Industrial equipment fault prediction and health management method based on multi-sensor fusion

Abstract

The present invention belongs to the field of equipment management technology. The present invention discloses an industrial equipment fault prediction and health management method based on multi-sensor fusion, including: acquiring multi-source sensor data of industrial equipment, performing signal decoupling analysis, and obtaining a decoupling characteristic spectrum; performing frequency domain conversion and modulation analysis to form a multi-dimensional characteristic spectrum; analyzing the modal correlation of the multi-dimensional characteristic spectrum to obtain a fault feature mapping network; constructing a hybrid time series prediction model, performing remaining life prediction and degradation trend assessment, and obtaining an equipment health trend map; establishing a health status evaluation index system, performing reliability assessment, and obtaining an equipment health status report; generating maintenance decision recommendations, and realizing real-time anomaly detection and maintenance recommendation push through edge computing. The present invention achieves early warning and accurate prediction of industrial equipment failures through multi-sensor data fusion and advanced analysis technology, significantly improving the operational reliability and production efficiency of industrial equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment management, and more specifically, to a method for industrial equipment fault prediction and health management based on multi-sensor fusion. Background Art

[0002] With the deepening advancement of industrial concepts and the rapid development of intelligent manufacturing technologies, industrial equipment health management systems have become a key support for ensuring stable production line operations. Current equipment fault prediction and health management systems primarily monitor parameters such as vibration, temperature, and sound to assess equipment operating status and predict potential failures. However, these systems still have significant shortcomings in the early identification and characterization of small bearing defects in high-frequency vibration environments.

[0003] Existing fault prediction systems often overlook the coupling relationship between the nonlinear harmonic signals generated by the initial stages of bearing microwear and equipment load fluctuations. This coupling phenomenon can cause fault signatures to be masked or distorted in the spectrum. Initial bearing microwear produces weak and unstable vibration signatures, which are modulated by periodic fluctuations in the equipment load. The lack of real-time decoupling analysis of the coupling effect between bearing microwear and load fluctuations makes it impossible to accurately capture the critical stages of early bearing degradation in actual industrial production, resulting in delayed fault warnings and the inability to effectively implement preventive maintenance strategies. Particularly in harsh working environments such as metallurgy and mining, the combined interference of bearing operating temperature and ambient vibration further reduces the identifiability of fault signatures. Existing multi-sensor fusion systems fail to account for the non-stationary and nonlinear characteristics of sensor signals under different operating conditions. Simple data fusion methods often amplify rather than suppress the interference of environmental noise, resulting in poor feature extraction. The limitations of this signal processing method have led to a significant decrease in the system's accuracy in identifying minor bearing defects under complex load conditions, shortening the average advance warning time by more than 30%, seriously affecting the timeliness and effectiveness of equipment health management decisions. In addition, different types of bearing materials have significant differences in their acoustic emission characteristics during the wear process. The current fault prediction platform lacks the ability to adaptively analyze the spectral characteristics of acoustic emission signals of bearings made of different materials, resulting in insufficient universality of the fault pattern recognition algorithm and an inability to cope with the signal characteristic variations caused by material diversity, which in turn affects the accuracy of health status assessment and the robustness of the prediction model.

[0004] In view of this, the present invention proposes an industrial equipment fault prediction and health management method based on multi-sensor fusion to solve the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned objectives, the present invention provides the following technical solution: a method for industrial equipment fault prediction and health management based on multi-sensor fusion, comprising:

[0006] Step S1: Acquire multi-source sensor data of industrial equipment, perform signal decoupling analysis on the multi-source sensor data, and obtain a decoupling characteristic spectrum;

[0007] Step S2: performing frequency domain conversion and modulation analysis on the decoupled characteristic spectrum to form a multi-dimensional characteristic spectrum, analyzing the modal correlation of the multi-dimensional characteristic spectrum, and obtaining a fault feature mapping network;

[0008] Step S3: Construct a hybrid time series prediction model, and perform remaining life prediction and degradation trend assessment on the fault feature mapping network based on the hybrid time series prediction model, thereby obtaining a device health trend map;

[0009] Step S4: Establish a health status evaluation indicator system, conduct reliability assessment on the equipment health trend map, and obtain an equipment health status report;

[0010] Step S5: Generate maintenance decision recommendations based on the equipment health status report, and implement real-time anomaly detection and maintenance recommendation push through edge computing.

