the integrated data management module as its core input and runs a preset maintenance optimization model. This model comprehensively considers multi-dimensional constraints, including maintenance costs (such as spare parts costs and labor costs), equipment availability (to meet production or task requirements), and fault risks (such as secondary losses caused by fault escalation). It generates an optimal maintenance plan and recommendations through a general maintenance decision algorithm. The plan includes maintenance time windows, a list of maintenance parts, and maintenance priority ranking. Simultaneously, the ground-based health management center unit can perform global optimization operations based on accumulated full data (historical data from multiple devices and diagnostic prediction results): when the accuracy of the fault diagnosis or life prediction model decreases (e.g., the prediction error exceeds a preset threshold), the model is retrained (e.g., by supplementing with new data to update deep learning model parameters and optimizing the distribution assumptions of the stochastic process model); if the optimized model needs to be applied to the tested device, the updated model parameters are sent to the corresponding fault diagnosis or life prediction module on the tested device via Ethernet. Finally, the maintenance decision results generated by the maintenance decision module are sent to the visualization module on the control end for user viewing.

[0073] The visualization module provides an intuitive graphical representation of data and results: for real-time equipment status, dynamic curves (such as current time-series curves and temperature change curves) show parameter trends; for fault diagnosis results, charts (such as fault type statistical bar charts and fault occurrence time axes) present fault distribution and development, and alarm information (such as pop-up prompts and audible / visual signal linkage) triggers fault warnings; for life prediction results, reliability curves and remaining life countdowns display the equipment's life status; for maintenance decision results, optimization schemes are presented through maintenance plan Gantt charts and priority ranking tables. Simultaneously, the human-machine interface module on the control end provides a visual operation interface, supporting users to perform multiple operations: view data and results output by each module (such as historical fault records and life prediction reports); issue control commands (such as starting comprehensive equipment testing and switching equipment operating modes); set system parameters (such as fault thresholds and data acquisition frequency); and trigger data queries (such as querying the operating data of a specific device by time range). User-issued control commands are transmitted to the tested device via the 1553B bus, enabling remote control of the tested device.

[0074] The device under test (DUT) receives control commands from the control unit via the 1553B bus. These commands include commands to start comprehensive testing, time synchronization, and mode switching. Upon receiving a command, the DUT's control module first parses the command content and adjusts the device's operating mode according to the command type: for example, upon receiving a "start comprehensive testing" command, the device switches from normal operation mode to comprehensive testing mode; upon receiving a "mode switching" command, it switches to the corresponding functional mode (such as data acquisition enhancement mode or low-power mode). Subsequently, the DUT executes operations matching the command: if it is a "start comprehensive testing" command, the control sensors and bus re-acquire full-dimensional data and perform specialized tests; if it is a "update model parameters" command (from globally optimized parameters), the new model parameters are written into the fault diagnosis or life prediction module, updating the module algorithm; if it is a "time synchronization" command, clock calibration is performed with the control unit. After the operation is completed, the DUT feeds back the execution results (such as test data, mode switching status, and parameter update confirmation information) to the control unit's comprehensive data management module via Ethernet, forming a closed loop of command execution.

[0075] In this embodiment, the control process of the embedded multimodal health management system is as follows:

[0076] S1. The tested health management unit collects multimodal raw data through sensors or a bus;

[0077] S2, the data preprocessing module performs real-time preprocessing operations on the multimodal raw data and outputs high-quality data;

[0078] S3. The fault diagnosis module, based on the preprocessed data, realizes real-time fault detection and diagnosis through the fault detection model, outputs the fault diagnosis results and stores them in the data storage module, and uploads them to the control terminal at the same time.

[0079] S4. The lifetime prediction module, based on the preprocessed data, performs performance degradation modeling and lifetime prediction through the remaining lifetime prediction model, outputs lifetime prediction results and stores them in the data storage module, and uploads them to the control terminal at the same time.

[0080] S5. The control terminal receives and integrates the data uploaded by the tested health management unit, generates a maintenance strategy through the maintenance decision module, displays it through the visualization module, and supports user operation and command issuance through the human-computer interaction module.

[0081] S6. The tested health management unit receives control commands from the control terminal, executes the corresponding operations, and feeds back the execution results, forming a closed-loop management system.

[0082] In some specific implementations, the preprocessing module achieves the above functions through the following process:

[0083] S1. Outlier Removal Process: Outlier removal is a crucial step in data preprocessing, aiming to identify and handle outliers that do not conform to the data distribution. The preprocessing module first identifies and removes outliers from the three input data types:

[0084] S101. Logical data typically represents control requirements, such as the requirement that the time interval between two logical signals must be within a certain range. Preprocessing algorithms can be used to identify the occurrence times of logical signals and normalize the data. Specifically, logical data employs time identification and time difference detection methods based on logical rules to verify the logical rationality of the signal sequence and eliminate abnormal data points that do not conform to the preset logical relationships.

[0085] S102. Time-series data is a sampling sequence of physical quantities that changes continuously over time, such as data collected by sensors during equipment operation, including temperature, pressure, and vibration acceleration. Time-series data is continuous and time-varying, reflecting the dynamic operating characteristics of equipment over time. However, it is large in volume and contains noise interference and possible abrupt changes. Specifically, time-series data uses the standard deviation algorithm or the Hampel function method based on a sliding window for dynamic outlier identification. The Hampel function method effectively identifies and removes impulse noise and random outliers by calculating the local median and absolute deviation.

[0086] S103. Status data represents the operating state indicated by the device status register at a certain moment, such as the current operating state or action state of the device. The status data, combined with state transition verification and box plot method, identifies and eliminates abnormal status values ​​that do not conform to the normal state transition rules or exceed the reasonable range.

[0087] In some optional implementations, the preprocessing module employs three outlier handling methods: standard deviation algorithm, box plot, and Hample function for identification and removal. The standard deviation algorithm, by calculating the sample standard deviation, reflects the degree of deviation of data points from the mean, thus removing outliers. Based on the statistical characteristics of the data, the standard deviation algorithm can effectively identify outliers that deviate from the mean by more than a certain multiple of standard deviation. It is relatively simple to calculate and easy to implement, but its drawback is its sensitivity to the assumption of a normal distribution. When the data exhibits non-normal distribution characteristics or contains many outliers, the estimated standard deviation may be distorted, affecting the accuracy of outlier removal. For data with a slow changing trend, it may mistakenly identify normal data abrupt changes as outliers. The standard deviation algorithm is suitable for datasets that are close to a normal distribution and have a relatively low proportion of outliers, such as large amounts of repetitive measurement data collected by equipment under stable operating conditions. The application scope of the standard deviation algorithm is limited to "stable operating conditions." The data preprocessing module determines the operating conditions and only activates the standard deviation algorithm when the data is stable to avoid misjudgment.

[0088] Box plots define outlier ranges using quartiles (Q1, Q3) and interquartile ranges (IQR). Data points outside this range are considered outliers. Their advantages include not relying on the assumption of normality, providing a more objective assessment of extreme values, strong resistance to data interference, and effective identification of outliers at the tails of the data distribution. However, their disadvantages include significant fluctuations in quartile estimations for small sample data, leading to decreased accuracy in outlier identification; and difficulty in handling outliers in data with complex distributions, potentially resulting in numerous misjudgments or omissions. Box plots are suitable for situations where the data distribution is unknown or non-normal, and the sample size is relatively moderate, such as data collected during equipment startup and shutdown. The box plot described in this embodiment is used in non-steady-state, medium-sample-size scenarios, such as equipment startup and shutdown processes. The data preprocessing module dynamically calls upon data after identifying data characteristics (such as sample size and distribution shape).

