Method and system for dynamic assessment of device health based on autoencoder

Abstract

The application discloses a device health degree dynamic evaluation method and system based on a self-encoder, relates to the technical field of health evaluation, and comprises the following steps: reading runtime sequence data; extracting data features after cleaning and screening; using the data features as training data to perform dynamic weight constraint learning on an LSTM self-encoder, and constructing a prediction model; receiving real-time data by using the model and calculating reconstruction error to establish a first score; configuring a sequence prediction branch, identifying gradual degradation by aligning short-term prediction error, and establishing a second score; identifying mutation points according to the reconstruction error, combining multi-scale sliding window analysis, and establishing a third score; and comprehensively evaluating the device health degree by using the three types of scores. The application solves the technical problems of high cost, poor adaptability and insufficient early warning capability caused by the dependence on invasive sensors and static rules in the prior art, and achieves the technical effects of improving the accuracy and foresight of device health evaluation.

Description

Technical Field

[0001] This invention relates to the field of health assessment technology, specifically to a method and system for dynamic assessment of device health based on an autoencoder. Background Technology

[0002] In modern industrial production, the stable and efficient operation of equipment is the core foundation for ensuring production safety and continuity. As industrial equipment becomes more complex and precise, the need for real-time monitoring and accurate assessment of equipment health status is becoming increasingly urgent. Traditional equipment health management methods mainly rely on physical sensor monitoring, expert rule threshold alarms, supervised learning prediction, and traditional unsupervised learning. However, these methods have significant limitations in practical applications: Physical sensor monitoring requires the installation of additional sensors on key components of the equipment, involving invasive installation and complex wiring, which not only results in high hardware and labor costs but also makes it difficult to modify a large number of existing operational equipment, and may even affect the original warranty and normal operation of the equipment; Expert rule threshold alarms rely on fixed thresholds and cannot adapt to the entire life cycle of the equipment (such as the break-in period and aging period) and the state changes under different operating conditions (high load, low load) and external environment (temperature and humidity fluctuations), which can easily lead to false alarms or missed alarms; Supervised learning prediction requires a large amount of labeled fault data to train the model, but the uptime of industrial equipment accounts for a very high proportion, and fault samples, especially early fault data, are scarce, making data acquisition extremely costly; Although traditional unsupervised learning does not require labels, it often treats time-series data as independent samples for analysis, ignoring the temporal dependencies of equipment operating parameters, and is insufficient in early warning capabilities for gradual faults (such as slow parameter degradation). Therefore, there is an urgent need for a health assessment method that does not rely on invasive sensors, fault tags, and static rules, and can dynamically integrate temporal characteristics and changes throughout the equipment's life cycle, in order to achieve accurate and forward-looking assessment of equipment status under complex operating conditions. Summary of the Invention

[0003] This application provides a method and system for dynamic evaluation of equipment health based on autoencoders. It addresses the technical problems of high cost, poor adaptability, and insufficient early warning capability caused by reliance on invasive sensors and static rules in existing technologies. The system achieves dynamic adaptive evaluation of equipment health status and early warning of gradual deterioration through dynamic weighting mechanism and multi-scale health scoring, thereby improving the accuracy and foresight of equipment health evaluation.

[0004] In view of the above problems, this application provides a method and system for dynamic evaluation of device health based on autoencoders.

[0005] The first aspect of this application provides a method for dynamic evaluation of device health based on an autoencoder. The method includes: reading historical operating data of the device over a predetermined period, wherein the historical operating data is runtime sequence data marked as normal operation; performing data cleaning and standard screening on the runtime sequence data, performing feature engineering on the retained data, and extracting data features, including basic physical features, temporal features, and physical derived features; using the data features as training data, performing convergence learning under dynamic weight constraints on an LSTM autoencoder network structure composed of an encoder and a decoder to construct a prediction model; using the prediction model to receive real-time operating data, performing reconstruction error calculation, and establishing a first health score; configuring a sequence prediction branch, using the sequence prediction branch to perform short-term prediction error alignment, identifying gradual degradation trends, and establishing a second health score; identifying abrupt change points based on the reconstruction error calculation, configuring multi-scale windows using the abrupt change points, performing multi-scale sliding window analysis of the real-time operating data, and establishing a third health score; and using the first health score, second health score, and third health score to perform dynamic evaluation of device health.

[0006] A second aspect of this application provides a dynamic device health assessment system based on an autoencoder. The system includes: a data reading module for reading historical operating data of the device over a predetermined period, wherein the historical operating data is runtime sequence data marked as normal operation; a feature extraction module for cleaning and standard filtering the runtime sequence data, performing feature engineering on the retained data, and extracting data features, including basic physical features, temporal features, and physical derived features; and a model learning module for using the data features as training data to perform convergence learning under dynamic weight constraints on an LSTM autoencoder network structure composed of an encoder and a decoder, thereby constructing a model. The system comprises: a prediction model; a first health assessment module, which receives real-time operational data using the prediction model, performs reconstruction error calculation, and establishes a first health score; a second health assessment module, which configures a sequence prediction branch, performs short-term prediction error alignment using the sequence prediction branch, identifies gradual degradation trends, and establishes a second health score; a third health assessment module, which identifies abrupt change points based on reconstruction error calculation, configures multi-scale windows using the abrupt change points, performs multi-scale sliding window analysis of real-time operational data, and establishes a third health score; and a dynamic health assessment module, which uses the first, second, and third health scores to dynamically assess the equipment's health.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0008] The system reads historical operating data of the equipment over a predetermined period, which is runtime sequence data marked as normal operation. The runtime sequence data undergoes data cleaning and standard filtering. Feature engineering is performed on the retained data to extract data features, including basic physical features, temporal features, and physical derived features. These data features are used as training data to perform convergence learning under dynamic weight constraints on an LSTM autoencoder network structure composed of an encoder and decoder, constructing a prediction model. The prediction model receives real-time operating data, performs reconstruction error calculation, and establishes a first health score. A sequence prediction branch is configured, and short-term prediction error alignment is performed using this branch to identify gradual degradation trends, establishing a second health score. Abrupt change points are identified based on the reconstruction error calculation. Multi-scale windows are configured using these abrupt change points, and multi-scale sliding window analysis of the real-time operating data is performed to establish a third health score. The first, second, and third health scores are used to dynamically assess the equipment's health. This achieves the technical effect of dynamically adaptively assessing the equipment's health status and providing early warning of gradual degradation through a dynamic weight mechanism and multi-scale health scores, improving the accuracy and foresight of equipment health assessment. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0010] Figure 1 This is a schematic diagram of the dynamic evaluation method for device health based on an autoencoder provided in an embodiment of this application.