[0011] The technical effects and advantages of the industrial equipment fault prediction and health management method based on multi-sensor fusion of the present invention are as follows:

[0012] The present invention uses frequency domain conversion and modulation analysis to form a multidimensional feature spectrum and fault feature mapping network, which can comprehensively capture the subtle change characteristics of the equipment in different operating stages, effectively identify early fault signs, and significantly improve the sensitivity and accuracy of fault diagnosis. By constructing a hybrid time series prediction model, it combines the advantages of a bidirectional long-short-term memory network and a self-attention mechanism, and can accurately predict the remaining service life and degradation trend of the equipment, providing a scientific basis for maintenance decision-making, and effectively avoiding the unexpected downtime of the production line and economic losses caused by the traditional passive maintenance model. By establishing a multidimensional health evaluation indicator system and combining multi-source data fusion with DS evidence theory, it can objectively evaluate the health status of the equipment, cope with complex failure modes under different operating conditions, and improve the robustness and adaptability of equipment health management. Edge computing technology is used to achieve real-time anomaly detection and maintenance recommendation push, which can provide timely warnings and provide accurate maintenance decision recommendations when the equipment is in an abnormal state, transforming equipment maintenance from traditional passive response to active prevention, significantly reducing maintenance costs and improving maintenance efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 Schematic diagram of the industrial equipment fault prediction and health management method based on multi-sensor fusion of the present invention;

[0014] Figure 2 Schematic diagram of the industrial equipment fault prediction and health management system based on multi-sensor fusion of the present invention. DETAILED DESCRIPTION

[0015] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0016] Example 1

[0017] See also Figure 1 As shown, the industrial equipment fault prediction and health management method based on multi-sensor fusion in this embodiment includes:

[0018] Step S1: Acquire multi-source sensor data of industrial equipment, perform signal decoupling analysis on the multi-source sensor data, and obtain a decoupling characteristic spectrum;

[0019] Step S2: performing frequency domain conversion and modulation analysis on the decoupled characteristic spectrum to form a multi-dimensional characteristic spectrum, analyzing the modal correlation of the multi-dimensional characteristic spectrum, and obtaining a fault feature mapping network;

[0020] Step S3: Construct a hybrid time series prediction model, and perform remaining life prediction and degradation trend assessment on the fault feature mapping network based on the hybrid time series prediction model, thereby obtaining a device health trend map;

[0021] Step S4: Establish a health status evaluation indicator system, conduct reliability assessment on the equipment health trend map, and obtain an equipment health status report;

[0022] Step S5: Generate maintenance decision recommendations based on the equipment health status report, and implement real-time anomaly detection and maintenance recommendation push through edge computing.

[0023] Preferably, the process of executing step S1 may specifically include the following steps:

[0024] Acquire multi-source sensor data of industrial equipment, including vibration signals, temperature signals, current signals, and acoustic emission signals; perform signal decomposition on the vibration signal to obtain vibration signal characteristics; perform time-series segmentation processing on the temperature signal to obtain temperature gradient change characteristics; perform envelope demodulation on the current signal to obtain current spectrum characteristics; perform spectrum analysis on the acoustic emission signal to obtain acoustic emission energy distribution characteristics; modulate and analyze the vibration signal characteristics and current spectrum characteristics to obtain load fluctuation signals; perform adaptive separation processing based on the vibration signal characteristics and load fluctuation signals to obtain intrinsic vibration signals; perform cross-validation based on intrinsic vibration signals, temperature gradient change characteristics, current spectrum characteristics, and acoustic emission energy distribution characteristics to construct a decoupled characteristic spectrum.

[0025] Specifically, the system first acquires multi-source sensor data from industrial equipment, including vibration, temperature, current, and acoustic emission signals. Vibration signals are collected using accelerometers installed in key equipment locations (such as bearings and gearboxes), with a sampling frequency set at 20kHz to capture high-frequency fault signatures. Temperature signals are collected using thermocouples or infrared temperature sensors at a sampling frequency of 1Hz to monitor temperature rise during equipment operation. Current signals are collected using Hall effect current sensors at a sampling frequency of 10kHz to monitor motor load changes. Acoustic emission signals are collected using piezoelectric acoustic emission sensors with a frequency range of 100kHz to 1MHz to detect microcracks and material damage.

