An electromechanical equipment health evaluation method based on unsupervised baseline learning and working condition decoupling and related equipment

Abstract

This application provides a method and related equipment for health assessment of electromechanical equipment based on unsupervised baseline learning and operating condition decoupling, belonging to the technical field of electromechanical equipment health assessment. The method includes: acquiring unlabeled historical operating data of electromechanical equipment under its health state, constructing a health benchmark embedding space and statistical baseline; extracting information features from standardized multivariate time series data blocks to obtain physical features characterizing the equipment state; calculating reconstruction error and representation distance to obtain deep features; fusing physical features and deep features to obtain an interpretable degradation feature vector; inputting the interpretable degradation feature vector and operating condition label data into an attention-based operating condition decoupling network to obtain operating condition-invariant degradation features; inputting the operating condition-invariant degradation features into a regressor to calculate a health score, and obtaining an electromechanical equipment health early warning report based on the health score. This application can achieve accurate assessment of the health state of electromechanical equipment.

Description

Technical Field

[0001] This application relates to the field of electromechanical equipment health assessment technology, and in particular to an electromechanical equipment health assessment method and related equipment based on unsupervised baseline learning and operating condition decoupling. Background Technology

[0002] In related technologies, traditional equipment maintenance strategies, such as "reactive maintenance" (repairing after a failure occurs) and "periodic maintenance" (maintenance at fixed intervals), are no longer sufficient to meet the requirements of modern civil and industrial buildings for high reliability, high availability, and low operating costs. Reactive maintenance leads to unplanned downtime, causing huge production losses; while periodic maintenance may result in "over-maintenance" (unnecessary dismantling and repair of equipment in good condition) or "under-maintenance" (failures occurring during maintenance intervals), leading to resource waste or unexpected downtime.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0004] The main objective of this application is to propose a method and related equipment for assessing the health status of electromechanical equipment based on unsupervised baseline learning and decoupling of operating conditions, which can achieve accurate assessment of the health status of electromechanical equipment.

[0005] To achieve the above objectives, one aspect of this application proposes a method for health assessment of electromechanical equipment based on unsupervised baseline learning and decoupling of operating conditions. The method includes the following steps: The multi-source sensor data and operating condition label data at the current moment are collected and preprocessed to form a standardized multivariate time series data block; Acquire unlabeled historical operating data of electromechanical equipment in its health status, train a multivariate time series representation learning network, construct a health baseline embedding space and statistical baseline, and store them in a health baseline library. Information features are extracted from the standardized multivariate time series data blocks to obtain physical features characterizing the device state; The standardized multivariate time series data block is input into the trained multivariate time series representation learning network, and the reconstruction error and representation distance are calculated using the health baseline library to obtain deep features; The physical features and the deep features are fused together, and a feature selection algorithm is used to select the feature subset with the highest correlation to degradation, so as to obtain an interpretable degradation feature vector. The interpretable degradation feature vector and the working condition label data are input into an attention-based working condition decoupling network to obtain working condition-invariant degradation features. The degradation characteristics under unchanged operating conditions are input into the regressor to calculate the health score, and a health warning report for electromechanical equipment is obtained based on the health score.

[0006] In some embodiments, the process of collecting and preprocessing multi-source sensor data and operational condition label data at the current moment to form a standardized multivariate time series data block includes: Data is collected synchronously from vibration sensors, temperature sensors, current sensors, voltage sensors, and equipment controllers to obtain multi-source sensor data at the current moment. Extract operating condition tag data from the manufacturing execution system or building equipment monitoring system; the operating condition tag data includes speed setpoint, load current, and operating mode code; The multi-source sensor data and the working condition tag data with different sampling frequencies are synchronized using timestamp alignment and resampling techniques to form data slices; Outliers are removed from the data slices using statistical methods or methods based on neighboring data, and missing values ​​are filled by linear interpolation or forward imputation to obtain standardized multivariate time series data blocks.

[0007] In some embodiments, the step of acquiring unlabeled historical operating data of the health status of electromechanical equipment, training a multivariate time series representation learning network, constructing a health baseline embedding space and statistical baseline, and storing it in a health baseline library includes: Acquire untagged historical operating data of electromechanical equipment during the initial stable operation period after installation and commissioning, and which are determined to be in a healthy state by preset rules; Using the unlabeled historical running data as training samples, a multivariate temporal representation learning network is trained; the multivariate temporal representation learning network is a stacked sparse autoencoder or a temporal contrastive learning network. The low-dimensional vector output by the multivariate temporal representation learning network is used as the health baseline embedding space, and the mean and variance upper and lower bounds of each channel of the health data are calculated as statistical baselines. The trained network model parameters and statistical baselines are stored together in the health baseline library. The health baseline embedding space is used as a low-dimensional benchmark reference system to represent the data distribution pattern of the device in a healthy state, and is used to quantify the device performance degradation by calculating the degree of deviation between the current data and the health baseline embedding space.

[0008] In some embodiments, the step of extracting information features from the standardized multivariate time series data block to obtain physical features characterizing the device state includes: The root mean square value, peak value, kurtosis, waveform factor, and impulse factor of the signal are calculated from the standardized multivariate time series data block to obtain the time domain characteristics; The signals in the standardized multivariate time series data block are subjected to Fast Fourier Transform to calculate the specific fault characteristic frequencies and their harmonic energy proportions that meet the preset conditions, thereby obtaining frequency domain characteristics; the specific fault characteristic frequencies include the bearing's passing frequency and the gear's meshing frequency. The wavelet packet energy spectrum is calculated for the non-stationary signal in the standardized multivariate time series data block, and the changes in energy distribution in different frequency bands are analyzed to obtain time-frequency domain characteristics. The time-domain features, the frequency-domain features, and the time-frequency-domain features are used as physical features characterizing the device state.

[0009] In some embodiments, the step of inputting the standardized multivariate time series data block into the trained multivariate time series representation learning network, calculating the reconstruction error and representation distance using the health baseline library, and obtaining deep features includes: The standardized multivariate time series data block is input into the trained multivariate time series representation learning network, and the activation values ​​of the intermediate or bottleneck layers of the network are obtained as deep representations. Using the health baseline embedding space in the health baseline library, calculate the reconstruction error between the reconstruction output and the original input after the standardized multivariate time series data block is input into the multivariate time series representation learning network, and the representation distance between the standardized multivariate time series data block and the health baseline embedding space; The depth representation, the reconstruction error, and the representation distance are used as depth features.

[0010] In some embodiments, fusing the physical features with the deep features and using a feature selection algorithm to filter out the feature subset most correlated with degradation to obtain an interpretable degradation feature vector includes: The physical features and the depth features are concatenated to form a high-dimensional hybrid feature vector; A feature selection algorithm based on mutual information or LASSO regression is used to select the feature subset that is most correlated with device degradation from the high-dimensional mixed feature vector; The aforementioned feature subset is used as an interpretable degenerate feature vector.

[0011] In some embodiments, inputting the interpretable degradation feature vector and the working condition label data into an attention-based working condition decoupling network to obtain working condition-invariant degradation features includes: The interpretable degradation feature vector and the working condition label data are input together into the working condition decoupling network based on the attention mechanism; The working condition decoupling network learns a set of attention weights and dynamically adjusts the weights of feature dimensions in the interpretable degradation feature vector that are below a preset sensitivity value for the current working condition but above a preset sensitivity value for degradation to a preset range. At the same time, it reduces the weights of feature dimensions that change drastically with the working condition but are unrelated to degradation to a preset range. The operating condition decoupling network outputs a pure degradation feature that is independent of the operating condition, which is then used as the degradation feature that remains unchanged by the operating condition.

[0012] In some embodiments, the step of inputting the degradation characteristics of the unchanged operating conditions into a regressor to calculate a health score, and obtaining a health warning report for electromechanical equipment based on the health score, includes: The degradation characteristics under unchanged operating conditions are input into a regressor to obtain a health score. The regressor performs forward computation using a fully connected neural network or an XGBoost model and outputs the health score in the range of 0 to 100, where 100 represents brand new and 0 represents complete failure. The regressor is trained using historical data with the goal of minimizing prediction error. The health score is used to reflect the degree of degradation of the device's current performance relative to a preset health baseline. The multiple health scores obtained through continuous calculation are arranged in chronological order to form a health time series. The health time series is fitted with exponential smoothing or linear regression to fit the trend, a failure threshold is set, and the time required for the health score to drop to the failure threshold is calculated by extrapolation based on the fitted curve. The time is used as the predicted value of the remaining service life. When the health score is lower than a preset threshold, a health warning report for electromechanical equipment is generated, which includes the current health score and the predicted value of the remaining service life.