[0011] Figure 2 This is a schematic diagram of the structure of a device health dynamic assessment system based on an autoencoder provided in an embodiment of this application.

[0012] Figure labeling: Data reading module 11, Feature extraction module 12, Model learning module 13, First health assessment module 14, Second health assessment module 15, Third health assessment module 16, Health dynamic assessment module 17. Detailed Implementation

[0013] This application provides a method and system for dynamic evaluation of equipment health based on autoencoders. It addresses the technical problems of high cost, poor adaptability, and insufficient early warning capabilities caused by reliance on invasive sensors and static rules in existing technologies. By using a dynamic weighting mechanism and multi-scale health scoring, it achieves dynamic adaptive evaluation of equipment health status and early warning of gradual deterioration, thereby improving the accuracy and foresight of equipment health evaluation.

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0015] Example 1, as Figure 1 As shown, this application provides a method for dynamic evaluation of device health based on an autoencoder, the method comprising:

[0016] Read the device's historical operating data for a predetermined period of time, wherein the historical operating data is runtime sequence data marked as normal operation.

[0017] In one embodiment, the device's operating records for a predetermined period (such as the most recent six months or one year) are first retrieved from the device's historical database or online data storage platform to form historical operating data. This historical operating data is time-series data that has been confirmed by maintenance personnel or the system to be in a normal state. That is, operating data collected under continuous operating conditions without faults, maintenance, or abnormal shutdowns, including but not limited to parameters such as current, voltage, power, temperature, humidity, pressure, and speed. It is stored sequentially according to the sampling frequency and indexed by timestamps. This runtime time-series data will be used for subsequent cleaning and feature extraction to help establish an accurate and reliable predictive model.

[0018] The runtime sequence data is cleaned and filtered according to standards. The retained data is then subjected to feature engineering to extract data features, which include basic physical features, time features, and physical derived features.

[0019] In one embodiment, after obtaining runtime sequence data for a predetermined period, to ensure data accuracy, the read runtime sequence data undergoes data cleaning and standard screening from three aspects: statistical outlier removal, fault record backtracking removal, and non-steady-state operation data removal. Statistical outlier removal eliminates sudden values ​​or data that do not conform to normal fluctuation ranges during normal equipment operation, ensuring the accuracy of subsequent data processing. Fault record backtracking removal eliminates data collected during fault periods, preventing data that does not conform to normal operating conditions from being input into the model. Non-steady-state operation data removal eliminates operating data during certain non-steady-state processes such as startup, shutdown, or load changes, improving the accuracy and generalization ability of model training. After data cleaning and standard screening, the remaining data undergoes feature engineering. This involves extracting key data features from the cleaned runtime sequence data according to predefined feature keys. These features include basic physical features, temporal features, and physical-derived features to reflect the equipment's health status. Basic physical features include core operating parameters such as current, power, temperature, and pressure. Temporal features include runtime and periodic variations. Physical-derived features include vibration frequency, fluctuation amplitude, power factor, temperature rise, and phase imbalance. These features further enrich the description of the equipment's operating status and serve as input data for subsequent model training, providing ample information for dynamically assessing equipment health.

[0020] In one possible implementation, the data cleaning and standard filtering of the runtime sequence data includes:

[0021] The runtime sequence data is divided into multiple local time windows, and the mean value of the data within each local time window is obtained. After creating a distribution constraint interval based on the mean value of the data, the distribution constraint interval is used to filter and remove data anomalies in the corresponding local time windows to complete the data cleaning.

[0022] Preferably, during the data cleaning process, the runtime sequence data is first divided into multiple local time windows according to a preset time window (e.g., hour, day). This preset time window is usually set by the periodic characteristics of device operation and actual business needs. By segmenting, the runtime sequence data can be broken down into smaller, relatively independent local time windows. Then, within each local time window, the mean of all data points within that window is calculated. These means represent the typical values ​​of the operating parameters within the corresponding local time window. Next, the standard deviation of each local time window is calculated, and the mean of each local time window is plus or minus the corresponding standard deviations (usually 2 or 3) to determine the distribution constraint interval for each local time window. Then, these distribution constraint intervals are applied to the corresponding local time windows to filter out outliers in the data contained within each local time window. Outliers that exceed the corresponding distribution constraint interval are removed to ensure that only true and valid normal data is retained in each local time window. This avoids the degradation of model training performance due to outlier interference, thereby providing a more accurate data foundation for subsequent feature engineering and model training.

[0023] In one possible implementation, after the data cleaning is completed, the method further includes:

[0024] Activate the backtracking exclusion mechanism, which is a mechanism that uses fault records as the starting point to perform adaptive backtracking data anomaly verification; use the backtracking exclusion mechanism to verify the probability of anomalies in runtime sequence data after data cleaning, and perform data removal management; obtain the device execution status and operation records, and perform non-standard operating condition data filtering on the runtime sequence data after data removal management based on the device execution status and operation records, so as to establish retained data.