[0089] The Hampel function is a sliding window-based outlier detection and removal method. Its core idea is to dynamically identify outliers using the local median and absolute deviation (MAD). It is suitable for real-time signal processing and non-Gaussian distributed data, exhibiting strong robustness against impulse noise and random outliers. A drawback is the need to select an appropriate sliding window size, which significantly impacts outlier detection performance. For data with long-term trends or periodic changes, it may mistakenly identify only a portion of the trend or periodic component as an outlier. The Hampel function is suitable for data stream processing scenarios with high real-time requirements, such as real-time outlier removal of rapidly changing data like vibration signals in online equipment monitoring systems. The Hampel function is used to process real-time data streams (such as vibration signals). The preprocessing module presets the optimal window size for different signal types through pre-analysis or experience, using the Hampel function to capture impulse noise rather than long-term trends.

[0090] S2, Data Warming and Feature Extraction

[0091] The preprocessing module employs features extraction and normalization to perform data normalization on logical data, time-series data, and state data after outlier removal; and uses EMD (Empirical Mode Decomposition) and PCA (Principal Component Analysis) for feature extraction.

[0092] EMD is a feature extraction method for nonlinear and non-stationary signals. It adaptively decomposes the original signal into multiple intrinsic mode functions (IMFs) and residual components, thereby revealing the intrinsic characteristics of the data. Its drawbacks include the potential for mode aliasing during decomposition, where signal components from different time scales are mixed within a single IMF component. It is also sensitive to high-frequency noise, which may result in a significant amount of noise in the decomposed IMF component. It is suitable for processing nonlinear and non-stationary time-series data, such as weak feature extraction in the early stages of equipment failures, separating fault features from background noise and highlighting subtle fault information. This implementation uses EMD as a deep feature extraction tool for nonlinear and non-stationary signals; it is used to extract weak early-stage fault features from signals such as vibration, complementing PCA.

[0093] PCA transforms the original features into linearly independent principal components through orthogonal transformation. These principal components are linear combinations of the original features and are ordered from largest to smallest variance, achieving the core objectives of maximizing variance and eliminating correlation, highlighting key features, and reducing data redundancy. The drawback of PCA is that it is merely a linear transformation and cannot capture non-linear relationships in the data; it is also sensitive to outliers, which may distort the direction of principal components and variance estimation. It is suitable for processing datasets with high multidimensional correlation, such as when data from multiple sensors on a device exhibit strong correlation. PCA can extract the main feature components, simplifying the data dimensions for subsequent analysis. In this embodiment, the data preprocessing module uses PCA to process large amounts of linearly correlated sensor data to quickly extract key features.

[0094] Simultaneously, a normalization method is employed for data regularization. This data preprocessing technique, which maps data to a specific range (such as [0,1] or [-1,1]) through mathematical transformations, aims to eliminate dimensional differences between different features, improve model training efficiency and accuracy, and is suitable for data fusion processing scenarios with different dimensions and numerical ranges. For example, after normalizing data of different dimensions such as temperature, pressure, and rotation speed of equipment, it can be used for unified fault diagnosis model training and prediction.

[0095] S3, Data Denoising Processing

[0096] The preprocessing module employs a moving average algorithm and SVD filtering to denoise the data. The moving average algorithm replaces the current data point with the average value of data within a fixed window in the data sequence, thus smoothing noise and preserving trend characteristics. Essentially a linear low-pass filter, it suppresses high-frequency noise but may introduce time delays. It has the advantages of being computationally simple and easy to implement, but its disadvantages include potential loss of detailed data information, such as smoothing out signal abrupt changes and high-frequency characteristic components. For data with non-stationary characteristics, the denoising effect may be unsatisfactory, and it may even introduce delay effects. It is suitable for processing data with relatively stable trends and high-frequency random noise, such as denoising temperature and pressure sensor data under stable operating conditions. For signals with high real-time requirements but low detail requirements, such as slowly varying temperature and pressure signals, the data preprocessing module uses data denoising processing.

[0097] Singular Value Decomposition (SVD) is a matrix factorization technique that decomposes any matrix into the product of three matrices, whose diagonal elements are singular values ​​(arranged in descending order). These values ​​reflect the main characteristics of the data; noise typically manifests as smaller singular values, while effective signals correspond to larger singular values. By truncating or reducing smaller singular values, noise can be filtered out while preserving the main structure of the data. The disadvantages of SVD include relatively high computational complexity, which may require a long computation time for processing large-scale data; and the need to select an appropriate truncation threshold to determine the number of singular values ​​to retain, as different threshold choices can significantly impact the denoising results. It is suitable for processing high-dimensional datasets with strong linear correlations, such as multi-channel vibration signal data from equipment. By applying SVD filtering to the data matrix, noise interference can be effectively removed, and the main characteristic components of the signal can be extracted. In this embodiment, when a need for in-depth analysis or processing of multi-channel vibration data is identified, if the current computing power allows, the data preprocessing module is controlled to activate the singular value decomposition algorithm. This method can achieve higher denoising fidelity, thereby providing a better data foundation for subsequent in-depth analysis or multi-channel vibration data processing.

[0098] In some specific implementations, the control flow of the fault diagnosis module is as follows:

[0099] S1. Perform dynamic characteristic discrimination on the preprocessed monitoring data;

[0100] S2. Faults are detected using a fault detection model based on local information increment and slow feature analysis;

[0101] S3. Use the training data to train the constructed model offline;

[0102] S4. Perform online testing using the test data. The process ends here.

[0103] Specifically, the fault detection model based on local information increment and slow feature analysis in the fault diagnosis module follows this process:

[0104] S1, Data Preprocessing

[0105] Suppose an input matrix X = [x1, x2, ..., xn] m ]∈R n×m There are n samples and m process variables. To achieve homogenization, each process variable x needs to be processed to have zero mean and zero standard deviation; that is, the processed process variable can be expressed as: in σ is the mean of the data, and σ is the standard deviation of the data. Subsequently, the input matrix X is divided into two parts: non-stationary variables... and steady-state variables And m = m ns+m s Among them, m ns and m s These represent the number of non-stationary variables and the number of stationary variables, respectively.

[0106] S2, Extraction of long-term steady-state equilibrium relationship

[0107] For an industrial process experiencing a failure, some variables become anomalous, disrupting the original long-term equilibrium. However, after modeling the non-stationary variables using cointegration analysis, the resulting residual sequence becomes stationary. For the non-stationary variables at time t, the corresponding cointegration model is established as follows:

[0108] ξ t =BX ns (t): t = 1, 2, ..., n (1)

[0109] Where, ξ t X is the steady-state residual obtained from the non-steady-state variable at time t. ns (t) is a non-stationary variable at time t, and the cointegration matrix B = (b1, b2, ..., bt) r ), where B′ is the transpose of the cointegrating matrix B. r is the number of cointegrating vectors obtained from the Johansen test.