[0013] To achieve the above objectives, another aspect of this application proposes a health assessment system for electromechanical equipment based on unsupervised baseline learning and operating condition decoupling, for implementing the method described above. The system includes: The multi-source heterogeneous data acquisition and preprocessing module is used to acquire and preprocess multi-source sensor data and operating condition label data at the current moment to form standardized multivariate time series data blocks. The unsupervised health baseline construction module is used to acquire unlabeled historical operating data of the health status of electromechanical equipment, train a multivariate time series representation learning network, construct a health baseline embedding space and statistical baseline, and store them in the health baseline library. An interpretable degradation feature extraction and fusion module is used to extract information features from the standardized multivariate time series data block to obtain physical features representing the device state; input the standardized multivariate time series data block into the trained multivariate time series representation learning network, calculate the reconstruction error and representation distance using the health baseline library to obtain deep features; fuse the physical features and the deep features, and use a feature selection algorithm to select the feature subset with the highest correlation to degradation to obtain an interpretable degradation feature vector; The working condition adaptive health calculation module is used to input the interpretable degradation feature vector and the working condition label data into an attention mechanism-based working condition decoupling network to obtain working condition invariant degradation features. The prediction, early warning, and decision output module is used to input the degradation characteristics of the unchanged operating conditions into the regressor to calculate the health score, and to obtain a health early warning report for electromechanical equipment based on the health score.

[0014] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0015] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described above.

[0016] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.

[0017] The embodiments of this application include at least the following beneficial effects: This application provides a method, system, electronic device, storage medium, and program product for health assessment of electromechanical equipment based on unsupervised baseline learning and operating condition decoupling. The method includes: acquiring unlabeled historical operating data of electromechanical equipment under health conditions, constructing a health benchmark embedding space and statistical baseline; extracting information features from standardized multivariate time series data blocks to obtain physical features characterizing the equipment state; calculating reconstruction error and characterization distance to obtain deep features; fusing physical features and deep features to obtain an interpretable degradation feature vector; inputting the interpretable degradation feature vector and operating condition label data into an attention-based operating condition decoupling network to obtain operating condition-invariant degradation features; inputting the operating condition-invariant degradation features into a regressor to calculate a health score, and obtaining an electromechanical equipment health early warning report based on the health score. This application can achieve accurate assessment of the health status of electromechanical equipment. Attached Figure Description

[0018] Figure 1 This is a flowchart of the electromechanical equipment health assessment method based on unsupervised baseline learning and operating condition decoupling provided in the embodiments of this application; Figure 2 This is a system architecture diagram provided in the embodiments of this application; Figure 3 This is a flowchart of the electromechanical equipment health assessment method provided in the embodiments of this application; Figure 4 This is a flowchart illustrating the standardized implementation process of migration deployment provided in this application embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0022] 1) Electromechanical equipment: The target objects to be monitored and maintained, generally referring to equipment or assets that integrate mechanical components and electrical / electronic control systems, and achieve specific functions through their coordinated operation. Examples include: fans, generators, compressors, pump sets, industrial robots, CNC machine tools, and transmission systems.

[0023] 2) Operational Data: During the normal operation of electromechanical equipment, a series of quantitative information reflecting the equipment's working status, performance, and environmental conditions are collected in real time or periodically by its built-in or external sensors and data acquisition systems. This includes, but is not limited to: vibration, noise, temperature, pressure, current, voltage, power, speed, control signals, alarm logs, etc.

[0024] 3) Predictive maintenance: Through continuous analysis and modeling of equipment operation data, potential failure modes are identified, the health status is assessed, and the future failure time is predicted before observable degradation of equipment performance occurs but before functional failure. This allows maintenance activities to be planned and executed at the most appropriate time, with the lowest cost and minimal production disruption.

[0025] 4) Health Status: A comprehensive evaluation index used to quantify the overall operating status of electromechanical equipment. It is not a simple binary judgment of "normal / faulty," but a continuous numerical value or level (e.g., a percentage of 0-100%, or a "excellent, good, average, poor, faulty" level) used to characterize the degree of deviation of the equipment's current performance from its new or standard health baseline. The lower the health status, the more significant the equipment performance degradation and the higher the risk of failure.

[0026] 5) Health Baseline: This refers to the standard reference model formed by the key operational data characteristics (such as statistical values, spectral characteristics, waveforms, etc.) of electromechanical equipment under a specific and recognized "healthy" or "normal operating" state (usually the initial stable operation period after installation and commissioning). This baseline serves as a benchmark for assessing the current health status and detecting anomalies.

[0027] 6) Degradation characteristics / fault characteristics: Quantitative indicators extracted from raw operating data through signal processing, time-frequency analysis, feature engineering, etc., which can effectively characterize the equipment performance degradation mode or specific type of fault, such as the energy value of vibration signal in a specific frequency band, the trend slope of temperature signal, and the harmonic distortion rate of current.

[0028] 7) Health Assessment Model: A computational model or model system based on data analysis and machine learning. This model takes real-time or historical operating data of the equipment as input, processes it through a specific algorithm, and outputs an assessment result of the equipment's current health status (such as health score, health level) and / or a prediction of future failures (such as remaining service life). The health assessment model in this application is a hybrid framework consisting of a stacked sparse autoencoder (for unsupervised health baseline construction), a mutual information feature selector (for feature filtering), an attention-based operating condition decoupling network (for removing operating condition interference), and an XGBoost regressor (for health calculation). In the AHU embodiment, multivariate Gaussian process regression is specifically used to construct the health baseline, and XGBoost is used as the health regressor, achieving electromechanical equipment health assessment that requires no fault data, is interpretable, and is resistant to operating condition interference.

[0029] 8) Machine Learning (ML): A branch of artificial intelligence that uses algorithms and models to enable computers to automatically learn from data and make predictions or decisions. Its core goal is to allow computers to analyze large amounts of data, identify patterns and rules, and build models that adapt to new data without explicit programming instructions. Machine learning includes different types such as supervised learning, unsupervised learning, and reinforcement learning.

[0030] 9) Remaining Useful Life (RUL): A key output metric for predictive maintenance, specifically referring to the length of time from the current assessment point until the equipment is expected to experience functional failure (i.e., be unable to perform its intended function). RUL prediction is based on the equipment's current degradation trajectory and historical failure data.

[0031] 10) Prognostics and Health Management (PHM): Significantly improves operational efficiency by monitoring equipment status in real time, predicting failures, and optimizing maintenance strategies.

[0032] 11) Robustness: The ability of a system, model, etc., to maintain its performance and operate stably under various disturbances or anomalies such as input errors, data disturbances, and environmental changes.

[0033] 12) XGBoost: Extreme Gradient Boosting is a machine learning algorithm based on decision tree ensemble. It iteratively trains multiple decision trees and combines them. Each new tree attempts to correct the prediction errors of all the previous trees, eventually resulting in a strong prediction model.

[0034] 13) AHU: Air Handling Unit, an electromechanical device used to regulate and circulate air. It is usually composed of components such as a fan, filter, cooling / heating coil, and humidifier. It is mainly used in central air conditioning systems to cool, heat, humidify, and filter air.

[0035] 14) Unsupervised baseline learning: A method that uses only unlabeled operational data of equipment in its health state (without requiring manual labeling of fault types) to automatically learn the inherent patterns of normal operating modes of the equipment through unsupervised algorithms such as autoencoders and contrastive learning, and constructs a health state reference benchmark.

[0036] 15) Operating condition decoupling: This refers to distinguishing changes in the operating status of equipment into "normal fluctuations caused by changes in operating conditions" and "abnormal changes caused by performance degradation", and using algorithms to eliminate the interference of operating condition changes on health assessment and extract pure degradation features that are unrelated to operating conditions.

[0037] 16) Health benchmark embedding space: A low-dimensional vector space extracted from health status data through unsupervised learning. It represents the standard reference system for the distribution of data in the health status of the device and is used to quantify the degree of deviation between the current data and the health status.

[0038] 17) Reconstruction Error: This refers to the degree of difference between the output data and the original input data after the input data has been encoded and decoded by a network such as an autoencoder. The greater the difference, the further the current data pattern deviates from the healthy data pattern learned during network training. For example, if the sensor data of the current AHU is input into a trained autoencoder, the network outputs reconstructed data. If the root mean square error between the original data and the reconstructed data is 0.05, while the error during the healthy period is only 0.01, then the increase in reconstruction error indicates that the device has deviated from a healthy state.

[0039] 18) Representation Distance: This refers to the spatial distance between the low-dimensional representation vector of the current data and the center of the health data representation distribution in the health baseline embedding space. The larger the distance, the greater the difference between the current equipment state and the healthy state, and the more severe the degradation. For example, the representation vector of the AHU data in a healthy state is clustered near the center point in the embedding space; when the fan belt is loose, the Euclidean distance between the current data representation vector and the center point increases from 0.2 to 1.5, indicating that the equipment has undergone significant degradation.