[0025] Optionally, after the initial data cleaning is completed, to avoid mistakenly using data containing potential fault information as normal data, a backtracking exclusion mechanism is first activated. This mechanism adaptively determines the start time point for backtracking exclusion based on the equipment's fault records (such as maintenance work orders, fault reports, etc.), and uses this time point as a reference to execute a data anomaly verification mechanism. That is, data within a certain period before the start time point is marked as sub-healthy. This period is usually set to one month before the fault occurred, but can be adjusted according to the equipment's fault mode and usage cycle. Subsequently, the backtracking exclusion mechanism is used to verify the runtime data anomaly probability of these data marked as sub-healthy. That is, the deviation degree of the data marked as sub-healthy is analyzed with the sample normal data. If the deviation of a certain data point exceeds the normal range, the data point is recorded, and the number of all recorded data points is divided by the number of data points marked as sub-healthy to obtain the runtime data anomaly probability. If this runtime data anomaly probability exceeds the probability threshold, it indicates a high risk, and at this time, these data marked as sub-healthy will be removed. Furthermore, if any deviation exceeds the threshold, data with sub-health markers will be directly removed without further verification of runtime data anomaly probabilities. After data removal management, the data undergoes further filtering. This involves first acquiring the device's execution status and operation records, then reading the operation records to determine if specific operations have been initiated, such as maintenance, debugging, shutdown, trial operation, sudden failure, or upgrades. These specific operations cause the device to enter non-standard operating conditions, resulting in significantly different device execution states compared to normal operating states. By utilizing these execution statuses and operation records, all non-standard operating condition data can be identified and removed to avoid impacting model training and health assessment results. After removing fault data, abnormal data, and non-standard operating condition data, a retained dataset is generated. This retained dataset serves as the basis for subsequent feature engineering and model training, ensuring the data during training is authentic and reliable, avoiding training bias caused by abnormal data, and ensuring more accurate and credible final evaluation results.

[0026] Using the data features as training data, a prediction model is constructed by performing convergence learning under dynamic weight constraints on the LSTM autoencoder network structure composed of the encoder and decoder.

[0027] In one embodiment, after obtaining the data features, these features are used as training data and input into a deep learning model based on a Long Short-Term Memory (LSTM) autoencoder network structure. The LSTM autoencoder consists of an encoder and a decoder. The encoder is responsible for compressing and encoding the input data features, extracting latent patterns and temporal information from the data; the decoder is responsible for reconstructing the original input data from the encoded information, ensuring that the output is as similar to the input as possible. During training, the input data features are transformed into a more concise, low-dimensional representation by the encoder, and then reconstructed by the decoder. The model's goal is to minimize the reconstruction error by learning the key temporal information extracted from the original data. The reconstruction error refers to the difference between the decoder's output data and the original input data; the smaller the error, the more the model has learned the inherent patterns and regularities of the data. To enable the model to adapt to the characteristics and operating environments of different devices, dynamic weight constraints are used to optimize the LSTM autoencoder. Specifically, a device age adjustment function and a device environment adjustment function are introduced to dynamically adjust the weights of each feature. During training, the model will give different attention to different features based on these dynamic weights, thereby enhancing the model's sensitivity and accuracy to changes in device health status. During training, the LSTM autoencoder continuously adjusts the weight parameters within the network to progressively optimize the model's performance, enabling it to better learn the device's behavior patterns during normal operation and minimize reconstruction errors, thus achieving convergence. Once the model is trained, a predictive model is obtained. This predictive model effectively represents the device's normal operating mode and exhibits high accuracy in reconstructing unknown data, making it suitable for subsequent device health assessments and fault prediction.

[0028] In one possible implementation, the convergence learning of the LSTM autoencoder network structure composed of the encoder and decoder under dynamic weight constraints includes:

[0029] Obtain the device age adjustment function and set the operating condition characteristic age influence coefficient according to the device's working conditions; obtain the device environment adjustment function and configure the environment characteristic age influence coefficient; calculate the dynamic weights based on the device age adjustment function, operating condition characteristic age influence coefficient, device environment adjustment function, and environment characteristic age influence coefficient to complete convergence learning.

[0030] Preferably, during the training and learning of the LSTM autoencoder network structure, the device age adjustment function and the device environment adjustment function are first obtained. The device age adjustment function reflects the degradation effect of the device as it ages, and can be expressed as the ratio of the difference between the device's current equivalent age and its minimum designed service age to the difference between its maximum designed service age and its minimum designed service age. It is typically between 0 and 1, representing the degree of device aging. The current equivalent age is a comprehensive measure considering factors such as maintenance and repair, and is usually the physical age minus the age adjusted for maintenance effectiveness. This age adjusted for maintenance effectiveness can be mapped based on a maintenance effectiveness rule base. The device environment adjustment function considers the impact of the external environment on the device's health. The impact of the device's environmental adjustment function can be implemented according to the accuracy requirements of the application scenario. For applications with lower accuracy requirements, a discrete lookup table method can be used. For example, the seasons can be segmented to construct a seasonal rule table, such as {Summer: 1.2, Transitional Season: 1.0, Winter: 0.8}. When it is determined that the current season is summer, the value of the device's environmental adjustment function is 1.2. For applications with higher accuracy requirements, a continuous function method can be used. For example, the device's environmental adjustment function can be defined as the deviation from the comfort zone, i.e., max(0, (current ambient temperature - 25°C) / 20°C). When the ambient temperature is higher than the set benchmark value (such as 25°C), the value of the device's environmental adjustment function increases accordingly, thereby more sensitively capturing anomalies caused by heat dissipation problems in high-temperature environments. In addition, operating condition characteristic age influence coefficients and environmental characteristic age influence coefficients are set according to the equipment's operating conditions. The operating condition characteristic age influence coefficient reflects the impact of usage intensity under different operating conditions on equipment aging, while the environmental characteristic age influence coefficient represents the degree of influence of environmental factors on equipment aging. These two influence coefficients are automatically learned using hyperparameter search algorithms such as Bayesian optimization, with the optimization objective of ensuring that the average reconstruction error distribution of equipment of different ages is as consistent as possible. Subsequently, the basic weights of each feature are obtained, and the dynamic weights of each feature are calculated using the formula: Dynamic Weight = Basic Weight × (1 + Operating Condition Characteristic Age Influence Coefficient × Equipment Age Adjustment Function + Environmental Characteristic Age Influence Coefficient × Equipment Environment Adjustment Function) to reflect the importance of each feature under different operating conditions, environments, and equipment aging stages. During model training, dynamic weights are used to influence the learning process of each feature. This allows the LSTM autoencoder network to converge more flexibly to the optimal model parameters. Each time weight constraints are applied, the network continuously adjusts its learning strategy based on changes in the equipment's age, operating conditions, and environment to minimize model loss.Ultimately, through this dynamically adjusted learning process, a final predictive model can be constructed. The model file, dynamic weight configuration file, preprocessor object, and other components corresponding to this predictive model are then packaged and stored in the model library for inference. This allows for accurate prediction of device health and identification of potential failure risks and health change trends.