[0110] Therefore, by fusing the steady-state residuals of both steady-state and non-steady-state variables, the joint matrix Vs is constructed as follows:

[0111]

[0112] ξ r Represents the r-th steady-state residual sequence;

[0113] Subsequently, an SFA model is established to extract slow features from the combined matrix, achieving feature-level data fusion. The following slow feature matrix S can be obtained through the SFA algorithm:

[0114] S = C_WVs (5)

[0115] Where C_W is the coefficient matrix.

[0116] However, including too many slow features can introduce too much noise and further degrade monitoring performance. To reduce the dimensionality of the slow feature matrix, a common approach is to select the top c slow features as the principal components. This is because the slower the variables change, the more intrinsic information they contain, and the slower the feature matrix Δ(x) becomes more complex. j This was selected as one of the criteria for choosing the number of slow features, as shown below:

[0117]

[0118] Among them, c eThis represents the number of slow features exceeding the threshold calculated using the slowness criterion, and card{·} represents the number of elements in a specific set. Represents the set {Δ(x)} j The q-quantile of )}, Δ(s i ) indicates the slowness, s i It is the i-th element of the slow characteristic matrix S.

[0119] Subsequently, the number c of the main slow features was calculated as follows:

[0120] c = m s +rc e (7)

[0121] In the formula, m s is the dimension of the steady-state variable, and r is the number of cointegrating vectors obtained from the Johansen test.

[0122] Therefore, the long-term steady-state equilibrium model matrix S can be extracted from the input matrix X. c It is expressed as follows:

[0123] S c = [s1, s2, ..., s k (8)

[0124] In the formula, s c These are the elements in the slow feature matrix S arranged in ascending order.

[0125] S3, Constructing Statistics

[0126] In practice, due to variable conditions, varying loads, and external noise, it is difficult to select monitoring statistics and implement accurate process monitoring based solely on current data. Therefore, this implementation proposes a local information increment by calculating a local covariance matrix established from current and historical data. The local information increment selects a fixed-length sampling window and uses a sliding window strategy to update the data. When the dynamic process enters different operating states, the inherent relationships within the signal are captured by the local covariance matrix. The structure of the local information increment is as follows: Figure 2 As shown.

[0127] For the slow eigenma matrix S k The sampling window length L is determined by cross-validation. In N t At any given time, the local data matrix formed by the sampling window is defined as follows:

[0128]

[0129] In the formula, It is the slow characteristic matrix S c The corresponding element in.

[0130] Similarly, in N t At time +1, the local data matrix formed by the sampling window is defined as follows:

[0131]

[0132] By combining and The common part, the local matrix is ​​defined as follows:

[0133]

[0134] Based on the local matrix, N t Local covariance matrix at time t The calculation is as follows:

[0135]

[0136] Among them, the intermediate variable matrix It is N t The average vector at time N. t The local covariance matrix at time +1 is calculated as follows:

[0137]

[0138] similar,

[0139] The local information increment matrix is ​​calculated as follows:

[0140]

[0141] In fact, This represents the difference matrix of the local covariance matrix. Considering redundant information and complex noise, it is unwise to observe every element of the local information increment matrix. Instead, the average local information increment is calculated as follows:

[0142]

[0143] In the formula, p c It is a difference matrix. The dimension. Obviously, It can effectively track the fluctuations of all variables. Therefore, it is the best choice. As a statistical measure.

[0144] S4. Constructing a dynamic threshold

[0145] This implementation uses an S-shaped membership function to convert basic input parameters into fuzzy variables; a dynamic threshold is established based on the fuzzy membership function; and to measure the correlation between the indicator and different evaluation criteria, a selection is made. As a membership degree, it is shown below:

[0146]

[0147] In the formula, κ represents the coefficient, and γ is the input variable of the membership function. Based on the above fuzzy membership function, the fused output value T is shown below:

[0148]

[0149] In the formula, C_W i It is a characteristic matrix element, ω i It is the membership degree, γ i It is the statistic selected in section S3.

[0150] Step 1: For the training data X = [x1, x2, ..., x...] m ]∈R n×m A sliding window is created and the corresponding mean of local information increments is calculated. Then, kernel density is used to estimate control limits and set as fixed thresholds.

[0151] Step 2: For the test data First, data preprocessing is performed based on the training data. Then, a new sliding window is constructed, including the last (L-1) samples of X and... The first sample. Combine the new sliding window with the previous one. The statistic can be calculated and defined as follows. p represents the dimension. Taking i-1 as an example, the sample... The corresponding sliding window is The previous sliding window was

[0152] Step 3: Considering that the closer to the current time, the stronger the representativeness of process monitoring, an S-shaped membership function is introduced to fully mine the hidden information of the sequence data, and a dynamic threshold is applied. The calculation is as follows:

[0153]

[0154] Step 4: Comprehensive dynamic threshold T D The calculation method is as follows:

[0155]

[0156] T D The final comprehensive threshold result that needs to be calculated is... It is the first type of basic dynamic threshold, where H is the threshold coefficient, satisfying 0 < H ≤ 1.

[0157] S5, Fault Detection

[0158] In the fault detection model based on local information increment and slow feature analysis, the fault detection logic strategy is as follows:

[0159]

[0160] Meanwhile, the updated strategy is as follows:

[0161]

[0162] Taking i=1 as an example:

[0163] (a) If So the sample Samples identified as normal are entered into a sliding window to update the data. Subsequently, the samples... The local information window is And the previous window is

[0164] (b) If So the sample The sample was identified as faulty and excluded from the sliding window. Subsequently, the sample... The local information window is And the previous window is P1 L =[x n-L+1 x n-L+2 , ..., x n ].

[0165] In some specific implementations, the control flow of the lifetime prediction module is as follows:

[0166] S1. Start the lifetime prediction process and load the data processed by the data preprocessing module;

[0167] S2. Preprocess the raw data, including smoothing, noise reduction, filtering, etc.

[0168] S3. Degradation trend judgment is performed on the preprocessed monitoring data;

[0169] S4. Determine whether the monitoring data shows a significant degradation trend: If the monitoring data does not show a significant degradation trend, proceed to S5; if the monitoring data shows a significant degradation trend, proceed to S9.

[0170] S5. Employ a deep learning-based method for predicting remaining lifetime.

[0171] S6. Use the training data to train the constructed network offline;

[0172] S7. Conduct online testing using test data;

[0173] S8. Obtain the remaining lifetime prediction results of the system under no degradation trend. The process ends.

[0174] S9. Lifetime prediction using a stochastic process-based remaining lifetime prediction model;

[0175] S10. Construct four degradation models (linear, power, logarithmic, and exponential);

[0176] S11. Lifetime prediction is performed using a deep learning-based remaining lifetime prediction model.

[0177] S12. Predict remaining lifetime based on the selected optimal model;

[0178] S13. Obtain the fusion result of the remaining lifespan of the system under the trend of degradation, and the process ends.

[0179] Specifically, in the lifespan prediction module, the residual life prediction model based on stochastic processes is grounded in probability and statistics theory. It uses a stochastic process model to model the evolution of equipment health status, thereby naturally obtaining the probability distribution of residual lifespan and effectively quantifying the uncertainty of residual lifespan estimation results. By constructing a stochastic degradation model that matches actual degradation behavior, it is possible not only to accurately depict the degradation trajectory of system performance over time but also to further derive the probability distribution characteristics of residual lifespan, providing theoretical support for equipment health management and maintenance decisions. In specific implementations, this invention uses linear Wiener process models, nonlinear Wiener process models, and exponential stochastic degradation models to describe the equipment health status degradation process, and evaluates and optimizes these candidate models based on historical monitoring data to determine the model structure that best reflects actual degradation characteristics. Under the first-passage time (FST) framework, an analytical expression for the residual lifespan under the corresponding stochastic degradation model is derived. Furthermore, by combining real-time acquired monitoring data, a parameter identification strategy that combines maximum likelihood estimation and Bayesian estimation is adopted to identify unknown parameters in the model offline and update them dynamically online, thereby achieving high-precision and adaptive prediction of the remaining service life of the equipment and ensuring the safety and reliability of the equipment during its service.