[0040] 19) Mutual information: In information theory, this is a measure of the degree of interdependence between two random variables. The larger the value, the stronger the correlation between the two variables, and the better it can capture linear and nonlinear relationships. For example, calculating the mutual information between "filter differential pressure rise rate" and "filter clogging degree" yields a high value of 0.85, indicating that this feature is highly correlated with clogging degradation; while the mutual information between "supply air temperature residual" and clogging is only 0.03, indicating a very low correlation, and it can be eliminated.

[0041] 20) LASSO (Minimum Absolute Shrinkage and Selection Operator) Regression: This is a linear regression method that compresses the regression coefficients of irrelevant or redundant features to zero by adding an L1 regularization term to the loss function, thereby achieving automatic feature selection.

[0042] 21) Stacked Sparse Autoencoders: These are deep neural networks composed of multiple sparse autoencoders stacked layer by layer. They learn hierarchical feature representations of the input data through unsupervised pre-training layer by layer, and introduce sparsity constraints to force the network to learn the most important features in the data. For example, 12-dimensional sensor data such as vibration, temperature, and current from an AHU can be input into a stacked sparse autoencoder. The first layer learns simple time-domain statistical features, the second layer learns frequency-domain energy distribution, and the final bottleneck layer outputs a 32-dimensional low-dimensional vector as the health baseline embedding space.

[0043] 22) Temporal Comparison Learning Network: This is a self-supervised learning method for processing time series data. It learns representations with good discriminative power by bringing health data from the same device at different times closer together in the representation space (positive sample pairs) and pushing healthy data away from abnormal data (negative sample pairs). For example, using AHU's health data from different dates as positive sample pairs, the network is required to map their representations to neighboring locations in the embedding space; using healthy data and simulated fault data as negative sample pairs, the network is required to push them away, thereby enhancing the model's sensitivity to abnormal states.

[0044] 23) PLC: Programmable Logic Controller, a digital computer used for industrial control. It monitors sensor inputs and controls actuator outputs by executing user-written logic programs to achieve automated control of equipment.

[0045] 24) DDC: Direct Digital Controller, a digital controller used in building automation systems. It directly collects and controls analog quantities such as temperature, humidity, and pressure through a built-in microprocessor and is used for local control of HVAC systems.

[0046] 25) BACnet / IP: A data communication protocol for building automation and control networks, based on IP network transmission, used for data exchange and interoperability between devices from different manufacturers in building automation systems.

[0047] 26) ITAE: Time-integrated-absolute-error, a performance evaluation index for control systems. It is used to evaluate the dynamic response performance of control loops by integrating the absolute error multiplied by time and giving greater penalty weight to persistent errors.

[0048] This invention takes into account that, with the development of technology, electromechanical equipment is evolving towards greater complexity, integration, and automation. Traditional equipment maintenance strategies, such as "reactive maintenance" (repairing after a failure occurs) and "periodic maintenance" (maintenance at fixed intervals), are no longer sufficient to meet the requirements of modern civil and industrial buildings for high reliability, high availability, and low operating costs. Reactive maintenance leads to unplanned downtime, causing significant production losses; while periodic maintenance may result in "over-maintenance" (unnecessary disassembly and repair of equipment in good condition) or "under-maintenance" (failures occurring during maintenance intervals), leading to resource waste or unexpected downtime.

[0049] Against this backdrop, predictive maintenance has emerged. The goal of PHM is to assess the health status of equipment and predict future failures through continuous monitoring and analysis of equipment operating data, thereby enabling the right maintenance to be performed on the right components at the right time. Its core technological approach involves building data-driven health assessment models to extract effective information characterizing equipment performance degradation from massive amounts of multi-dimensional operating data.

[0050] The mainstream methods for data-driven equipment health assessment are mainly divided into two categories: physical model-based methods and data-driven model-based methods. Among them, data-driven model-based methods have become the focus of research and application because they do not require precise knowledge of physical failure mechanisms and can learn directly from historical data.

[0051] In view of this, this application provides a method and related equipment for health assessment of electromechanical equipment based on unsupervised baseline learning and decoupling of operating conditions. This scheme utilizes unlabeled historical operating data of equipment in its health state to construct a health benchmark embedding space, solving the problem of dependence on a large amount of fault labeling data in traditional methods. Through interpretable degradation feature extraction and fusion, time-frequency domain features with clear physical meaning are combined with deep features (reconstruction error, representation distance) based on the health baseline to generate feature vectors with traceable degradation evidence, solving the "black box" problem of the model. Through an operating condition decoupling network based on an attention mechanism, features sensitive to degradation are dynamically amplified and feature variations related to operating conditions are suppressed, outputting pure degradation features unrelated to operating conditions, solving the problem of false alarms and missed alarms caused by operating condition fluctuations. Finally, through regressors and trend prediction, continuous health scores, predicted values ​​of remaining service life, and interpretable early warning reports are output, realizing a complete closed loop from quantitative assessment of equipment status to predictive maintenance decisions. At the same time, its modular design supports rapid migration and deployment across equipment, reducing the technical threshold and operation and maintenance costs of implementing predictive maintenance for large-scale electromechanical equipment groups.

[0052] Figure 1 This is an optional flowchart of a method for health assessment of electromechanical equipment based on unsupervised baseline learning and decoupling of operating conditions, provided in an embodiment of this application. Figure 1 The method may include, but is not limited to, steps S101 to S107.

[0053] Step S101: Collect and preprocess the multi-source sensor data and operating condition label data at the current moment to form a standardized multivariate time series data block; Step S102: Obtain unlabeled historical operating data of the health status of electromechanical equipment, train a multivariate time series representation learning network, construct a health baseline embedding space and statistical baseline, and store them in the health baseline library. Step S103: Extract information features from the standardized multivariate time series data blocks to obtain physical features characterizing the equipment status; Step S104: Input the standardized multivariate time series data block into the trained multivariate time series representation learning network, and use the healthy baseline library to calculate the reconstruction error and representation distance to obtain deep features; Step S105: The physical features and deep features are fused, and a feature selection algorithm is used to select the feature subset with the highest correlation to degradation to obtain an interpretable degradation feature vector. Step S106: Input the interpretable degradation feature vector and the working condition label data into the working condition decoupling network based on the attention mechanism to obtain the working condition invariant degradation features. Step S107: Input the degradation characteristics with unchanged operating conditions into the regressor to calculate the health score, and obtain the health warning report of the electromechanical equipment based on the health score.

[0054] Steps S101 to S107 as illustrated in the embodiments of this application construct a baseline using unsupervised health baseline learning and unlabeled health status data to address the dependence of cold start on fault data; through the fusion of physical and deep features, traceable degradation evidence is output to improve model interpretability; through an attention mechanism, operating condition interference is decoupled to output pure degradation features independent of operating conditions, enhancing robustness under varying operating conditions; finally, a health and remaining service life report is generated through regression prediction, realizing a complete closed loop from quantitative assessment to predictive maintenance, and supporting rapid migration across devices to reduce the cost of large-scale deployment.

[0055] In some embodiments, step S101 may include, but is not limited to, steps S111 to S114: Step S111: Data is collected synchronously from vibration sensors, temperature sensors, current sensors, voltage sensors, and the device controller to obtain multi-source sensor data at the current moment; Step S112: Extract operating condition tag data from the manufacturing execution system or building equipment monitoring system; the operating condition tag data includes speed setpoint, load current and operating mode code; Step S113: Use timestamp alignment and resampling technology to synchronize the channels of multi-source sensor data and working condition tag data with different sampling frequencies to form data slices; Step S114: Remove outliers from the data slices using statistical methods or methods based on neighboring data, and perform linear interpolation or forward imputation on missing values ​​to obtain standardized multivariate time series data blocks.

[0056] Steps S111 to S114, as illustrated in the embodiments of this application, involve synchronously collecting multi-source sensor data and operating condition label data, performing timestamp alignment, outlier removal, and missing value imputation to form standardized multivariate time series data blocks. This provides high-quality, uniformly formatted input data for subsequent feature extraction. For example, in an air handling unit, data is synchronously collected through vibration sensors, temperature sensors, and building automation systems. After alignment and cleaning, standardized data slices are formed at a rate of once per minute, avoiding subsequent calculation errors caused by different sampling frequencies or missing data.

[0057] In some embodiments, step S102 may include, but is not limited to, steps S201 to S203: Step S201: Obtain untagged historical operating data of electromechanical equipment during the initial stable operation period after installation and commissioning, and which is determined to be in a healthy state by preset rules; Step S202: Using unlabeled historical running data as training samples, a multivariate temporal representation learning network is trained; the multivariate temporal representation learning network is a stacked sparse autoencoder or a temporal contrastive learning network. Step S203: The low-dimensional vector output by the multivariate temporal representation learning network is used as the health baseline embedding space, and the mean and variance upper and lower bounds of each channel of the health data are calculated as statistical baselines. The trained network model parameters and statistical baselines are stored together in the health baseline library. The health baseline embedding space is used as a low-dimensional benchmark reference system to represent the data distribution pattern of the device in a healthy state. It is used to quantify the device performance degradation by calculating the degree of deviation between the current data and the health baseline embedding space.