[0031] Table 1: Example Table of Repair Effectiveness Calculation Rule Base

[0032]

[0033] Table 1 above is an example table of the maintenance effect conversion rule library. This table shows the maintenance items of different maintenance levels and their corresponding conversion years, providing a reference for maintenance effect conversion.

[0034] The prediction model is used to receive real-time operating data, perform reconstruction error calculation, and establish a first health score.

[0035] In one embodiment, real-time operating data of the device, including current, voltage, temperature, pressure, and rotational speed, is acquired through a reserved interface. This real-time operating data is then input into a predictive model. The predictive model reconstructs the input data using an encoder, that is, based on its learned normal operating mode, it regenerates an output that is as consistent as possible with the input data. A decoder then generates a predictive output. The reconstruction error is calculated by comparing the error between the input and output data feature by feature using mean squared error (MSE) or other error metrics. The magnitude of this reconstruction error reflects the degree of deviation between the device's current operating state and its normal operating mode. If the real-time data matches the device's normal mode, the error will be very small, indicating good device health. Conversely, a large error indicates that the device may have an anomaly or poor health. Subsequently, based on the calculated reconstruction error, these errors are fused using a weighted method, and then a mapping function is used to convert the reconstruction error into a first health score from 0 to 100. The higher the first health score, the closer the device's operating state is to normal and the better its health. The lower the score, the more likely the device has a fault, deterioration, or other health problems. In this way, the operating status of the equipment can be monitored in real time, and the health of the equipment can be assessed by reconstructing errors, so as to identify potential problems in a timely manner and provide a basis for subsequent maintenance and repair decisions.

[0036] In one possible implementation, the step of using the prediction model to receive real-time operational data, perform reconstruction error calculation, and establish a first health score includes:

[0037] The system receives data packets uploaded by the device in real time via the MQTT protocol. These data packets contain real-time running data. The system then calls the preprocessor in the prediction model to perform data preprocessing on the real-time running data in the data packets and loads the prediction model and configuration file. After inputting the preprocessed real-time running data into the calculation unit of the prediction model, the system obtains reconstructed data. The system calculates the feature-by-feature reconstruction error between the reconstructed data and the original data. After calculating the equivalent age of the device and environmental conditions, and configuring dynamic feature weights, the system performs a weighted calculation of the reconstruction error and outputs a first health score through a preset mapping function.

[0038] Optionally, the device periodically collects operational data via installed sensors and packages it into data packets, which are then uploaded via the MQTT protocol. These data packets include various real-time operational data of the device. Upon receiving the data packets, the preprocessor in the prediction model is invoked. This preprocessor cleans and standardizes the real-time operational data to ensure data quality and consistency. The preprocessing steps are the same as described above, thereby transforming the received real-time operational data into real-time data features and removing any abnormal or invalid data that may affect the model's accuracy. Subsequently, the pre-trained prediction model and related configuration files are loaded from the model library, and the pre-processed real-time data features are input into the prediction model's computation unit. In this computation unit, the prediction model compresses the data into low-dimensional feature representations using an encoder, and then attempts to reconstruct the original data based on the learned normal operating mode using a decoder. This results in reconstructed data that matches the input data as closely as possible, representing the device's expected performance under normal conditions. Afterward, the per-feature error between the reconstructed data and the real-time data features is calculated using the mean squared error. Then, based on the equivalent age and environmental conditions, the dynamic weight of each feature is calculated using the device age adjustment function, the device environment adjustment function, the operating condition feature age influence coefficient, and the environmental feature age influence coefficient. Then, based on the dynamic weights of these features, the feature-by-feature errors are weighted to obtain a comprehensive reconstruction error value. Finally, using a preset mapping function (such as 100 minus the ratio of error to threshold), the weighted comprehensive reconstruction error value is converted into a health score between 0 and 100. The higher the score, the better the equipment's health status and the closer the equipment is to normal operation. A lower score indicates that the equipment may have a fault or health problem and requires further attention. Through this series of processes, the health status of the equipment can be monitored in real time, abnormal conditions can be detected in a timely manner, and corresponding health assessments can be provided, providing strong support for subsequent maintenance decisions and fault early warnings.

[0039] Configure a sequence prediction branch, use the sequence prediction branch to align short-term prediction errors, identify gradual deterioration trends, and establish a second health score.