[0026] Decompose the vibration signal to obtain its characteristics. This involves selecting a wavelet basis function suitable for bearing fault characteristics, such as the Symlet wavelet (sym8), and performing wavelet packet decomposition on the vibration signal to obtain sub-signals in different frequency bands. Typically, the signal is decomposed into five layers, generating 32 sub-bands. The energy value of each sub-signal in each frequency band is calculated to form a frequency band energy distribution diagram. The energy calculation formula is: ,in Indicates the The energy of the sub-band, Represents the signal of the corresponding sub-band. Based on the frequency band energy distribution diagram, the information entropy criterion is used to select the key frequency band containing fault information. M frequency bands with energy values ​​greater than a preset energy threshold and information entropy greater than a preset entropy value are selected. The selected key frequency band signals are reconstructed to obtain the vibration sub-band signals. The time domain statistical characteristics of the reconstructed vibration sub-band signals are calculated, including the kurtosis value (reflecting the signal's impact), the skewness value (reflecting the signal's asymmetry), and the root mean square value (reflecting the signal's energy level), to form the vibration signal feature vector.

[0027] The temperature signal is processed in time-series segments to obtain temperature gradient variation characteristics. First, the temperature signal is smoothed and filtered to eliminate transient noise interference. Then, an adaptive segmentation algorithm is used to divide the temperature curve into multiple characteristic segments, each representing the temperature variation pattern of the equipment under different operating conditions. The temperature change rate and temperature fluctuation amplitude are calculated for each characteristic segment. Finally, temperature anomalies and mutation points are extracted based on preset thresholds to form a temperature gradient variation characteristic. The temperature gradient variation characteristic is important for detecting equipment friction anomalies, lubrication status changes, and overload conditions.

[0028] The current signal is envelope demodulated to obtain the current spectrum characteristics. First, the current signal is bandpass filtered to retain the frequency band containing the fault characteristics. The filtered signal is then Hilbert transformed to obtain the analytical signal. The amplitude of the analytical signal is calculated to obtain the current signal envelope. The envelope signal is Fourier transformed to obtain the envelope spectrum. The characteristic frequency components and their amplitudes are extracted from the envelope spectrum to form the current spectrum characteristics. The current spectrum characteristics can reflect motor load fluctuations and electrical faults, and are particularly sensitive to rotor faults and power quality issues.

[0029] Spectral analysis of the acoustic emission signal is performed to obtain the acoustic emission energy distribution characteristics. First, the acoustic emission signal is preprocessed, including noise removal and signal enhancement. Then, the time domain parameters of the acoustic emission signal, such as amplitude distribution and energy accumulation, are calculated. Short-time Fourier transform is used to perform time-frequency analysis on the acoustic emission signal to obtain the spectral characteristics of the acoustic emission signal. Statistical analysis of the spectral characteristics is performed to calculate the energy distribution and peak frequency of each frequency band, forming the acoustic emission energy distribution characteristics. The acoustic emission energy distribution characteristics are highly sensitive to early material damage and microcracks and can serve as an important indicator for early warning of faults.

[0030] The vibration signal characteristics and current spectrum characteristics are modulated and analyzed to obtain the load fluctuation signal. Based on the physical model of the equipment, the modulation relationship between the vibration signal and the current signal is analyzed. Demodulation techniques are used to extract the modulation frequency components of the two types of signals. Correlation analysis is used to determine the common modulation frequencies, which are often related to the equipment's workload fluctuations. The load fluctuation signal is reconstructed based on the modulation frequencies. This signal reflects the load changes during equipment operation.

[0031] Adaptive separation processing is performed based on the vibration signal characteristics and the load fluctuation signal to obtain the intrinsic vibration signal. An adaptive filter is constructed with the load fluctuation signal as the reference input and the vibration signal characteristics as the desired output. The filter parameters are optimized using the minimum mean square error criterion to separate the load-related vibration components. The separated signal is then differentiated from the original vibration signal to obtain the intrinsic vibration signal, which is unaffected by the load. The intrinsic vibration signal primarily reflects the fault characteristics of the equipment itself, eliminating interference caused by load fluctuations.

[0032] A decoupled signature spectrum was constructed through cross-validation based on intrinsic vibration signals, temperature gradient characteristics, current spectrum characteristics, and acoustic emission energy distribution characteristics. A data fusion algorithm was used to perform correlation analysis and consistency testing on the four features. Noise and outliers were removed, and feature selection and dimensionality reduction were performed on each feature. Finally, a decoupled signature spectrum was constructed that incorporates multi-source information and is mutually independent. This decoupled signature spectrum provides a reliable feature foundation for subsequent fault diagnosis and health assessment.