[0058] Steps S201 to S203, as illustrated in the embodiments of this application, acquire unlabeled historical operating data of the device's health status, perform unsupervised training of a multivariate temporal representation learning network, and construct a health baseline embedding space and statistical baseline, storing them in a health baseline library. This allows the model to establish a reference benchmark for health status without fault-labeled data. For example, in a newly installed AHU, 30 days of stable operation data with a completely new filter in spring are selected, and a health baseline embedding space is obtained through training with a stacked sparse autoencoder. Any subsequent data deviating from this space can be identified as degradation.

[0059] In some embodiments, step S103 may include, but is not limited to, steps S301 to S304: Step S301: Calculate the root mean square value, peak value, kurtosis, waveform factor, and impulse factor of the signal from the standardized multivariate time series data block to obtain the time domain characteristics; Step S302: Perform a fast Fourier transform on the signal in the standardized multivariate time series data block to calculate the specific fault characteristic frequency and its harmonic energy ratio that meet the preset conditions, and obtain the frequency domain characteristics; the specific fault characteristic frequency includes the bearing's passing frequency and the gear's meshing frequency. Step S303: Calculate the wavelet packet energy spectrum of the non-stationary signal in the standardized multivariate time series data block, analyze the changes in energy distribution in different frequency bands, and obtain the time-frequency domain characteristics; Step S304: Use time-domain features, frequency-domain features, and time-frequency-domain features as physical features to characterize the device state.

[0060] Steps S301 to S304, as illustrated in the embodiments of this application, involve signal processing of standardized data blocks, performing parallel calculations of time-domain, frequency-domain, and time-frequency-domain features to obtain features with clear physical meaning (such as root mean square value, kurtosis, characteristic frequency energy, and wavelet packet energy spectrum), enabling the degradation assessment results to be traced back to specific physical phenomena. For example, calculating the kurtosis value from the AHU vibration signal and inferring an increase in kurtosis can directly suggest the presence of an impact fault (such as bearing pitting); calculating the proportion of characteristic frequency energy of the wind turbine from the current signal can determine whether an electrical anomaly exists.

[0061] In some embodiments, step S104 may include, but is not limited to, steps S401 to S403: Step S401: Input the standardized multivariate time series data block into the trained multivariate time series representation learning network, and obtain the activation values ​​of the intermediate layer or bottleneck layer of the network as deep representations. Step S402: Using the health baseline embedding space in the health baseline library, calculate the reconstruction error between the reconstruction output and the original input after inputting the standardized multivariate time series data block into the multivariate time series representation learning network, and the representation distance between the standardized multivariate time series data block and the health baseline embedding space. Step S403: Depth representation, reconstruction error, and representation distance are used as depth features.

[0062] Steps S401 to S403, as illustrated in the embodiments of this application, involve inputting standardized data blocks into a trained multivariate temporal representation learning network to calculate the reconstruction error and representation distance. This yields deep features that quantify the degree to which the current data pattern deviates from the healthy pattern, capturing complex nonlinear degradation patterns that physical features cannot reflect. For example, when scale forms inside the AHU cooling coil, the physical characteristics may not change significantly, but the reconstruction error continues to increase, indicating that the equipment's operating mode has deviated from the healthy baseline, providing an early warning of the potential risk of decreased heat exchange efficiency.

[0063] In some embodiments, step S105 may include, but is not limited to, steps S501 to S503: Step S501: Concatenate the physical features and the depth features to form a high-dimensional hybrid feature vector; Step S502: Use a feature selection algorithm based on mutual information or LASSO regression to select the feature subset with the highest correlation to device degradation from the high-dimensional mixed feature vector; Step S503: Use the feature subset as the interpretable degenerate feature vector.

[0064] Steps S501 to S503, as illustrated in the embodiments of this application, involve concatenating and fusing physical features and deep features, and then executing a feature selection algorithm based on mutual information or LASSO regression. This process selects the subset of features with the highest correlation to degradation as the interpretable degradation feature vector, reducing feature dimensionality, improving computational efficiency, and enhancing the model's generalization ability. For example, in an AHU filter clogging scenario, the mutual information method automatically selects "filter differential pressure rise rate" as the most relevant feature, while "supply air temperature residual" is removed, allowing the model to focus on core degradation indicators and avoiding redundant information interference.

[0065] In some embodiments, step S106 may include, but is not limited to, steps S601 to S603: Step S601: Input the interpretable degradation feature vector and the working condition label data into the working condition decoupling network based on the attention mechanism. Step S602: The working condition decoupling network learns a set of attention weights and dynamically adjusts the weights of feature dimensions in the interpretable degradation feature vector that are below the preset sensitivity value for the current working condition but above the preset sensitivity value for degradation to a preset range. At the same time, it reduces the weights of feature dimensions that change drastically with the working condition but are unrelated to degradation to a preset range. Step S603: Use the operating condition decoupling network to output pure degradation features that are independent of the operating condition, as degradation features that are invariant to the operating condition.

[0066] Steps S601 to S603, as illustrated in this embodiment, involve inputting the interpretable degradation feature vector and operating condition label data into an attention network. This process learns attention weights and performs feature weighting, dynamically amplifying degradation-sensitive features while suppressing operating condition-related feature variations. The output is a pure degradation feature independent of the operating condition, eliminating interference from load, speed, and other operating condition fluctuations in the evaluation. For example, in cooling and ventilation modes, the normal fluctuation range of the fan current in an AHU differs. The attention network automatically learns the weight difference of the "current" feature in the two operating conditions, suppressing the contribution of the current feature in cooling mode to avoid misjudging degradation due to a normal increase in current.

[0067] In some embodiments, step S107 may include, but is not limited to, steps S701 to S704: Step S701: Input the degradation features with unchanged operating conditions into the regressor to obtain a health score; The regressor performs forward calculation through a fully connected neural network or XGBoost model and outputs a health score in the range of 0 to 100, where 100 represents brand new and 0 represents complete failure; The regressor is trained using historical data with the goal of minimizing prediction error; The health score is used to reflect the degree of degradation of the current performance of the equipment relative to the preset health baseline; Step S702: Arrange the multiple health scores obtained from continuous calculation in chronological order to form a health time series; Step S703: Apply exponential smoothing or linear regression to fit the trend of the health time series, set the failure threshold, extrapolate the fitted curve to calculate the time required for the health score to drop to the failure threshold, and use the time as the predicted value of the remaining service life. Step S704: When the health score is lower than a preset threshold, generate an electromechanical equipment health early warning report that includes the current health score and the predicted value of the remaining service life.

[0068] Steps S701 to S704, as illustrated in the embodiments of this application, input the degradation characteristics with unchanged operating conditions into the regressor, perform continuous calculation of the health score, time series trend fitting, and extrapolation prediction of remaining service life, and generate a multi-level early warning report containing the health score, the predicted value of remaining service life, and key degradation evidence, providing accurate maintenance decision support for operation and maintenance personnel. For example, when the AHU health drops to 72, the system generates an early warning report: Health 72 (medium level, pay attention to warning), remaining service life prediction 15 days (extrapolated based on the filter differential pressure rise curve), the main degradation evidence is a 12% decrease in fan efficiency from the baseline and an 8% decrease in coil heat exchange efficiency, and the maintenance recommendations are to check the fan belt, replace the filter within two weeks, and clean the coil.

[0069] This application also provides a health assessment system for electromechanical equipment based on unsupervised baseline learning and operating condition decoupling, for implementing the method described above. The system includes: The multi-source heterogeneous data acquisition and preprocessing module is used to acquire and preprocess multi-source sensor data and operating condition label data at the current moment to form standardized multivariate time series data blocks. The unsupervised health baseline construction module is used to acquire unlabeled historical operating data of the health status of electromechanical equipment, train a multivariate time series representation learning network, construct a health baseline embedding space and statistical baseline, and store them in the health baseline library. The interpretable degradation feature extraction and fusion module is used to extract information features from standardized multivariate time series data blocks to obtain physical features representing the device state; the standardized multivariate time series data blocks are input into a trained multivariate time series representation learning network, and the reconstruction error and representation distance are calculated using a health baseline library to obtain deep features; the physical features and deep features are fused, and a feature selection algorithm is used to select the feature subset with the highest correlation to degradation to obtain the interpretable degradation feature vector; The working condition adaptive health calculation module is used to input the interpretable degradation feature vector and working condition label data into the working condition decoupling network based on the attention mechanism to obtain the working condition invariant degradation features. The prediction, early warning, and decision output module is used to input the degradation characteristics with unchanged operating conditions into the regressor to calculate the health score, and to obtain a health early warning report for electromechanical equipment based on the health score.