[0040] In one embodiment, the configuration layer of the sequence prediction branch is configured based on the window length and fluctuation complexity characteristics of the real-time operating data. Then, based on the configured sequence prediction branch, short-term predictions are made on the real-time operating data. The predicted results are then aligned with the real-time operating data to calculate the short-term prediction error, which reflects the gap between the actual operating state of the equipment and the predicted result. Next, based on the calculated error, a gradual degradation trend is identified, and a second health score is calculated. This second health score provides a quantitative indicator for the short-term health assessment of the equipment. A low second health score indicates a significant gradual degradation trend and poor health, while a high second health score indicates that the equipment has not experienced significant degradation problems in the short term and is operating well. In this way, not only can the status of the equipment be monitored in real time, but the gradual deterioration trend of the equipment can also be effectively identified, providing early warnings to maintenance personnel to help them intervene in a timely manner.

[0041] In one possible implementation, the step of using the sequence prediction branch for short-term prediction error alignment, identifying gradual deterioration trends, and establishing a second health score includes:

[0042] The real-time running data is used as input data and fed into the sequence prediction branch; the window length feature of the real-time running data is obtained, and a first constraint is established based on the window length feature; the fluctuation complexity feature of the real-time running data is obtained, and a second constraint is established based on the fluctuation complexity feature; after receiving the first constraint and the second constraint using the configuration layer of the sequence prediction branch, the prediction window of the sequence prediction branch is configured; after performing short-term prediction based on the prediction window, the short-term prediction results are aligned with the real-time running data to construct an analysis dataset; a prediction trend score is performed on the analysis dataset to identify gradual deterioration trends and establish a second health score.

[0043] Optionally, after obtaining real-time operational data, this data is input into the sequence prediction branch (constructed using an LSTM). In this branch, the window length feature of the real-time operational data is first extracted—that is, the size of each data window in the time series—to determine a suitable time scale for prediction. Based on this time scale, a first constraint is set to limit the amount of data used in prediction, thereby improving the accuracy of the prediction results. Subsequently, the volatility complexity of the real-time operational data is analyzed. Volatility complexity refers to the degree of fluctuation in feature data over time during equipment operation; complex fluctuations may represent changes in equipment status or potential performance degradation. By calculating the frequency (obtained using a Fast Fourier Transform) and amplitude (standard deviation) of the fluctuations, volatility complexity features can be derived, and a second constraint is established to limit the fluctuation range of the equipment operational data, thus affecting the learning ability and accuracy of the prediction model. Afterward, the first and second constraints are passed to the configuration layer of the sequence prediction branch. In the configuration layer, the size and attributes of the prediction window are configured according to the constraints. The prediction window determines the prediction range for future time periods and affects the accuracy of the prediction results. After configuring the prediction window, the sequence prediction branch performs short-term predictions based on real-time operational data. During this process, the sequence prediction branch predicts the future state of the device based on learned knowledge and real-time operational data, obtaining short-term prediction results. These short-term prediction results are aligned with the real-time operational data to form an analysis dataset for verifying the model's prediction accuracy. Then, based on the analysis dataset, the prediction error is calculated using the absolute difference, and the mean of the prediction error is input into the second health score formula to score the prediction trend, identifying gradual deterioration trends. This second health score formula is as follows:

[0044] ;

[0045] Where HD2 is the second health score; γ is the weighting coefficient used to adjust the sensitivity of the health score; σ(•) is the Sigmoid function, which ensures that the output health score is between 0 and 100; It is the mean of the prediction error; and These are the mean and standard deviation under normal operating conditions, respectively. Using the second health score formula, the prediction error can be mapped to a second health score. This second health score reflects the equipment's health status in the short term. If the equipment's prediction error is small and shows no obvious trend, the health score is high; conversely, if the error gradually increases, it indicates a possible trend of equipment health decline, and the score is low. Through this process, the second health score helps the system accurately identify gradual equipment degradation and provides strong support for subsequent maintenance and repair decisions.

[0046] Based on the reconstruction error calculation, abrupt change points are identified. After configuring multi-scale windows using the abrupt change points, multi-scale sliding window analysis of real-time running data is performed to establish a third health score.

[0047] In one embodiment, based on the calculated reconstruction errors, a statistical change point detection algorithm is used to analyze these errors and identify abrupt changes. These abrupt changes typically represent drastic fluctuations in the equipment's operating status, potentially indicating a malfunction or anomaly. Subsequently, based on the identified abrupt changes, multi-scale windows are configured, and a sequence prediction branch is used to perform short-term predictions for each scale window. The prediction results are then aligned to obtain the analysis dataset for each scale window. The obtained analysis datasets are then sequentially input into the third health score formula to calculate the third health score. This third health score comprehensively considers the equipment's health status at different time scales, reflecting the overall health of the equipment. A higher third health score indicates a more stable health status across all time scales and a lower risk of abrupt changes, while a lower score indicates a greater risk of fluctuations or sudden failures, requiring timely intervention and maintenance. Through this process, the third health score can accurately reflect whether the equipment has experienced a sudden failure in the short term, while simultaneously assessing the equipment's health stability over a longer period, providing strong data support for equipment health management.

[0048] In one possible implementation, the step of identifying abrupt change points based on reconstruction errors, configuring multi-scale windows using these abrupt change points, performing multi-scale sliding window analysis on real-time running data, and establishing a third health score includes:

[0049] Obtain the reconstruction error sequence in the reconstruction error calculation; perform abrupt detection of the reconstruction error sequence based on the statistical change point detection algorithm to establish a set of abrupt change points; calculate the intensity feature of each abrupt change point in the set of abrupt change points, the intensity feature being obtained by pre-estimating the error change amplitude, direction, and duration; use the intensity feature as a configuration constraint for a multi-scale window to complete the multi-scale window configuration.