[0033] Preferably, the process of executing step S2 may specifically include the following steps:

[0034] Perform Hilbert-Huang transform on the decoupled characteristic spectrum to obtain the time-frequency domain feature representation; based on the time-frequency domain feature representation, extract the instantaneous frequency and instantaneous amplitude features to form the instantaneous characteristic spectrum; perform frequency domain conversion on the instantaneous characteristic spectrum to obtain a spectrum feature graph; perform fault mode recognition based on the spectrum feature graph and establish a fault feature library; perform tensor decomposition on the vibration signal features, temperature gradient change features, current spectrum features and acoustic emission energy distribution features to obtain a cross-modal feature representation; based on the cross-modal feature representation, calculate the correlation matrix between each mode and construct a modal correlation network; perform key feature screening based on the modal correlation network to obtain a fault diagnosis feature set; perform feature matching based on the fault diagnosis feature set and the fault feature library to construct a fault feature mapping network.

[0035] Specifically, the decoupled characteristic spectrum is subjected to a Hilbert-Huang transform to obtain a time-frequency domain representation. The Hilbert-Huang transform is an advanced time-frequency analysis method that combines empirical mode decomposition and the Hilbert transform. First, the decoupled characteristic spectrum is decomposed into multiple intrinsic mode functions (IMFs) using empirical mode decomposition. Each IMF represents a natural oscillation mode in the signal. Each IMF is then subjected to a Hilbert transform to obtain its analytical signal. Finally, the instantaneous frequency and instantaneous amplitude of each IMF are calculated to form the Hilbert spectrum, which forms the time-frequency domain representation of the signal. The Hilbert-Huang transform effectively processes nonlinear and nonstationary signals and is suitable for fault feature extraction under complex operating conditions of industrial equipment. Based on the time-frequency domain representation, the instantaneous frequency and instantaneous amplitude features are extracted to form the instantaneous characteristic spectrum. For each IMF, the statistical characteristics of the instantaneous frequency, including the mean, standard deviation, maximum, and minimum value, are calculated. The distribution characteristics of the instantaneous amplitude, including the crest factor, crest factor, and impulse factor, are also extracted. These time-varying features are combined into an instantaneous characteristic spectrum to characterize the dynamic changes in the equipment status over time. The instantaneous characteristic spectrum can capture the transient behavior and non-stationary characteristics of the equipment, which is of great value for early fault diagnosis.

[0036] The instantaneous characteristic spectrum is transformed into the frequency domain to obtain a spectrum feature map. Fast Fourier transform is used to transform the instantaneous characteristic spectrum into the frequency domain to obtain the initial spectrum. Second-order spectrum analysis is performed on the initial spectrum to calculate the bispectrum, which can reveal the second-order nonlinear characteristics and phase coupling phenomena in the signal. An adaptive threshold segmentation algorithm is used on the bispectrum to extract significant characteristic frequency points, which are related to the equipment's fault characteristic frequency. Finally, a spectrum feature map is constructed based on the extracted significant characteristic frequency points to intuitively display the distribution and intensity of the fault characteristic frequency.

[0037] Fault pattern recognition is performed based on spectral signature graphs, and a fault signature library is established. Characteristic spectrum samples of different fault types (such as bearing inner race faults, outer race faults, rolling element faults, and gear tooth breakage) are collected and annotated. The characteristic frequency patterns and amplitude distribution characteristics of each fault are extracted. A fault signature library is constructed containing various fault signature templates. Each template includes information such as the fault type, characteristic frequency, amplitude ratio, and confidence interval. The fault signature library provides a reference standard for subsequent fault identification and matching.

[0038] The vibration signal features, temperature gradient change features, current spectrum features, and acoustic emission energy distribution features are subjected to tensor decomposition to obtain a cross-modal feature representation. The four types of features are organized into a third-order tensor, with the three dimensions representing feature type, feature parameter, and time series, respectively. Using tensor decomposition methods such as Tucker decomposition or CP decomposition, the high-dimensional tensor is decomposed into a core tensor and a factor matrix. The main characteristic patterns of each mode are extracted through the factor matrix, and the interaction between modes is expressed through the core tensor. Finally, a cross-modal feature representation with reduced dimensionality and preserved inter-modal correlation is obtained. Tensor decomposition can effectively handle the fusion of multimodal data and preserve the structural relationships between different sensing modes.

[0039] Based on cross-modal feature representation, the correlation matrix between modalities is calculated, and a modal correlation network is constructed. The strength of the intermodal association is quantified by calculating the mutual correlation coefficients between different modal features. The results are organized into a correlation matrix, with the matrix elements representing the degree of correlation between corresponding modal pairs. A weighted undirected graph is constructed based on the correlation matrix, with nodes representing different sensing modalities and edge weights representing the strength of the intermodal correlation. A community discovery algorithm is used to identify highly correlated modal groups, each of which represents a modal correlation network. The modal correlation network reveals the inherent connections between different sensor signals, providing a basis for multi-source information fusion.