[0070] This application takes into account the following problems with the relevant technology: (1) Strong dependence on high-quality labeled data and the "cold start" problem: Deep learning methods in related technologies are usually supervised learning, and their performance depends heavily on a large amount of complete and accurate labeled historical data (i.e., it is necessary to know the health status or remaining lifespan of the equipment corresponding to each data point). However, in civilian or industrial settings, it is extremely difficult and costly to obtain full lifecycle data of equipment from brand new to failure. In most cases, only a large amount of "unlabeled" normal operation data and a small amount of fault moment data can be obtained. This leads to a "cold start" dilemma in model training. Without sufficient labeled data, the model is difficult to build or performs poorly.

[0071] (2) Poor interpretability of the model and low trust among maintenance personnel: Deep learning models are like "black boxes," with opaque internal decision-making logic. When the model outputs a warning of "declining health," maintenance engineers cannot know which component, parameter, or physical degradation mode caused the judgment. The lack of interpretability makes it difficult for field engineers to quickly locate the root cause of the problem and to build trust in the model's prediction results, thus affecting the effective response to warning information and decision-making.

[0072] (3) Insufficient robustness to changes in operating conditions: The operating status of electromechanical equipment often changes dynamically with factors such as production plans, load, and environment (i.e., multi-condition operation). The models in related technologies are usually trained on fixed operating condition data. When the operating conditions of the equipment change, the data distribution will shift, causing the model to misjudge normal operating condition changes as performance degradation, generating a large number of false alarms, which seriously affects the practicality and reliability of the model.

[0073] (4) Weak model generalization and transfer capabilities: Models trained for specific models or devices are often difficult to apply directly to other devices of the same type but different models or with different installation environments. This means that each new device may require the collection of full lifecycle data and the training of a new model, resulting in long project implementation cycles, high costs, and difficulty in rapid large-scale deployment.

[0074] The purpose of this application is to overcome the limitations of the prior art through an innovative model architecture and algorithm design, and to achieve more accurate, reliable, practical and economical predictive maintenance of electromechanical equipment. This application includes the following embodiments: 1. Overview of the overall technical solution: The core of this invention lies in abandoning the traditional, end-to-end "black box" modeling approach and instead adopting a "phased, interpretable, and knowledge-guided" hybrid modeling framework. This framework decouples the health assessment task into three logically clear and functionally defined stages, and innovatively introduces unsupervised health baseline learning, physical information feature enhancement, and attention-based work condition decoupling mechanisms, thereby systematically solving the four major problems raised in the background art.

[0075] 2. System Architecture and Module Details: like Figure 2 As shown, the health assessment system of the present invention mainly includes the following five core modules: multi-source heterogeneous data acquisition and preprocessing module, unsupervised health baseline construction module, interpretable degradation feature extraction and fusion module, working condition adaptive health calculation module, and prediction, early warning and decision output module.

[0076] 2.1 Multi-source heterogeneous data acquisition and preprocessing module: This module is responsible for acquiring raw operating data from the target electromechanical equipment, performing preliminary cleaning and formatting to provide high-quality input for subsequent analysis.

[0077] Data sources include, but are not limited to, vibration sensors, current / voltage sensors, temperature sensors, pressure sensors, noise sensors, digital / analog signals output by equipment controllers (such as PLCs and DDCs), and production logs (such as load, speed, and operating mode) related to equipment operation in Manufacturing Execution Systems (MES) / Building Equipment Monitoring Systems (BAS).

[0078] Data Synchronization and Alignment Unit: Since the sampling frequency and data timestamps of different sensors may be different, this unit adopts timestamp alignment and resampling technology to ensure that the data of all channels are strictly synchronized in time, forming a unified "data slice".

[0079] Outlier handling and missing value filling unit: Using statistical methods (such as the 3σ principle) or methods based on neighboring data, outliers caused by momentary sensor failure are identified and removed, and linear interpolation or forward filling is performed on a small number of missing data.

[0080] 2.2 Unsupervised Health Baseline Construction Module (Corresponding to technical problem 1): This module is key to reducing the reliance on fault labeling data in this invention. Its core lies in automatically learning the "health fingerprint" of the device by utilizing only "unlabeled" operational data from the early known healthy state of the device (such as the first N hours of stable operation after installation and commissioning).

[0081] Health Data Selector: Filters data from historical databases that have been judged as "healthy" by domain experts or simple rules.

[0082] Multivariate Temporal Representation Learning Network: This is a core innovative submodule. We designed a deep learning model based on a stacked sparse autoencoder or a temporal contrastive learning network. This model takes multi-sensor synchronous temporal data from a healthy state as input and learns its high-dimensional, low-redundancy latent representations in an unsupervised manner. The training objective is to enable the model to regenerate the input data with extremely low reconstruction error, or to map the health data of the same device at different times into compact clusters in the representation space. The low-dimensional vector output by the "bottleneck layer" of this trained network constitutes the baseline embedding space of the data distribution of the device in its healthy state.

[0083] Health baseline library: Stores the device health baseline embedding space model learned by the module, as well as key statistical features of health data (such as the mean and upper and lower bounds of variance for each channel). This constitutes the "benchmark" for subsequent evaluation.

[0084] 2.3 Explainable Degradation Feature Extraction and Fusion Module (Corresponding to technical problem 2): This module aims to extract a set of features from the current running data that can both characterize degradation and have clear physical meaning, rather than "black box" features.

[0085] The physical information feature engineering unit is divided into time-domain feature extraction, frequency-domain feature extraction, and time-frequency-domain feature extraction.

[0086] Time-domain feature extraction: Calculating classic time-domain metrics such as RMS (root mean square), peak value, kurtosis, waveform factor, and impulse factor of a signal. These metrics are directly related to physical phenomena such as vibration energy and impact intensity.

[0087] Frequency domain feature extraction: Perform a Fast Fourier Transform on the signal to calculate the specific fault characteristic frequencies (such as the bearing's passing frequency or the gear's meshing frequency) and the energy proportion of their harmonics. This is directly related to the specific fault type.

[0088] Time-frequency domain feature extraction: For non-stationary signals, calculate the wavelet packet energy spectrum and analyze the changes in energy distribution in different frequency bands.

[0089] The data-driven deep feature extraction unit inputs the current data slice into a pre-trained multivariate temporal representation learning network, but instead of using a "bottleneck layer," it extracts the activation values ​​of one or several intermediate layers as deep features. Simultaneously, it calculates the reconstruction error and representation distance of the current data in the baseline embedding space of a healthy baseline database. The reconstruction error and representation distance are themselves powerful and interpretable degradation metrics, quantifying the degree to which the current data pattern deviates from the "healthy pattern."

[0090] Feature Fusion and Selector: Physical features from the Physical Information Feature Engineering Unit and deep features from the Data-Driven Deep Feature Extraction Unit, along with distance metrics, are concatenated to form a high-dimensional hybrid feature vector. Then, a feature selection algorithm based on mutual information or LASSO regression is used to automatically select a subset of features most correlated with device degradation, avoiding the curse of dimensionality and improving model efficiency. The final output is an interpretable degradation feature vector, where each dimension (or combination of dimensions) can be traced back to a specific physical quantity or data pattern deviation.

[0091] 2.4. Adaptive Working Condition Health Calculation Module (Corresponding to technical problem 3): This module is the core of improving the robustness of the model, and its goal is to eliminate the interference of operating condition fluctuations on health assessment.

[0092] Operating condition identification unit: Extracts operating condition label variables that directly characterize the equipment's operating status from raw data or MES or BAS system logs, such as speed setpoint, load current, and operating mode code. These variables are known and measurable.

[0093] An attention-based working condition decoupling network takes interpretable degenerate feature vectors and working condition label variables as input. Its function is to learn a set of attention weights that dynamically "amplify" feature dimensions that are insensitive to the current working condition but sensitive to degradation, while "suppressing" feature dimensions that change drastically with the working condition but are unrelated to degradation. Essentially, this network learns a mapping from "original features + working condition" to "working condition-invariant degenerate features".

[0094] Health Regressor: This regressor takes the clean, condition-independent degradation features generated by the condition-decoupled network and inputs them into a fully connected neural network or gradient boosting decision tree (such as XGBoost) for regression. The output of this regressor is a continuous health score (e.g., 0-100, where 100 represents brand new and 0 represents complete failure). This score is a measure of the purity of the inherent performance degradation of the equipment after filtering out condition interferences.

[0095] 2.5 Forecasting, Early Warning, and Decision Output Module: This module performs trend analysis and decision support based on calculated health scores.

[0096] Health score series and trend fitting unit: The health scores calculated over a period of time are arranged in chronological order to form a health score time series. Exponential smoothing or linear / nonlinear regression is used to fit this series to predict its short-term future trend.

[0097] RUL Prediction Unit: Sets a failure threshold (e.g., health = 20). Based on the fitted curve of the health time series, extrapolates the time required for the health to drop to that threshold, i.e., the predicted remaining useful life.