[0180] S1, Degeneracy Modeling

[0181] S101. The linear Wiener process model mainly describes the degradation of a system that increases linearly with time and exhibits random fluctuations. In this implementation, the degradation process {X(t), t≥0} of the linear stochastic degradation system is described by the Wiener process.

[0182] X(t) = λt + σ B B(t) (22)

[0183] Where λ is the drift coefficient, σ B >0 is the diffusion coefficient, and {B(t), t≥0} is the standard Brownian motion that reflects the time-varying randomness of the degradation process.

[0184] Degradation modeling based on the Wiener process {X(t), t≥0} shows that the expected value of the system degradation is a linear function of time, i.e., E[X(t)]=λt. Therefore, the drift parameter λ is the degradation rate of the system, which is closely related to the degradation evolution process. Furthermore, the variance of the corresponding degradation process is var[X(t)]=σ B 2 t represents the degradation uncertainty related to time t.

[0185] S102, Exponential Stochastic Degradation Model

[0186] The model corresponding to the exponential stochastic degradation process {X(t), t≥0} in this embodiment can be described as follows:

[0187]

[0188] Where φ is a constant, σ B >0 represents the diffusion coefficient, a deterministic model parameter, while θ′ and β′ are random variables characterizing individual differences within the system, and {B(t), t≥0} represents standard Brownian motion.

[0189] For exponential models, S(t) is defined at time t as follows:

[0190]

[0191] Where, let θ = lnθ′, θ is the natural logarithm of the random variable θ′, and β is the parameter of the random variable after drift correction.

[0192] To simplify the analysis, this implementation takes φ = 0. Of course, φ can be any constant. Even if φ ≠ 0, it can still be transformed into a workable form through translation.

[0193] S103, Nonlinear Wiener Process Degradation Model

[0194] In practice, especially when the system's operating environment, load, or working conditions change (for example, the propagation of fatigue cracks may accelerate or decelerate during crack propagation), the system's degradation rate will change over time, exhibiting characteristics of a non-uniform degradation rate. This implementation uses a nonlinear stochastic process {X(t), t≥0} to describe the degradation process of a nonlinear stochastic degradation system. Specifically, X(t) represents the degradation at time t, and the nonlinear Wiener process model is defined as follows:

[0195]

[0196] The degradation process X(t) is driven by the standard BM process B(t); μ(τ; θ w ) and σ B Representing the drift term and diffusion coefficient of the degradation process, respectively, τ is the integration variable, and θ is the time variable; w For the unknown parameter vector; μ(τ; θ) w μ(τ; Δ) is a nonlinear function of time t, used to characterize the nonlinear features of the model. Furthermore, μ(τ; Δ) has different functional forms. w It can describe different forms of nonlinear stochastic degradation processes. Without loss of generality, we mainly consider the case of X(0) = x0 = 0.

[0197] Nonlinear Wiener process models are typically considered to have two typical forms: power function form and exponential function form. Their drift coefficient expressions are as follows: In these two forms of drift coefficients in nonlinear Wiener process models, a is a scale parameter used to adjust the magnitude of the drift; b is a shape (or rate) parameter used to control the rate or form of drift change over time.

[0198]

[0199] S2, Parameter Identification

[0200] To construct the aforementioned degradation model and predict its remaining life, it is necessary to estimate the unknown parameters of the model. Based on historical equipment measurement data, this implementation method uses maximum likelihood estimation to obtain initial estimates of the model parameters. For field-tested equipment, when new data is collected, Bayesian parameter estimation can be used to update the model parameters. Here, this implementation method takes the nonlinear Wiener process model as an example and uses a fusion method of maximum likelihood and Bayesian estimation to estimate and update the unknown parameters of the model. The specific steps of the algorithm are as follows:

[0201] Step 1: Estimate model parameters based on historical data

[0202] Assume there are N systems, and the degradation data of the nth system are respectively in At time, m is obtained, where m n Let represent the number of measurements for the nth system, where n = 1, ..., N. Therefore, the nth system at measurement time t... n，j The degenerate state can be represented as

[0203]

[0204] Where j = 1, ..., m n a nLet a be the iid (independent and identically distributed) realization of a, and follow a normal distribution, μ a The mean, Let Variance be the variance.

[0205] To simplify notation, define functions In the model μ(τ;θ) w ) = abt b-1 and model μ(τ; θ) w )=ab exp(bt) are respectively and make T n This represents the time set of the nth system degradation data; T n，j The time function value of the j-th measurement of the nth system; Let X represent the m-th monitoring point of the n-th device, and ()' denotes the transpose of the matrix within the parentheses. n Let X be the data sequence of the total lifecycle degradation of the nth system. Based on the independent increment property of standard Brownian motion, we know that X... n It follows a multidimensional normal distribution, and its mean is The sum of the covariance matrix ∑ n They are respectively:

[0206]

[0207] Among them, Q n Ω is the time structure matrix. n It is a random covariance matrix.

[0208]

[0209] Since the degradation processes of different individual systems are independent of each other, then in data X n Below about parameter vectors The log-likelihood function can be expressed as:

[0210]

[0211] in, Let ∑ be the covariance matrix n The inverse matrix;

[0212] Consider the case where the degradation measurement time and the number of measurements are the same for all systems, i.e., for all individual systems, m n It is a constant, and for any n, l = 1, ..., N, we have t n，j =t l，j .

[0213] In this case, T nΩ n ,∑ n The subscripts can all be removed. Therefore, the above equation regarding μ a and standard deviation σ a The first-order partial derivative can be simplified to:

[0214]

[0215] Then, for each For the specific values ​​of μ and b, let the two partial derivatives on the equation be equal to 0, then with respect to μ... a and σ a The maximum likelihood estimation result can be expressed as:

[0216]

[0217] Based on this, σ B and b about μ a and σ a The likelihood function of the maximum likelihood estimate can be expressed as:

[0218]

[0219] Then, σ B The maximum likelihood estimates of σ and b can be obtained by maximizing the likelihood function of the above profile using a two-dimensional search method. Then, the obtained σ... B Substituting the maximum likelihood estimate of b back into μ a and σ a The expression for μ can be used to obtain the corresponding μ. a and σ a The maximum likelihood estimate.

[0220] This implementation provides three degradation models for selection. To choose the optimal degradation prediction model, the model form can be determined using the AIC criterion or the mean squared error (MSE). g represents the number of unknown parameters, i.e.

[0221]

[0222] l n，k This represents the remaining lifetime of the nth system during the kth prediction. Indicates the actual remaining lifespan.

[0223] Step 2: Updating Model Parameters Based on Bayesian Estimation

[0224] Let the prior distribution of the random parameter λ be p(λ). According to the properties of the standard BM, for a given λ, X 0:k The sample distribution is a multivariate normal distribution, and its joint PDF is

[0225]

[0226] P() is the probability density function, k represents the k-th measurement, and t j This indicates the specific time of the j-th measurement.