[0050] Preferably, all calculated reconstruction errors are first acquired and stored in chronological order to form a reconstruction error sequence. Then, a statistical change point detection algorithm is used to analyze the reconstruction error sequence. During this process, the mean and standard deviation of the reconstruction errors are calculated, and a threshold interval is constructed by adding or subtracting multiple standard deviations from the mean. This threshold interval allows for the selection of abrupt change points from the reconstruction error sequence, creating a set of abrupt change points. Subsequently, for each abrupt change point in the set, the error change amplitude, direction, and duration are calculated. The error change amplitude is the magnitude of the error change before and after the abrupt change point; the direction indicates whether the abrupt change point jumps upwards or downwards, i.e., whether it suddenly increases or decreases; and the duration is the length of time the abrupt change lasts. Short-term error fluctuations may indicate instantaneous shocks, while long-term deviations may indicate gradual equipment degradation or systemic failures. By normalizing the error change amplitude, direction, and duration, and then weighting the results, the intensity characteristics of each abrupt change point can be quantified, providing important reference for subsequent multi-scale window configuration. Subsequently, based on the intensity characteristics of the abrupt change points, multiple windows of different scales are configured for real-time data analysis. Specifically, the intensity characteristics are compared with intensity ranges. If the intensity characteristic is large, it indicates that the equipment has experienced a severe instantaneous impact. In this case, a short-scale window is configured to improve the system's response speed, and its weight is increased to give greater attention to rapidly changing anomalies. If the intensity characteristic is small but the duration is long, it indicates that the equipment's state has changed slowly. In this case, a medium-scale window is configured to accommodate this slow change trend, and its weight is adjusted accordingly to highlight this long-term trend of continuous deviation. If the intensity characteristic is small and the duration is even longer, it indicates that the equipment's health status has deviated from the normal range for a long period. In this case, a long-scale window is configured to capture the long-term trend of the equipment's changes. Through this series of steps, the health status of the equipment can be dynamically and accurately assessed, and sudden failures and gradual degradation can be detected in a timely manner, providing strong support for subsequent maintenance and early warning.

[0051] In one possible implementation, the multi-scale sliding window analysis of real-time running data to establish a third health score includes:

[0052] The third health score is calculated using the following formula:

[0053] ;in, Characterized by the third health score, This is the global scaling factor. Characterization scale sliding window, These respectively represent the short-scale window set, the medium-scale window set, and the long-scale window set. Characterizing the sliding window scale is Scale weights dynamically determined by the characteristics of mutation points. The smoothing mapping function is used to map deviation values ​​to the interval between 0 and 1. Characterizing the sliding window scale is The deviation of the sliding window statistics at that time The sliding window scale, representing the normal scale, is The mean of the statistical measure at time, The sliding window scale, representing the normal scale, is The standard deviation of the statistical measure over time.

[0054] Optionally, after obtaining multiple analysis datasets using the sequence prediction branch and multi-scale sliding window, the mean prediction error of each analysis dataset is calculated. The calculated mean is then input into the third health score formula to calculate the third health score. The specific third health score formula is as follows:

[0055] ;

[0056] in, Characterized by the third health score, This is the global scaling factor. Characterization scale sliding window, These respectively represent the short-scale window set, the medium-scale window set, and the long-scale window set. Characterizing the sliding window scale is Scale weights dynamically determined by the characteristics of mutation points. The smoothing mapping function is used to map deviation values ​​to the interval between 0 and 1. Characterizing the sliding window scale is The deviation of the sliding window statistics at that time The sliding window scale, representing the normal scale, is The mean of the statistical measure at time, The sliding window scale, representing the normal scale, is The standard deviation of the statistical measure over time.

[0057] The equipment health is dynamically assessed using the first health score, the second health score, and the third health score.

[0058] In one embodiment, after obtaining the first health score, the second health score, and the third health score, the three scores are weighted and fused together, and then the early warning management is carried out based on the fusion result. This makes it easier for monitoring personnel to discover potential problems with the equipment in a timely manner, so as to carry out corresponding maintenance or optimization measures, avoid equipment failure, or extend the service life of the equipment.

[0059] In one possible implementation, the dynamic assessment of device health using the first health score, the second health score, and the third health score includes:

[0060] The first health score, the second health score, and the third health score are weighted and fused to establish a comprehensive health score; the comprehensive health score is used to identify and issue health warnings, generate health warning signals, and implement warning management.

[0061] Preferably, after obtaining the first, second, and third health scores, these scores are merged according to the preset weights of each health score to obtain a comprehensive health score. Subsequently, the comprehensive health score is subject to tiered threshold judgments. These tiered thresholds include normal threshold ranges, warning threshold ranges, and alarm threshold ranges. When the comprehensive health score falls into the warning threshold range, a health warning signal is generated, and a yellow alert is triggered on the management platform. When the score falls into the alarm threshold range and continues to exceed the set window length, a red health alarm signal is generated, and an alarm management process is triggered, such as automatically pushing maintenance work orders, linking the control system to reduce load, or performing emergency shutdown protection, thereby avoiding major equipment failures and improving equipment availability and lifespan.

[0062] Example 2, based on the same inventive concept as the self-encoder-based dynamic device health assessment method in the foregoing examples, such as... Figure 2 As shown, this application provides a dynamic device health assessment system based on an autoencoder. The system and method embodiments in this application are based on the same inventive concept. The system includes: a data reading module 11: reading historical operating data of the device over a predetermined period, wherein the historical operating data is runtime sequence data marked as normal operation; a feature extraction module 12: performing data cleaning and standard screening on the runtime sequence data, performing feature engineering on the retained data, and extracting data features, wherein the data features include basic physical features, time features, and physical derived features; a model learning module 13: using the data features as training data, performing convergence learning under dynamic weight constraints on the LSTM autoencoder network structure composed of an encoder and decoder to construct a prediction model; and a first health assessment module 14. The system employs the following modules: 1) Receive real-time operational data using the prediction model, calculate reconstruction error, and establish a first health score; 2) Configure a sequence prediction branch, use the sequence prediction branch to align short-term prediction errors, identify gradual degradation trends, and establish a second health score; 3) Identify abrupt change points based on reconstruction error calculations, configure multi-scale windows using the abrupt change points, perform multi-scale sliding window analysis on real-time operational data, and establish a third health score; 4) Perform dynamic health assessment using the first, second, and third health scores.