[0098] Adaptive Early Warning and Report Generation Unit: Based on the health status, RUL, and the exceeding of key physical features in the degradation feature vector, multi-level early warnings (such as "Caution", "Abnormal", "Danger") are set. An interpretable report is automatically generated, including health score, RUL prediction, major degradation features (such as "significant increase in vibration kurtosis" and "bearing outer ring frequency energy percentage exceeds threshold"), and maintenance recommendations.

[0099] 3. Overview of the Method and Flow: like Figure 3 As shown, the method includes the following steps: STEP 1: Data Acquisition and Preprocessing. Simultaneously acquire time-series data from various sensors of the target device, along with operational condition labels, and align, clean, and slice the data to form standardized multivariate time-series data blocks.

[0100] STEP 2: Unsupervised Health Baseline Learning (Offline Stage). Using historical health data of the device, train the multivariate temporal representation learning network to construct the device's health baseline embedding space and statistical baseline, and store them in the health baseline library.

[0101] STEP 3: Real-time / Online Degradation Feature Extraction and Fusion. For the current data block to be evaluated, execute in parallel: STEP3a: Calculate classical physical information features (time domain, frequency domain, time-frequency domain).

[0102] STEP3b: Using the network trained by STEP2, calculate the reconstruction error and representation distance of the current data block in the healthy baseline embedding space.

[0103] STEP3c: Integrates and filters physical features and depth distance features to form an interpretable degradation feature vector.

[0104] STEP4: Adaptive health calculation under operating conditions.

[0105] STEP4a: Identify the operating condition label corresponding to the current data block.

[0106] STEP4b: Input the degraded feature vector and the working condition label together into the working condition decoupling network based on the attention mechanism to obtain the pure degraded features after filtering out the working condition interference.

[0107] STEP4c: Input the pure degradation features into the health regressor to calculate the health score at the current moment.

[0108] STEP5: Forecasting, Early Warning and Output.

[0109] STEP5a: Aggregates historical health sequences to perform trend fitting and RUL prediction.

[0110] STEP5b: Based on preset rules, determine whether an alert is triggered and generate an interpretable report containing health status, RUL, and degradation characteristic descriptions.

[0111] STEP5c: Outputs evaluation results and reports to support maintenance decisions.

[0112] 4. Implementation methods for model generalization and transfer (corresponding to technical problem 4): This invention significantly improves the generalization and transfer capabilities of the model through its modular and decoupled design concept and the reusability of its core components, specifically in the following three aspects: ①. Modular decoupling at the architecture level supports "plug and play" migration.

[0113] The overall architecture of this invention is clearly divided into five modules: data preprocessing, baseline construction, feature extraction, health calculation, and early warning output. This design allows for loose coupling between the core algorithm logic of the model and the data patterns of specific devices. When migrating this solution to a new device of the same type but a different model, engineers can reuse most of the modules (such as feature engineering logic, attention network architecture, and health regression model structure), while only needing to retrain or adjust a few device-related modules. This significantly reduces the workload of redevelopment.

[0114] ②. The universality of core technologies reduces the requirements for specific data on equipment.

[0115] Unsupervised baseline learning is universally applicable. The method used in this invention to construct a health baseline is a domain-unsupervised feature learning approach. It does not rely on any device-specific fault mode labels; its learning objective is to better represent the normal patterns of health data. This objective is common to any rotating machinery, reciprocating equipment, etc., therefore the algorithm framework itself has extremely strong cross-device generalization potential.

[0116] The decoupling concept of attention mechanisms is universally applicable. Attention-based condition decoupling networks learn the general ability to distinguish between features related to operating conditions and those related to degradation. This core problem is similar for many industrial devices affected by load and rotational speed. Therefore, a decoupling network pre-trained on one device can be quickly adapted to another device through transfer learning (such as fine-tuning), accelerating the convergence of the new model.

[0117] ③. Standardize the implementation process to ensure rapid deployment.

[0118] This invention provides a standardized implementation process, such as... Figure 4 As shown. For new devices, migration deployment simply requires following these clear steps: STEP A (Data Preparation): Collect short-term health operation data for the new equipment.

[0119] STEP B (Baseline Personalization): Input the data from A into the unsupervised health baseline building module to train a health baseline model specific to this device. This is the only core step that needs to be customized for a new device and does not require fault data.

[0120] STEP C (Model Loading and Fine-tuning): Loads the feature engineering logic from the interpretable degradation feature extraction module, which has been trained on other similar devices, as well as the attention network and regressor model from the condition-adaptive health calculation module. It allows for fine-tuning of the model using limited data from the new device to quickly adapt to individual differences.

[0121] STEP D (Go Online): Integrate the results of B and C into the system and begin online evaluation.

[0122] In summary, this invention effectively addresses the pain points of existing technologies, such as "one model per machine" and difficulty in large-scale deployment, through its innovative modular architecture, universally applicable core algorithms, and standardized migration process. It achieves the goal of "develop once, deploy multiple times quickly," significantly reducing the overall technical threshold and long-term operation and maintenance costs of implementing predictive maintenance on large-scale equipment clusters.

[0123] This embodiment uses a combined air handling unit (AHU) used in a large commercial complex as an example to describe the specific implementation process of the present invention in detail. The AHU includes a mixing section, a pre-filter / medium-efficiency filter section, a cooling / heating coil section, a humidification section, a fan section, and a noise reduction section. Its core function is to ensure stable temperature, humidity, cleanliness, and airflow of the supplied air.

[0124] 1. System Structure and Data Preparation: Target equipment: One unit with an air volume of 50,000 m³ / h 3 The combined AHU with a frequency conversion fan is / h.

[0125] Data source and measurement point layout (corresponding to the multi-source heterogeneous data acquisition and preprocessing module): Airflow / differential pressure sensor: static pressure difference between the inlet and outlet of the fan, and pressure difference before and after the filter (used to calculate airflow and filter blockage).

[0126] Temperature and humidity sensors: dry bulb temperature and relative humidity of fresh air, return air, and supply air.

[0127] Actuator feedback: opening degree of surface cooling / heating water valve, humidifier opening degree, and fan frequency feedback.

[0128] Current / Power Sensor: Collects the operating current and real-time power of the fan motor.

[0129] Control system data: Read setpoints (such as supply air temperature setpoint and static pressure setpoint) and equipment operating modes from the building automation system (BAS).

[0130] Data synchronization: All sensor and BAS data are connected to the edge computing gateway via the BACnet / IP protocol, with unified timestamps, forming multivariate time series data with 1-minute intervals.

[0131] 2. Unsupervised health baseline construction (offline stage, corresponding to the unsupervised health baseline construction module): Health data selection: 30 consecutive days of stable AHU operation during spring (when ambient temperature and humidity are moderate and there are no extreme loads) with newly replaced filters were selected as health data. During this period, the equipment had no alarms, and the supply air temperature and humidity control accuracy met the requirements.

[0132] Baseline model training: Data preprocessing: Health data is cleaned, removing data from transitional periods during equipment start-up and shutdown. For different seasonal operating conditions, separate "summer cooling mode baselines" and "winter heating mode baselines" can be established. This embodiment uses "cooling mode" as an example.

[0133] Model Construction: Multivariate Gaussian process regression is used to model the health data. Gaussian process regression can learn not only the mean trend of health data, but also its fluctuation range (covariance). Input features are operating parameters: [fresh air temperature, return air temperature, fan frequency]. Output are key performance parameters: [supply air temperature, fan current, filter differential pressure].

[0134] Training: Using fitted health data, a set of mapping relationship models is obtained. This model can not only predict the expected performance value of a healthy AHU under a given operating condition, but also provide the prediction confidence interval.

[0135] Baseline storage: Store the trained model parameters, the distribution range of healthy operating condition data, and the benchmark range of fan efficiency (air volume / power) and coil heat exchange efficiency (heat exchange / flow) under healthy conditions into the healthy baseline library.

[0136] 3. Real-time interpretable degradation feature extraction (online stage, corresponding to the interpretable degradation feature extraction and fusion module): For real-time data streams, perform the following operations: 3.1 Calculation of physical information characteristics: Energy efficiency characteristics: Real-time calculation of fan efficiency and surface cooling coil sensible heat exchange efficiency, compared with baseline efficiency.

[0137] Control performance characteristics: Calculate the absolute error of the time integral between the supply air temperature and the set value to evaluate the performance of the temperature control loop.

[0138] Resistance characteristics: Calculate the net pressure difference of the filter (pressure difference before and after) and observe its upward trend over time.

[0139] Air volume balance characteristics: Based on the fan law, at the same frequency, compare the deviation of the current air volume (estimated by pressure difference) with the healthy baseline air volume.

[0140] 3.2 Data-driven deep feature calculation: Input the current operating condition [T_new, T_return, Freq_fan] into the trained model.