[0227] Within the Bayesian framework, to compute the posterior distribution P(λ|X) of the random parameters 0：k Assume the prior distribution of λ is... That is, λ follows a mean of μ0 and a variance of μ. It follows a normal distribution.

[0228] At the current time t k Based on monitoring data X 0：k And in the description of the random degradation process, the posterior distribution of the random parameter λ can be updated using the Bayesian theorem, specifically:

[0229]

[0230] ∝ represents the proportional relationship between the posterior distribution and the prior distribution multiplied by the likelihood function in Bayesian inference; x j -x j-1 It is the increment of degradation between the j-th and (j-1)-th measurements; μ θ，k It is the posterior mean of the model parameter θ at time k; σ is the posterior variance of the random parameter λ; λ，k It is the posterior variance of the random parameter λ.

[0231] Due to λ|X 0：k It is a conditional random variable, P(λ|X) 0：k λ|X is the probability density function; therefore, based on the characteristics of a normally distributed random variable, we know that λ|X 0：k It is normally distributed, and has

[0232]

[0233] in,

[0234]

[0235] Among them, t k Let σ represent the k-th monitoring time. k μ represents the degradation data corresponding to the k-th monitoring point. λ，k σ represents the mean value of the drift coefficient corresponding to the k-th monitoring point. λ，k It is the standard deviation corresponding to the k-th monitoring point λ. λ is the initial value of the random parameter λ; from the above equation, it can be seen that the posterior estimate of the random parameter λ can be updated after new state monitoring data is acquired.

[0236] S3. Derivation of Remaining Lifetime Distribution

[0237] For control systems with degradation trends, in order to achieve lifetime prediction, this implementation method considers the use of a stochastic process approach, namely a lifetime prediction model based on degradation process modeling.

[0238] The lifetime prediction model based on degradation process modeling aims to model the degradation process using acquired performance test data and establish the relationship between test data and remaining lifetime. It proposes to use linear Wiener process, exponential stochastic process and nonlinear Wiener process among stochastic processes to analyze the test data. When the performance test data has a linear degradation trend, the linear Wiener process or exponential stochastic process is selected for modeling. When the performance test data has nonlinear degradation characteristics, the nonlinear Wiener process is selected for modeling.

[0239] (1) Linear Wiener process model

[0240] The degradation model based on the Wiener process can be described as follows:

[0241] X(t)=X(0)+λt+σ B B(t) (45)

[0242] Where t represents time, X(0) is the initial degradation level, and λ and σ B Let X(0) represent the drift coefficient and diffusion coefficient, respectively, and B(t) represent standard Brownian motion. Without loss of generality, let X(0) = 0. For cases where X(0) is not zero, we can transform it into the case where X(0) = 0. The model parameters, including the drift coefficient and diffusion coefficient, can be determined based on historical test data. The equipment lifetime in the sense of first arrival time can be defined as:

[0243] T=inf{t:X(t)≥ω|x0<ω} (46)

[0244] Where the failure threshold is ω; then the probability density function PDF and cumulative distribution function CDF with lifetime distribution are expressed as follows:

[0245]

[0246] Φ represents the cumulative distribution function of the standard normal distribution;

[0247] Similarly, based on the concept of the first arrival time of a stochastic process {X(t), t≥0}, the system is considered to have reached its end of life when the failure threshold ω is first reached. Therefore, based on the observed data X... 0k ={x0, x1, x2, ..., x k The system at time t k Remaining lifespan Lk Defined as

[0248] L k =inf{l k :X(l k +t k )≥ω|X 0：k} (48)

[0249] l k For the system at time t k Remaining lifespan L k The variable is inf; inf is the infimum.

[0250] At this point, the PDF and CDF of the remaining lifetime distribution under fixed parameter conditions are expressed as follows:

[0251]

[0252] In order to characterize P(λ) which represents the uncertainty in the drift coefficient estimation k |X 0：k Integrating it into the remaining lifetime estimation, at time t k Using X 0：k and λ k The posterior probability distribution can provide the PDF and CDF of the remaining lifetime distribution under random parameters.

[0253]

[0254] P k|k It means that the random parameter λ k In t k The posterior variance at time t is used to quantize the parameter λ. k The degree of uncertainty in the estimate Represents the random parameter λ k The posterior mean.

[0255] (2) Exponential stochastic degradation model

[0256] In general, the model corresponding to the exponential stochastic degenerate process {X(t), t≥0} can be described as follows:

[0257]

[0258] Where φ is a constant, σ B >0 represents the diffusion coefficient, a deterministic model parameter, while θ′ and β′ are random variables characterizing individual differences within the system, and {B(t), t≥0} represents standard Brownian motion.

[0259] For exponential models, S(t) is defined at time t as the time function after a logarithmic transformation of the original degenerate data. The nonlinear exponential degradation process is as follows:

[0260]

[0261] Where θ = lnθ′, θ is the natural logarithm of the random variable θ′, and β is the parameter of the random variable after drift correction.

[0262] For t k For a control system at a given time, if θ, β, and the current observation s(k) = ln x k This can transform the original degradation process into...

[0263] S(t) = s k +β(tt k )+σ B (B(t)-B(t k )), t≥t k (54)

[0264] Given θ, β and the current observation s(k), t k The remaining lifetime at time ω can be calculated by determining the first arrival time exceeding the threshold ω for the following stochastic process:

[0265] S′(l k ) = s k +βl k +σ B W(l k ), l k ≥0 (55)

[0266] Among them, W(l k )=B(t k +l k )-B(t k ), representing t k +l k The degenerate state after the change of time.

[0267] At this time, {W(l k ), l k ≥0} is still a standard Brownian motion. Therefore, {S′(l k ), l k ≥0} is also the drift portion, βl k The initial value is S′(0)=s k A stochastic process driven by a standard BM. Under fixed parameter conditions, t... k The conditional PDF and conditional CDF for the remaining lifetime estimated at each time point are as follows:

[0268]

[0269] Indicates remaining lifetime L k The conditional probability density function (PDF) of the standard normal distribution, and the cumulative distribution function (CDF) of the standard normal distribution; Conditional cumulative distribution function (CDF) of remaining lifetime.

[0270] Similarly, for the exponential stochastic model, the PDF and CDF of the remaining lifetime estimate under the stochastic parameters obtained based on s(k) are as follows:

[0271]

[0272] (3) Nonlinear Wienex process model

[0273] The nonlinear Wiener process model is defined as follows:

[0274]

[0275] The degradation process X(t) is driven by the standard BM process B(t); μ(τ; θ w ) and σ B θ represents the drift coefficient and diffusion coefficient of the degradation process, respectively; w A parameter vector with unknown parameters; μ(τ; θ) w μ(τ; θ) is a nonlinear function of time t, used to characterize the nonlinear features of the model. Furthermore, μ(τ; θ) in different functional forms... w It can describe different forms of nonlinear stochastic degradation processes.

[0276] Based on the concept of first arrival time, lifetime T can be defined as...

[0277] T=inf{t:X(t)≥ω|x0<ω} (60)

[0278] Without considering the random parameter θ w Under the influence of the following conditions, for a random degradation process {X(t), t≥0}, the PDF approximate analytical representation of the corresponding lifetime distribution is as follows:

[0279]

[0280] Among them, S B (t) represents the time-varying boundary.