[0063] Furthermore, the feature extraction module 12 also includes:

[0064] The runtime sequence data is divided into multiple local time windows, and the mean value of the data within each local time window is obtained. After creating a distribution constraint interval based on the mean value of the data, the distribution constraint interval is used to filter and remove data anomalies in the corresponding local time windows to complete the data cleaning.

[0065] Furthermore, the feature extraction module 12 also includes:

[0066] Activate the backtracking exclusion mechanism, which is a mechanism that uses fault records as the starting point to perform adaptive backtracking data anomaly verification; use the backtracking exclusion mechanism to verify the probability of anomalies in runtime sequence data after data cleaning, and perform data removal management; obtain the device execution status and operation records, and perform non-standard operating condition data filtering on the runtime sequence data after data removal management based on the device execution status and operation records, so as to establish retained data.

[0067] Furthermore, the model learning module 13 also includes:

[0068] Obtain the device age adjustment function and set the operating condition characteristic age influence coefficient according to the device's working conditions; obtain the device environment adjustment function and configure the environment characteristic age influence coefficient; calculate the dynamic weights based on the device age adjustment function, operating condition characteristic age influence coefficient, device environment adjustment function, and environment characteristic age influence coefficient to complete convergence learning.

[0069] Furthermore, the first health assessment module 14 also includes:

[0070] The system receives data packets uploaded by the device in real time via the MQTT protocol. These data packets contain real-time running data. The system then calls the preprocessor in the prediction model to perform data preprocessing on the real-time running data in the data packets and loads the prediction model and configuration file. After inputting the preprocessed real-time running data into the calculation unit of the prediction model, the system obtains reconstructed data. The system calculates the feature-by-feature reconstruction error between the reconstructed data and the original data. After calculating the equivalent age of the device and environmental conditions, and configuring dynamic feature weights, the system performs a weighted calculation of the reconstruction error and outputs a first health score through a preset mapping function.

[0071] Furthermore, the second health assessment module 15 also includes:

[0072] The real-time running data is used as input data and fed into the sequence prediction branch; the window length feature of the real-time running data is obtained, and a first constraint is established based on the window length feature; the fluctuation complexity feature of the real-time running data is obtained, and a second constraint is established based on the fluctuation complexity feature; after receiving the first constraint and the second constraint using the configuration layer of the sequence prediction branch, the prediction window of the sequence prediction branch is configured; after performing short-term prediction based on the prediction window, the short-term prediction results are aligned with the real-time running data to construct an analysis dataset; a prediction trend score is performed on the analysis dataset to identify gradual deterioration trends and establish a second health score.

[0073] Furthermore, the third health assessment module 16 also includes:

[0074] Obtain the reconstruction error sequence in the reconstruction error calculation; perform abrupt detection of the reconstruction error sequence based on the statistical change point detection algorithm to establish a set of abrupt change points; calculate the intensity feature of each abrupt change point in the set of abrupt change points, the intensity feature being obtained by pre-estimating the error change amplitude, direction, and duration; use the intensity feature as a configuration constraint for a multi-scale window to complete the multi-scale window configuration.

[0075] Furthermore, the third health assessment module 16 also includes:

[0076] The third health score is calculated using the following formula: ;in, Characterized by the third health score, This is the global scaling factor. Characterization scale sliding window, These respectively represent the short-scale window set, the medium-scale window set, and the long-scale window set. Characterizing the sliding window scale is Scale weights dynamically determined by the characteristics of mutation points. The smoothing mapping function is used to map deviation values ​​to the interval between 0 and 1. Characterizing the sliding window scale is The deviation of the sliding window statistics at that time The sliding window scale, representing the normal scale, is The mean of the statistics at that time. The sliding window scale, representing the normal scale, is The standard deviation of the statistical measure over time.

[0077] Furthermore, the health dynamic assessment module 17 also includes:

[0078] The first health score, the second health score, and the third health score are weighted and fused to establish a comprehensive health score; the comprehensive health score is used to identify and issue health warnings, generate health warning signals, and implement warning management.

[0079] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0080] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0081] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A method for dynamic evaluation of device health based on a self-encoder, characterized in that, The method includes: Read the device's historical operating data for a predetermined period of time, wherein the historical operating data is runtime sequence data marked as normal operation; The runtime sequence data is cleaned and filtered according to standards. The retained data is then subjected to feature engineering to extract data features, which include basic physical features, time features, and physical derivative features. Using the data features as training data, a prediction model is constructed by performing convergence learning under dynamic weight constraints on the LSTM autoencoder network structure composed of encoder and decoder. The prediction model is used to receive real-time operational data, perform reconstruction error calculation, and establish a first health score. Configure a sequence prediction branch, use the sequence prediction branch to align short-term prediction errors, identify gradual deterioration trends, and establish a second health score. Based on the reconstruction error calculation, abrupt change points are identified. After configuring multi-scale windows using the abrupt change points, multi-scale sliding window analysis of real-time running data is performed to establish a third health score. The first health score, the second health score, and the third health score are used to dynamically assess the health of the equipment. The process of using the sequence prediction branch to align short-term prediction errors, identify gradual deterioration trends, and establish a second health score includes: The real-time running data is used as input data and input into the sequence prediction branch; Obtain the window length feature of the real-time running data, and establish a first constraint based on the window length feature; Obtain the fluctuation complexity characteristics of the real-time running data, and establish a second constraint based on the fluctuation complexity characteristics; After receiving the first constraint and the second constraint using the configuration layer of the sequence prediction branch, the prediction window of the sequence prediction branch is configured. After performing short-term predictions based on the prediction window, the short-term prediction results are aligned with real-time running data to construct an analysis dataset. A predictive trend score is performed on the analyzed dataset to identify gradual deterioration trends and establish a second health score. The process involves calculating and identifying mutation points based on reconstruction errors, configuring multi-scale windows using these mutation points, performing multi-scale sliding window analysis on real-time running data, and establishing a third health score, including: Obtain the reconstruction error sequence in the reconstruction error calculation; A mutation detection algorithm based on statistical change point detection is used to reconstruct the error sequence and establish a set of mutation points. Calculate the intensity characteristics of each mutation point in the mutation point set. The intensity characteristics are obtained by pre-estimating the error change amplitude, direction, and duration. The intensity features are used as configuration constraints for the multi-scale window to complete the multi-scale window configuration.