[0141] Calculate the residuals: Obtain the model's predicted values ​​and confidence intervals for [T_supply, I_fan, dP_filter]. Calculate the standardized residuals between the actual observed values ​​and the predicted mean. For example, (T_supply_actual - T_supply_predicted) / prediction_std. The statistical properties of this residual sequence (such as mean shift and increased variance) are important indicators of degradation.

[0142] Calculate the likelihood probability: Calculate the marginal log-likelihood of the current observed data point under the model. A continuously decreasing value indicates that the current operating pattern is increasingly deviating from the statistical distribution of healthy data.

[0143] 3.3 Feature Fusion and Selection: The energy efficiency characteristics, control characteristics, and drag characteristics are fused with the model's residual characteristics and likelihood probability characteristics. Mutual information is used to select the feature subset most relevant to known common AHU faults (such as filter clogging, coil scaling, and fan belt slippage). The resulting interpretable feature vector may be: [fan efficiency reduction rate, filter differential pressure increase rate, supply air temperature control ITAE, supply air temperature residual, log-likelihood...].

[0144] 4. Adaptive Health Calculation under Operating Conditions (corresponding to the Adaptive Health Calculation module under Operating Conditions): Operating Condition Identification: The operating condition of the AHU is mainly defined by the fresh air temperature and the operating mode (cooling / heating / ventilation). This embodiment discretizes the operating condition into several typical ranges, such as "fresh air temperature < 15°C and heating mode" and "15°C ≤ fresh air temperature < 26°C and cooling mode".

[0145] Operating condition decoupling: Since Gaussian process regression has been used to explicitly model the operating conditions (using the operating conditions as input), the model itself already possesses a certain degree of adaptability to operating conditions. To further optimize, one-hot encoding of the operating condition category is added to the feature vector, and a lightweight gradient boosting tree model is input. During training, this model automatically learns the weight differences in the impact of each feature on health under different operating conditions, achieving more refined decoupling.

[0146] Health Regression: The decoupled features are input into another XGBoost regressor, which outputs a health score. The training data for this model can be partly derived from the reverse annotation of historical alarm logs (such as "filter blockage alarm"), where data from a period prior to the alarm is labeled as "sub-healthy".

[0147] 5. Prediction, Early Warning, and Decision Output (corresponding to the Prediction, Early Warning, and Decision Output module): Health sequence: The system calculates health every 15 minutes and forms a sequence.

[0148] RUL Prediction: For critical consumables (such as filters), the decline in their health is usually strongly correlated with time or pressure differential. When the health begins to decline due to filter clogging, the remaining time to reach the filter's pressure differential threshold (such as final resistance) is predicted by fitting the rise curve of the filter's pressure differential using linear or exponential regression. This is the RUL of the filter.

[0149] Explainable alert and optimization report generation: When the health level falls below a threshold, the system triggers a report. Health: 72; Key evidence of degradation: 1. The fan efficiency decreased by 12% compared to the baseline, and the current fluctuation increased, which is suspected to be due to a loose fan belt or poor bearing lubrication, resulting in increased energy consumption for delivering the same amount of air.

[0150] 2. The sensible heat exchange efficiency of the cooling coil decreased by 8%, and the residual error of the supply air temperature control remained positive. It is suspected that scale buildup on the surface of the coil or internal dirt blockage affected the heat exchange effect, causing the chilled water valve to remain at a large opening for a long time and still struggle to control the temperature.

[0151] 3. The filter differential pressure rises faster than expected, and the predicted remaining service life (RUL) is 15 days.

[0152] Maintenance and optimization suggestions: 1. Immediate inspection: Check the tension of the fan belt and the condition of the bearings.

[0153] 2. Scheduled maintenance: Replace the filter within the next two weeks.

[0154] 3. Efficiency optimization: It is recommended to clean the surface cooling coils. After cleaning, the energy efficiency is expected to be restored, and the fan energy consumption is expected to be reduced by about 8%.

[0155] Compared with related technologies, the electromechanical equipment health assessment model and method provided by this invention, through its innovative "phased, interpretable, and knowledge-guided" hybrid modeling framework, brings the following outstanding and verifiable beneficial effects: 1. Significantly reduces reliance on labeled data, effectively solving the problems of "cold start" and "small sample size".

[0156] This invention achieves a key technological breakthrough through an unsupervised health baseline construction module. This module utilizes readily available, unlabeled operational data on the early "health status" of equipment, and automatically learns the inherent patterns of healthy operation through unsupervised / self-supervised methods such as autoencoders or contrastive learning to construct a "health fingerprint" benchmark. This enables the construction of effective evaluation models even when full lifecycle labeled data (especially fault data) is lacking, lowering the data threshold and cycle time for model construction, allowing predictive maintenance to be implemented at the early stages of a project, and solving the "cold start" problem.

[0157] 2. Significantly improves the interpretability of the model, enhancing the trust and decision support capabilities of operations and maintenance personnel.

[0158] This invention, through the design of an interpretable degradation feature extraction and fusion module, overcomes the drawbacks of traditional "black box" models. It innovatively integrates classic physical information features (such as vibration kurtosis and characteristic frequency energy) with depth distance features based on a health baseline (such as reconstruction error). The final output feature vector has clear physical or statistical significance in each key dimension, enabling the model's "health decline" warning to simultaneously provide specific and traceable evidence of degradation. Maintenance engineers can directly obtain clues about fault modes, quickly locate potentially problematic components, and formulate precise maintenance strategies, enhancing their trust in and willingness to use the intelligent assessment system.

[0159] 3. It possesses excellent adaptability to operating conditions, significantly improving the robustness and accuracy of the assessment.

[0160] This invention features a unique attention-based operating condition decoupling network that effectively distinguishes the intrinsic causes of data changes. This network automatically learns and "strips away" normal data fluctuations caused by changes in equipment load, speed, and other operating conditions, "focusing" on the characteristic components reflecting equipment performance degradation. This makes the final calculated health score a "pure" metric that filters out operating condition interference and reflects the essential degradation state of the equipment. Therefore, in real-world operating environments with multiple operating conditions and varying loads, this model can significantly reduce false alarms and false negatives, maintain the stability and accuracy of evaluation results, address the core pain point of existing models where performance deteriorates sharply due to changes in operating conditions, and improve the model's practicality.

[0161] 4. Enhanced model generalization and transfer capabilities, facilitating large-scale deployment.

[0162] This invention features a phased and modular design with good transferability. The core unsupervised health baseline construction module and the decoupled network architecture are universal. When applied to new equipment of the same type, only a small amount of health data from that equipment needs to be collected, and the health baseline module can be retrained or fine-tuned; the core algorithm and model architecture are reusable. This avoids designing and training entirely new and complex end-to-end models for each device from scratch, greatly shortening the model deployment cycle, reducing implementation and maintenance costs, and providing a feasible technical path for the rapid and large-scale predictive maintenance deployment of massive amounts of similar equipment in public buildings and factories.

[0163] 5. It has achieved a complete closed loop from "condition assessment" to "predictive maintenance", resulting in significant overall benefits.

[0164] This invention not only outputs a single health score, but also integrates remaining service life prediction and automatic generation of interpretable reports through a predictive warning and decision-making output module. The system can extrapolate predictions based on health trends and provide multi-level warnings including descriptions of key degradation characteristics. This upgrades maintenance decision-making from "manual analysis based on model output" to "system-assisted automated decision suggestions," helping users move from a perceptual level of "what happened" to a cognitive and predictive level of "why it happened" and "when failure might occur." Ultimately, through earlier, more accurate, and more actionable warnings, it guides users to develop optimal maintenance plans, minimize unplanned downtime, avoid catastrophic failures, optimize spare parts inventory, and thus achieve a systematic reduction in total operating costs, demonstrating the core value of predictive maintenance.

[0165] Key points and protections of this application: 1. A dynamic health baseline construction method based on unsupervised learning: The core of this method is to learn the "health fingerprint" of a device using only unlabeled data on its health status through specific unsupervised models (such as stacked sparse autoencoders or temporal contrastive learning networks), and then calculate the "reconstruction error" and "representation distance" as core degradation indicators. This solves the data dependency problem.

[0166] 2. An interpretable degradation feature generation mechanism: This mechanism fuses and filters time / frequency / time-frequency domain features with clear physical meaning with deep distance features based on a healthy baseline, forming feature vectors that combine data-driven capabilities and physical interpretability. This solves the "black box" problem.

[0167] 3. Attention-based working condition decoupling network structure: This is a specific neural network architecture that takes the original degenerate features and working condition labels as input, and dynamically generates weights through an internal attention network to suppress working condition-related variations in the features, thereby outputting "working condition-invariant degenerate features". This solves the robustness problem.

[0168] 4. A phased, explainable, and knowledge-guided overall system framework: This framework connects unsupervised baseline construction, explainable feature extraction, operational condition decoupling, and health regression into a complete solution in a specific order and logic. This framework itself is a key systemic invention.