[0281]

[0282] If there is a model M1μ(τ;θ) w ) = abt b-1 Then the PDF of the lifetime distribution is

[0283]

[0284] If there is a model M2μ(τ;θ) w If ) = ab exp(bt), then the PDF of the lifetime distribution is

[0285]

[0286] According to the definition of system lifetime, the system at the current time t is defined as... k Remaining lifespan L k for

[0287] L k =inf{l k :X(t) k +l k )≥ω} (65)

[0288] Under the same conditions, if the system at t k The degenerate state at time x k =X(t) k ), in t k The PDF of the remaining lifetime of model M1 at time point is as follows

[0289]

[0290] in,

[0291]

[0292] η(l k () represents the time function transformation term of the remaining lifetime in a power-law degradation model. Similarly, the PDF of the estimated remaining lifetime under model M2 is...

[0293]

[0294] In the above formula, for ease of representation, where, Because the degradation trajectories differ among individuals, θ w In this context, parameter 'a' represents a random parameter used to characterize the differences between individuals, and it follows a mean of μ. α variance is The distribution is Gaussian, and b is a fixed parameter whose value is the same for each individual, used to characterize the common features of similar systems.

[0295] Therefore, for a stochastic degenerate process {X(t), t≥0}, considering models M1 and M2 respectively, if The estimated system lifetime PDFs under model M1 and model M2 are as follows:

[0296]

[0297] In the above formula, for ease of representation, γ(t) = exp{bt}-1, β(t) = exp{bt}-bt exp{bt}-1.

[0298] Furthermore, for a stochastic degenerate process {X(t), t≥0}, considering models M1 and M2 respectively, if... The system at t k The degenerate state at time x k =X(t) k ), in t k The PDF of the remaining lifetime of model M1 at time point is as follows

[0299]

[0300] In the above formula, for ease of representation, where, ω k =ω-X(t) k ).

[0301] Similarly, the PDF of the estimated remaining lifetime under model M2 is...

[0302]

[0303] In the above formula, for ease of representation, where,

[0304] In practical engineering, if there is not enough prior information to determine which type of nonlinearity the degradation process is, several common nonlinear models can be provided. Then, the AIC criterion or MSE can be used to select the model form that is closest to the actual degradation. Finally, based on the above conclusions, the equipment life and remaining life can be predicted.

[0305] Specifically, the deep learning-based remaining life prediction model first acquires historical test data of key performance indicators and statistical data on failures of key components through a multi-sensor system deployed on critical equipment, and then uses expert knowledge to perform qualitative analysis of the information. Next, it preprocesses the raw data using normalization methods to transform it into a quantitatively analyzable data format, taking into account the characteristics of different data types. Finally, deep learning models are built based on the performance indicators. The trained networks are then used to predict the remaining life of critical equipment under the same operating conditions. For different operating conditions, model fine-tuning and other methods are used to continuously optimize the model structure and parameters, thereby achieving the prediction of the remaining life of critical equipment.

[0306] The deep learning-based remaining lifespan prediction model is specifically as follows:

[0307] S1. Acquire historical test data of key performance indicators and statistical data of key component failures through a multi-sensor system deployed on key equipment, and perform qualitative analysis of the information using expert knowledge.

[0308] Based on the characteristics of different data, data cleaning, noise reduction and normalization methods are used to preprocess the raw data and transform it into a data form that can be quantitatively analyzed.

[0309] Using NumPy and Pandas library functions, data cleaning is performed for different data sizes and scales. Valid data is selected, outliers are removed, and missing values ​​are filled to improve data quality and the accuracy of subsequent data analysis. Necessary noise reduction methods such as mean filtering, wavelet transform, and singular value decomposition are employed. Necessary normalization methods such as min-max standardization and Z-score standardization are also used.

[0310] When normalizing the raw data, it is considered that different monitoring quantities have different dimensions, and that the activation function has a limited sensitivity range. The normalization formula is as follows:

[0311]

[0312] Where, x i，j (1≤j≤n) represents the monitoring data of the j-th group belonging to the i-th device. x represents i，j The normalized value, Let represent the maximum and minimum values ​​of the raw measurement data from the i-th device, respectively. The normalized data range is [0, 1].

[0313] S2. For data without degradation trends, build deep learning models based on performance metrics. The deep learning models used include DBN, CNN, LSTM, and GRU. Specifically:

[0314] 1) Deep Belief Network (DBN)

[0315] DBN is a probabilistic generative model, a deep network composed of multiple stacked Restricted Boltzmann Machines (RBMs) and a classification (or regression) layer. It achieves optimal model training through greedy forward learning combined with gradient descent for backward fine-tuning. Taking a DBN consisting of three RBMs as an example, the network structure is as follows... Figure 3 As shown.

[0316] Each RBM consists of a visible layer and a hidden layer. The visible layer receives data, and the hidden layer extracts features. The output of the previous RBM layer serves as the input to the next layer. Figure 4 Taking RBM1 as an example, its input is the complete generated data. The state vector corresponding to the visible layer v0 is v = (v1v2, ..., v2v3). m ), v mThe visible layer represents neurons, and the hidden layer's state vector is represented as h = (h1, h2, ..., h...). n Let ), representing a neuron in the hidden layer. Then the energy function of RBM1 is:

[0317] E n (v,h)=-a s 'hb s ′vv′w s h (73)

[0318] In the formula, a s b represents the bias vector of the view layer. s w represents the bias vector of the hidden layer. s v' represents the weight matrix between the visible layer and the hidden layer, and v′ represents the reconstructed state vector.

[0319] Based on the energy function, the model probability distribution can be obtained as follows:

[0320]

[0321] Z represents the partition function, which is the normalization constant that yields a probability distribution sum of 1;

[0322] It can be seen that P(v, h) and the energy function E n (v, h) are inversely proportional, and the smaller the energy, the more stable the system. Therefore, the optimization objective of pre-training is:

[0323] L oss =maxP(v,h) (75)

[0324] L oss As the loss function, directly calculating P(v, h) has a very high computational complexity. However, RBMs have inter-layer constraints, and neurons in the same layer are independent of each other. Therefore, the conditional distribution function can be obtained as follows:

[0325]

[0326] Where h represents a single node (neuron); h j b represents the activation probability of a single node. sj Here is the hidden layer bias matrix; a si w is the view layer bias matrix; ij v is the visual layer weight matrix; i This is the hidden layer weight matrix.

[0327] Let there be two sets of unknown parameters ψ = (a s b s The objective function can be equivalently expressed as:

[0328]

[0329] Among them, v i Let ψ represent the visible layer state vector of the i-th training sample, and let ψ represent the set of all unknown parameters of the restricted Boltzmann machine.

[0330] 2) Convolutional Neural Network (CNN)

[0331] CNN networks mainly consist of convolutional layers and pooling layers, possessing superior performance characteristics such as parameter sharing and sparse connections. Their powerful feature extraction capabilities give them unique advantages in data denoising, making them suitable for processing massive amounts of equipment monitoring data in industry. One-dimensional CNNs are primarily used for processing and analyzing time-related sequences. The specific process of convolutional operations is as follows... Figure 4 As shown. Figure 4 The four convolutional kernels represent four feature extractors. The four convolutional kernels with a size of 3×1 perform traversal convolution operations on the input (7×1 in length) with a stride of 1. Since no zero padding is set, four output features with a length of 5×1 can be obtained in the end.