2. The method for dynamic evaluation of device health based on an autoencoder as described in claim 1, characterized in that, The multi-scale sliding window analysis of the real-time running data is used to establish a third health score, including: The third health score is calculated using the following formula: ； in, Characterized by the third health score, This is the global scaling factor. Characterization scale sliding window, These respectively represent the short-scale window set, the medium-scale window set, and the long-scale window set. Characterizing the sliding window scale is Scale weights dynamically determined by the characteristics of mutation points. The smoothing mapping function is used to map deviation values ​​to the interval between 0 and 1. Characterizing the sliding window scale is The deviation of the sliding window statistics at that time The sliding window scale, representing the normal scale, is The mean of the statistics at that time. The sliding window scale, representing the normal scale, is The standard deviation of the statistical measure over time.

3. The method for dynamic evaluation of device health based on an autoencoder as described in claim 1, characterized in that, The data cleaning and standard filtering of the runtime sequence data includes: The runtime sequence data is divided into multiple local time windows, and the average value of the data within each local time window is obtained; After creating a distribution constraint interval based on the data mean, the distribution constraint interval is used to filter and remove data anomalies in the corresponding local time window to complete the data cleaning.

4. The method for dynamic evaluation of device health based on an autoencoder as described in claim 3, characterized in that, After the data cleaning is completed, the process also includes: Activate the backtracking exclusion mechanism, which is a mechanism that uses the fault record as the starting point of time to perform adaptive backtracking data anomaly verification; The backtracking and exclusion mechanism is used to verify the probability of anomalies in runtime sequence data after data cleaning, and data removal management is performed. The device execution status and operation records are obtained. Based on the device execution status and operation records, the runtime sequence data after data removal management is filtered for non-standard operating conditions in order to establish retained data.

5. The method for dynamic evaluation of device health based on an autoencoder as described in claim 1, characterized in that, The convergence learning of the LSTM autoencoder network structure composed of the encoder and decoder under dynamic weight constraints includes: Obtain the equipment age adjustment function and set the operating condition characteristic age influence coefficient according to the equipment's working conditions; Obtain the device's environment adjustment function and configure the environmental characteristic age influence coefficient; Dynamic weights are calculated based on the equipment age adjustment function, the operating condition characteristic age influence coefficient, the equipment environment adjustment function, and the environmental characteristic age influence coefficient to complete convergent learning.

6. The method for dynamic evaluation of device health based on an autoencoder as described in claim 1, characterized in that, The process of receiving real-time operational data using the prediction model, performing reconstruction error calculation, and establishing a first health score includes: Data packets uploaded by the device are received in real time via the MQTT protocol, wherein the data packets contain real-time running data; The preprocessor in the prediction model is invoked to perform data preprocessing on the real-time running data in the data packet, and the prediction model and configuration file are loaded. After the preprocessed real-time running data is input into the calculation unit of the prediction model, the reconstructed data is obtained; The feature-by-feature reconstruction error between the reconstructed data and the original data is calculated. After configuring the equivalent age and environmental conditions of the computing device and the dynamic weights of the features, the weighted calculation of the reconstruction error is performed, and the first health score is output through the preset mapping function.

7. The method for dynamic evaluation of device health based on an autoencoder as described in claim 1, characterized in that, The dynamic assessment of equipment health using the first health score, the second health score, and the third health score includes: A comprehensive health score is established by weighting and fusing the first, second, and third health scores. The comprehensive health score is used to identify and issue health warnings, generate health warning signals, and implement warning management.

8. A dynamic health assessment system for equipment based on a self-encoder, characterized in that, The system is used to execute the device health dynamic assessment method based on an autoencoder as described in any one of claims 1 to 7, the system comprising: Data reading module: Reads historical operating data of the device for a predetermined period of time, wherein the historical operating data is runtime sequence data marked as normal operation; Feature extraction module: performs data cleaning and standard screening on the runtime sequence data, performs feature engineering on the retained data, and extracts data features, including basic physical features, time features, and physical derived features; Model learning module: Using the data features as training data, the module performs convergence learning under dynamic weight constraints on the LSTM autoencoder network structure composed of encoder and decoder to build a prediction model. First health assessment module: Utilizes the prediction model to receive real-time operational data, performs reconstruction error calculation, and establishes a first health score; Second health assessment module: Configure sequence prediction branch, use the sequence prediction branch to perform short-term prediction error alignment, identify gradual deterioration trend, and establish a second health score; The third health assessment module: Based on the reconstruction error, it identifies mutation points, configures multi-scale windows using the mutation points, performs multi-scale sliding window analysis on real-time running data, and establishes a third health score. Health dynamic assessment module: Uses the first health score, the second health score, and the third health score to perform dynamic assessment of equipment health.

Images