[0169] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0170] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0171] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0172] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0173] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0174] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0175] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0176] This application provides a method, system, electronic device, storage medium, and program product for health assessment of electromechanical equipment based on unsupervised baseline learning and decoupling of operating conditions. By utilizing unlabeled data on equipment health status, and through interpretable feature extraction and a decoupled network that adapts to operating conditions, it achieves accurate assessment of the health status of electromechanical equipment, traceable analysis of degradation causes, and prediction of remaining service life. Ultimately, it generates an early warning report with clear maintenance recommendations, reducing the dependence of predictive maintenance on fault data and lowering the implementation threshold, and supporting the rapid deployment of large-scale equipment groups.

[0177] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0178] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0179] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0180] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0181] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0182] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0183] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0184] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0185] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0186] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0187] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for health assessment of electromechanical equipment based on unsupervised baseline learning and operating condition decoupling, characterized in that, The method includes the following steps: The multi-source sensor data and operating condition label data at the current moment are collected and preprocessed to form a standardized multivariate time series data block; Acquire unlabeled historical operating data of electromechanical equipment in its health status, train a multivariate time series representation learning network, construct a health baseline embedding space and statistical baseline, and store them in a health baseline library. Information features are extracted from the standardized multivariate time series data blocks to obtain physical features characterizing the device state; The standardized multivariate time series data block is input into the trained multivariate time series representation learning network, and the reconstruction error and representation distance are calculated using the health baseline library to obtain deep features; The physical features and the deep features are fused together, and a feature selection algorithm is used to select the feature subset with the highest correlation to degradation, so as to obtain an interpretable degradation feature vector. The interpretable degradation feature vector and the working condition label data are input into an attention-based working condition decoupling network to obtain working condition-invariant degradation features. The degradation characteristics under unchanged operating conditions are input into the regressor to calculate the health score, and a health warning report for electromechanical equipment is obtained based on the health score.

2. The method according to claim 1, characterized in that, The process of collecting and preprocessing multi-source sensor data and operational condition label data at the current moment to form standardized multivariate time series data blocks includes: Data is collected synchronously from vibration sensors, temperature sensors, current sensors, voltage sensors, and equipment controllers to obtain multi-source sensor data at the current moment. Extract operating condition tag data from the manufacturing execution system or building equipment monitoring system; the operating condition tag data includes speed setpoint, load current, and operating mode code; The multi-source sensor data and the working condition tag data with different sampling frequencies are synchronized using timestamp alignment and resampling techniques to form data slices; Outliers are removed from the data slices using statistical methods or methods based on neighboring data, and missing values ​​are filled by linear interpolation or forward imputation to obtain standardized multivariate time series data blocks.

3. The method according to claim 1, characterized in that, The process of acquiring unlabeled historical operating data of the health status of electromechanical equipment, training a multivariate time-series representation learning network, constructing a health baseline embedding space and statistical baseline, and storing it in a health baseline library includes: Acquire untagged historical operating data of electromechanical equipment during the initial stable operation period after installation and commissioning, and which are determined to be in a healthy state by preset rules; Using the unlabeled historical running data as training samples, a multivariate temporal representation learning network is trained; the multivariate temporal representation learning network is a stacked sparse autoencoder or a temporal contrastive learning network. The low-dimensional vector output by the multivariate temporal representation learning network is used as the health baseline embedding space, and the mean and variance upper and lower bounds of each channel of the health data are calculated as statistical baselines. The trained network model parameters and statistical baselines are stored together in the health baseline library. The health baseline embedding space is used as a low-dimensional benchmark reference system to represent the data distribution pattern of the device in a healthy state, and is used to quantify the device performance degradation by calculating the degree of deviation between the current data and the health baseline embedding space.

4. The method according to claim 1, characterized in that, The step of extracting information features from the standardized multivariate time series data block to obtain physical features characterizing the device state includes: The root mean square value, peak value, kurtosis, waveform factor, and impulse factor of the signal are calculated from the standardized multivariate time series data block to obtain the time domain characteristics; The signals in the standardized multivariate time series data block are subjected to Fast Fourier Transform to calculate the specific fault characteristic frequencies and their harmonic energy proportions that meet the preset conditions, thereby obtaining frequency domain characteristics; the specific fault characteristic frequencies include the bearing's passing frequency and the gear's meshing frequency. The wavelet packet energy spectrum is calculated for the non-stationary signal in the standardized multivariate time series data block, and the changes in energy distribution in different frequency bands are analyzed to obtain time-frequency domain characteristics. The time-domain features, the frequency-domain features, and the time-frequency-domain features are used as physical features characterizing the device state.

5. The method according to claim 1, characterized in that, The process involves inputting the standardized multivariate time series data block into the trained multivariate time series representation learning network, calculating the reconstruction error and representation distance using the health baseline library, and obtaining deep features, including: The standardized multivariate time series data block is input into the trained multivariate time series representation learning network, and the activation values ​​of the intermediate or bottleneck layers of the network are obtained as deep representations. Using the health baseline embedding space in the health baseline library, calculate the reconstruction error between the reconstruction output and the original input after the standardized multivariate time series data block is input into the multivariate time series representation learning network, and the representation distance between the standardized multivariate time series data block and the health baseline embedding space; The depth representation, the reconstruction error, and the representation distance are used as depth features.

6. The method according to claim 1, characterized in that, The process involves fusing the physical features with the deep features, and using a feature selection algorithm to filter out the feature subset most correlated with degradation, resulting in an interpretable degradation feature vector, including: The physical features and the depth features are concatenated to form a high-dimensional hybrid feature vector; A feature selection algorithm based on mutual information or LASSO regression is used to select the feature subset that is most correlated with device degradation from the high-dimensional mixed feature vector; The aforementioned feature subset is used as an interpretable degenerate feature vector.

7. The method according to claim 1, characterized in that, The step of inputting the interpretable degradation feature vector and the working condition label data into an attention-based working condition decoupling network to obtain working condition-invariant degradation features includes: The interpretable degradation feature vector and the working condition label data are input together into the working condition decoupling network based on the attention mechanism; The working condition decoupling network learns a set of attention weights and dynamically adjusts the weights of feature dimensions in the interpretable degradation feature vector that are below a preset sensitivity value for the current working condition but above a preset sensitivity value for degradation to a preset range. At the same time, it reduces the weights of feature dimensions that change drastically with the working condition but are unrelated to degradation to a preset range. The operating condition decoupling network outputs a pure degradation feature that is independent of the operating condition, which is then used as the degradation feature that remains unchanged by the operating condition.

8. The method according to claim 1, characterized in that, The step of inputting the degradation characteristics of the unchanged operating conditions into the regressor to calculate the health score, and obtaining the electromechanical equipment health early warning report based on the health score, includes: The degradation characteristics under unchanged operating conditions are input into a regressor to obtain a health score. The regressor performs forward computation using a fully connected neural network or an XGBoost model and outputs the health score in the range of 0 to 100, where 100 represents brand new and 0 represents complete failure. The regressor is trained using historical data with the goal of minimizing prediction error. The health score is used to reflect the degree of degradation of the device's current performance relative to a preset health baseline. The multiple health scores obtained through continuous calculation are arranged in chronological order to form a health time series. The health time series is fitted with exponential smoothing or linear regression to fit the trend, a failure threshold is set, and the time required for the health score to drop to the failure threshold is calculated by extrapolation based on the fitted curve. The time is used as the predicted value of the remaining service life. When the health score is lower than a preset threshold, a health warning report for electromechanical equipment is generated, which includes the current health score and the predicted value of the remaining service life.

9. A health assessment system for electromechanical equipment based on unsupervised baseline learning and operating condition decoupling, used to implement the method as described in any one of claims 1 to 8, characterized in that, The system includes: The multi-source heterogeneous data acquisition and preprocessing module is used to acquire and preprocess multi-source sensor data and operating condition label data at the current moment to form standardized multivariate time series data blocks. The unsupervised health baseline construction module is used to acquire unlabeled historical operating data of the health status of electromechanical equipment, train a multivariate time series representation learning network, construct a health baseline embedding space and statistical baseline, and store them in the health baseline library. An interpretable degradation feature extraction and fusion module is used to extract information features from the standardized multivariate time series data block to obtain physical features representing the device state; input the standardized multivariate time series data block into the trained multivariate time series representation learning network, calculate the reconstruction error and representation distance using the health baseline library to obtain deep features; fuse the physical features and the deep features, and use a feature selection algorithm to select the feature subset with the highest correlation to degradation to obtain an interpretable degradation feature vector; The working condition adaptive health calculation module is used to input the interpretable degradation feature vector and the working condition label data into an attention mechanism-based working condition decoupling network to obtain working condition invariant degradation features. The prediction, early warning, and decision output module is used to input the degradation characteristics of the unchanged operating conditions into the regressor to calculate the health score, and to obtain a health early warning report for electromechanical equipment based on the health score.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 8.

Images