[0332] The formula for calculating convolutional layers is as follows:

[0333]

[0334] In the formula, For convolution operations, For the i-th weight on the j-th convolutional kernel, φ s Learnable bias, f(·) is the ReLU activation function. Represents the i-th feature of the l-th layer. After convolution kernel The i-th feature of the (l+1)-th layer obtained after operation and activation by the activation function.

[0335] 3) Long Short-Term Memory (LSTM) network

[0336] LSTM was designed to address the problem of long-term dependencies. It employs a gated structure where each memory cell's temporal data is controlled by input, forget, and output gates. This allows for flexible storage and retrieval of input data, increasing the temporal utilization and prediction length, establishing a long-term connection between future moments and historical data, and resolving the vanishing and exploding gradient problems. The hidden layer structure of an LSTM network is a Long Short-Term Memory (LSTM) block, which consists of three control gates and a single cell structure. The specific structure is as follows: Figure 5 As shown;

[0337] An LSTM network is used to build the prediction model. For each training sample, the forget gate first selectively forgets the input of the previous node:

[0338] f t =σ s (W f x t +U f h t-1 +b f (79)

[0339] In the formula, σ s The sigmoid activation function maps the input to the (0, 1) interval, providing the "gating logic" for the forget gate. t h represents the input sequence data at time t. t-1 W represents the hidden layer state at time t-1. f U f b f These are parameters to be updated.

[0340] The input gate selectively remembers the current input and updates the cell state:

[0341] i t =σ s (W i x t +U i h t-1 +b i (80)

[0342] C t =f t *C t-1 +i t *tU C h t-1 +b C )

[0343] In the formula, x t b represents the input sequence data at time t. i i represents the bias matrix of the corresponding layer. t C is the output of the input gate at time t. t C t-1 Represent the cell states at the current time step and the previous time step, respectively. * denotes the vector inner product. W i U i b i W C U C b C These are parameters to be updated.

[0344] Output gates control the output of updated states:

[0345]

[0346] In the formula, W o U o b o For parameters to be updated, o t h represents the output value of the output gate. t This represents the hidden state of the LSTM.

[0347] The output of the last layer of neurons is connected to the regression layer to obtain the prediction result:

[0348]

[0349] Let W be the output mapping function of the LSTM network, and let X be the weight matrix. i+d-1 This represents the actual input data to the LSTM; b s This represents the bias matrix of the corresponding layer.

[0350] 4) Gated Recurrent Unit (GRU)

[0351] GRU merges the input gate and forget gate of LSTM into an update gate. From the perspective of internal connections, GRU converts the cell state c in LSTM into an update gate. t The hidden state h of LSTM t The merging of GRU neurons can, to some extent, be understood as a variant of the LSTM network. The mathematical expression for a GRU neuron is as follows: Figure 6 As shown:

[0352]

[0353] In the formula, U rh W rx b r To reset the weight matrix of the gate (U) rh Corresponding historical hidden state h t-1 W rx Corresponding to the current input x t ) and bias term b r The sigmoid activation function provides the decision logic for the reset gate. It is the weight matrix of the candidate hidden state ( Corresponding historical hidden state h t-1 , Corresponding to the current input x t and bias terms Candidate information is generated using the tanh activation function. zh W zx b z Update the gate weight matrix (U) zh Corresponding historical hidden state h t-1 W zxCorresponding to the current input x t ) and bias term b z The sigmoid activation function provides the decision logic for the update gate. t With z t These represent the reset door and the update door, respectively. r represents the memory state of a cell at time t. t The memory portion can be controlled to discard irrelevant information in order to update the final hidden state h. t .

[0354] GRU has a simpler structure than LSTM, resulting in higher computational efficiency, but their model performance is similar. In practical applications, LSTM networks have stronger learning capabilities for larger datasets, while GRU networks are typically used for processing small-batch time-series datasets.

[0355] This implementation uses a Bidirectional Long Short-Term Memory (BiLSTM) network as the prediction network in the deep learning model. BiLSTM networks can simultaneously learn the forward and backward temporal information of the input data, enabling deeper mining of the temporal information contained in degraded data. This solves the problems of vanishing gradients, exploding gradients, and long-term dependencies. Its network structure is as follows: Figure 7 As shown. Compared to traditional Long Short-Term Memory (LSTM) networks, this network processes the input sequence through a network and a feedforward layer, extracting deep features of the sequence while simultaneously representing short-term dependencies, thus fully utilizing the information contained in past and future data.

[0356] The implementation process of BiLSTM network is shown in the following equation:

[0357]

[0358] In the formula, This indicates the output value of the forward output gate. Indicates the input value of the input gate. This represents the forgotten value from the forget gate, with → and ← indicating forward and backward propagation respectively. The network output is:

[0359]

[0360] The offline historical data processed by S2 were subjected to sliding time windows to construct training and testing data. The input structure is generally three-dimensional (batch size, time step, input dimension), the prediction method is single-step prediction, and the output is the predicted remaining lifetime (RUL). Based on the above prototype software system, a BiLSTM network was built using Python for RUL prediction. The ReLU function was chosen as the non-linear activation function. Through model fine-tuning and multiple training iterations, appropriate network parameters and the model were selected and retained.

[0361] S3. Determine the probability density function of the RUL of the degraded device, etc. Traditional deep learning models using dropout are equivalent to corresponding Bayesian deep learning models based on variational inference, which can characterize the uncertainty of the prediction results through the randomness of the weights and give an interval estimate of the predicted RUL.

[0362] Specifically, the PDF of the RUL prediction results under the deep learning model is as follows:

[0363]

[0364] Where P() is the conditional probability density function; For newly entered data, For the predicted RUL, X represents the input data of the training set, Y represents the corresponding RUL label in the training set, and N represents the total number of samples. This is a first-order sampling result based on the optimal variational distribution.

[0365] After training the RUL prediction network using the training set, inputting new monitoring data will yield the corresponding RUL prediction results. For ease of subsequent representation, the probability density function for predicting RUL is simplified to f(l k |X 1：k ), where l k For t k RUL, X, time-predicted 1：k The input monitoring data. From the probability density function of RUL, the cumulative distribution function F(l) of RUL can be further obtained. k |X 1：k ) and reliability function R(l k |X 1：k The following are the details:

[0366]

[0367] R(l k |X 1：k )=1-F(l k |X 1：k (88)

[0368] The joint decision-making model proposed in this embodiment is based on the RUL prediction information of multi-dimensional degraded equipment. Therefore, as long as the remaining life PDF is obtained, the corresponding cumulative distribution function and reliability function can be further derived, and a reasonable maintenance strategy can be formulated on this basis to ensure the safe and stable operation of the equipment.

[0369] S4 calculates the trust score of the deep learning model. Through similarity analysis of offline historical data and online field data, the trust score of the deep learning model in this embodiment is calculated. Multiplying this result by the RUL obtained from training with Droupt yields the final RUL value of the deep learning model.

[0370] For one-dimensional monitoring sequences, the Pearson correlation coefficient, also known as the Pearson product-moment correlation coefficient, can be used. It is a linear correlation coefficient and one of the most commonly used. Denoteed as R, it reflects the degree of linear correlation between two variables X and Y. The R value ranges from -1 to 1; a larger absolute value indicates a stronger correlation. It is applicable to continuous variables. Correlation coefficients and their strength are generally classified into five categories: 0.8-1.0 (extremely strong correlation), 0.6-0.8 (strong correlation), 0.4-0.6 (moderate correlation), 0.2-0.4 (weak correlation), and 0.0-0.2 (extremely weak correlation or no correlation).

[0371] Specifically, the Pearson correlation coefficient between two variables is defined as the product of the covariance of the two variables and their standard deviations:

[0372]

[0373] μ X With μ Y σ represents the mean of variables X and Y. X and σ Y represents the standard deviation of variables X and Y; E represents the expectation operator, used to calculate the statistical expectation of a function of random variables; X and Y represent two random variables to be analyzed (such as equipment degradation monitoring values, operating time, or variables between different degradation characteristics), which are the objects of Pearson correlation coefficient analysis, used to measure the degree of linear correlation between them.

[0374] The above formula defines the population correlation coefficient, commonly represented by the lowercase Greek letter ρ. By estimating the sample covariance and standard deviation, the sample correlation coefficient (sample Pearson coefficient) can be obtained, denoted by the uppercase letter R.

[0375]

[0376] R can also be derived from (X) i Y i Estimating the mean of the standard scores of the sample points yields an expression equivalent to the above formula:

[0377]

[0378] in, and σ X They are X i The standard score, sample mean, and sample standard deviation of the sample; and σ Y They are Y i The standard score, sample mean, and sample standard deviation of the sample; n is the total number of samples (e.g., the number of degenerate data points), and i is the sample index (from 1 to n).

[0379] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. An embedded multimodal device health management microsystem, characterized in that, The device includes a health management unit under test, used to predict the lifespan and diagnose faults of the target health management device, and send the prediction or diagnosis results to the control terminal. The health management unit under test includes a data preprocessing module, a fault diagnosis module, a lifespan prediction module, and a data storage module; the health management unit under test establishes a communication connection with the control terminal. The data preprocessing module is used to receive multimodal raw data and perform real-time preprocessing, and output the processed data to the fault diagnosis module and the life prediction module. The fault diagnosis module is used to perform real-time detection and diagnosis of equipment faults from the preprocessed data and output the fault diagnosis results. The life prediction module is used to perform equipment performance degradation modeling and remaining life prediction on the preprocessed data, and output the life prediction results. The data storage module is used to store preprocessed data, fault diagnosis results, and life prediction results, and to interact bidirectionally with the control terminal.

2. The embedded multimodal device health management microsystem according to claim 1, characterized in that, The multimodal raw data includes logic data, timing data, and state data; the logic data includes device switching status and command response signals; the timing data includes current-voltage change curves and rotational speed timing sequences; and the state data includes steady-state parameters such as temperature and vibration amplitude.

3. The embedded multimodal device health management microsystem according to claim 1, characterized in that, The real-time preprocessing operations of the data preprocessing module include outlier identification and removal, data normalization and feature extraction, and data denoising. Outlier identification and removal are used to remove outliers from the data, data normalization and feature extraction are used to extract key features and perform normalization, and data denoising is used to reduce the impact of environmental interference on data quality.

4. The embedded multimodal device health management microsystem according to claim 1, characterized in that, The diagnostic scope of the fault diagnosis module includes the identification of minor faults and the monitoring of multivariate statistical processes. The fault diagnosis module is deployed with a fault detection model based on local information increment and slow feature analysis. The fault diagnosis results include fault type, occurrence time and severity level.

5. The embedded multimodal device health management microsystem according to claim 1, characterized in that, The lifetime prediction module supports multi-indicator fusion evaluation and deploys a lifetime prediction model based on stochastic processes and a lifetime prediction model based on deep learning; the lifetime prediction model based on stochastic processes includes a linear Wiener process model, a nonlinear Wiener process model, and an exponential stochastic degradation model.

6. An embedded multimodal health management system, characterized in that, It includes the tested health management unit and control terminal as described in any one of claims 1-5; the control terminal includes a basic management module, a comprehensive data management module, a maintenance decision module, a visualization module, a human-computer interaction module, and a health management center unit; The basic management module is used for device information management, user permission management, and model parameter configuration. The integrated data management module is used to receive and integrate multi-source data uploaded by the tested health management unit, and provides data query, structured storage and backup functions; The maintenance decision module generates maintenance plans and recommendations based on fault diagnosis results and life prediction results. The visualization module displays data from the entire health management process in a graphical format; The human-computer interaction module provides an operation interface and supports interactive functions such as command issuance and parameter setting. The health management center unit also includes a data preprocessing module, a fault diagnosis module, a life prediction module, a data storage module, and a maintenance decision module, which are used to achieve global optimization and model training; the tested health management unit and the control terminal transmit control commands through a 1553B bus and transmit data through Ethernet.

7. A control method for an embedded multimodal device health management microsystem as described in any one of claims 1-5, characterized in that, Includes the following steps: S 1. The tested health management unit collects multimodal raw data through sensors or a bus; S2, the data preprocessing module performs real-time preprocessing operations on the multimodal raw data and outputs high-quality data; S3. The fault diagnosis module, based on the preprocessed data, realizes real-time fault detection and diagnosis through the fault detection model, outputs the fault diagnosis results and stores them in the data storage module, and uploads them to the control terminal at the same time. S4. The lifetime prediction module, based on the preprocessed data, performs performance degradation modeling and lifetime prediction through the remaining lifetime prediction model, outputs lifetime prediction results and stores them in the data storage module, and uploads them to the control terminal at the same time. S5. The control terminal receives and integrates the data uploaded by the tested health management unit, generates a maintenance strategy through the maintenance decision module, displays it through the visualization module, and supports user operation and command issuance through the human-computer interaction module. S6. The tested health management unit receives control commands from the control terminal, executes the corresponding operations, and feeds back the execution results, forming a closed-loop management system.

8. The embedded multimodal health management and control method according to claim 7, characterized in that, In step S2, the real-time preprocessing operation includes outlier removal, data normalization and feature extraction, and data denoising. The outlier removal method employs the standard deviation algorithm, box plot method, or Hampel function method. The data normalization and feature extraction employ empirical mode decomposition and principal component analysis, combined with normalization processing; The data denoising employs a moving average algorithm or singular value decomposition filtering.

9. The embedded multimodal health management and control method according to claim 7, characterized in that, In step S3, the specific process of fault diagnosis includes: S31. Perform dynamic characteristic discrimination on the preprocessed monitoring data; S32. Fault detection is performed using a fault detection model based on local information increment and slow feature analysis. The model is trained offline using training data and then used for online detection using test data. S33. Calculate the mean of local information increments as a statistic, and construct a dynamic threshold based on the S-shaped membership function; S34. Determine whether a sample is a faulty sample based on the comparison results of the statistics and the dynamic threshold, and output the fault diagnosis results.

10. The embedded multimodal health management and control method according to claim 7, characterized in that, In step S4, the specific process for lifetime prediction includes: S41. Degradation trend determination is performed on the preprocessed monitoring data; S42. If the monitoring data shows no significant degradation trend, a deep learning-based remaining life prediction model is used, and the prediction results are output after offline training and online testing. S43. If the monitoring data shows a significant degradation trend, construct four types of degradation models: linear, power, logarithmic, and exponential. Use a residual lifetime prediction model based on stochastic processes for prediction and combine the results of the deep learning model to obtain a fused prediction result. S44. Select the optimal model using the AIC criterion or mean square error, and output the final lifetime prediction result.

Images