[0040] Key features are screened based on the modal correlation network to obtain a fault diagnosis feature set. Based on the modal correlation network, centrality metrics for each feature, including degree centrality, are calculated. Features with degree centrality greater than a preset selection threshold are selected as key features, as these have strong diagnostic capabilities. Feature redundancy is also considered, and highly correlated redundant features are removed. Ultimately, a streamlined and effective fault diagnosis feature set is formed. This fault diagnosis feature set includes the most critical multimodal features for fault identification, improving diagnostic accuracy and efficiency.

[0041] Based on feature matching between the fault diagnosis feature set and the fault feature library, a fault feature mapping network is constructed. Similarity is calculated between the fault diagnosis feature set and the templates in the fault feature library using metrics such as Euclidean distance, cosine similarity, or Mahalanobis distance. Based on the similarity results, a mapping relationship between features and fault types is established, ultimately forming a fault feature mapping network. This network describes the probabilistic mapping relationship from multimodal features to various fault types. The fault feature mapping network provides the knowledge foundation for subsequent health assessment and prediction.

[0042] Preferably, the process of executing step S3 may specifically include the following steps:

[0043] A bidirectional long-short-term memory network model is constructed to capture the long-term dependency characteristics of equipment degradation; a self-attention mechanism model is constructed to identify key timing nodes in the degradation process; the bidirectional long-short-term memory network model is integrated with the self-attention mechanism model to construct an initial timing prediction model; the initial timing prediction model is trained using historical fault data to obtain a hybrid timing prediction model; a sliding time window method is used to extract features from the fault feature mapping network, and the extracted features are input into the hybrid timing prediction model to obtain the remaining life prediction results of the equipment; based on the remaining life prediction results, a degradation trend curve is constructed; the degradation trend curve is compared and analyzed with the preset health benchmark to obtain an equipment health trend map.

[0044] Specifically, a bidirectional long short-term memory (LSTM) network model was constructed to capture the long-term dependencies of device degradation. This model consists of two LSTM layers, one forward and one backward, capable of simultaneously considering both historical and future information, making it suitable for handling the temporal dependencies of device degradation. The LSTM unit incorporates three gating mechanisms: an input gate, a forget gate, and an output gate. These gates selectively memorize and forget information, effectively addressing the vanishing gradient problem of traditional RNNs. The input of the bidirectional LSTM network model is a temporal feature sequence, and its output is a hidden state vector that encodes the long-term dependencies of the device degradation process. The model structure includes an input layer (with the same dimension as the number of features), two Bi-LSTM layers (64 units each), and a fully connected layer. Dropout (ratio 0.2) is used to prevent overfitting.

[0045] A self-attention mechanism model is constructed to identify key time points in the degradation process. The self-attention mechanism calculates the correlation between elements within the sequence and assigns different weights to each time point, highlighting the characteristic information of key moments. The self-attention module first converts the input sequence into three matrices: query, key, and value. It then calculates the dot product of the query and key to obtain a similarity matrix. After softmax normalization, it obtains attention weights. Finally, the weighted sum of the value matrix is ​​used to obtain a weighted feature representation. The self-attention mechanism can adaptively focus on anomalies and turning points in the degradation process, improving the model's ability to perceive key events.

[0046] The initial time series prediction model is constructed by fusing a bidirectional long short-term memory network model with a self-attention mechanism model. The specific architecture is as follows: the input feature sequence is first processed through a Bi-LSTM layer to obtain an encoded hidden state sequence; the hidden state sequence is then input into the self-attention module, where attention weights are calculated and a weighted feature representation is obtained; the weighted feature representation is then input into a fully connected layer, which outputs the prediction result. This fusion of the two models leverages the advantages of the Bi-LSTM in capturing long-term dependencies and the self-attention mechanism in identifying critical moments, improving the model's ability to model complex degradation processes.

[0047] The initial time series prediction model was trained using historical failure data to obtain a hybrid time series prediction model. A complete data set of the equipment's historical operation until failure was collected, including various sensor signals and corresponding remaining life labels. The data set was divided into training and validation sets in a certain ratio (8:2). Mean squared error was selected as the loss function, and the Adam optimizer was used for parameter optimization with a learning rate of 0.001 and a batch size of 32. Training was terminated when the validation set loss function value did not decrease for 10 consecutive cycles. The model parameters with the best performance were selected as the final hybrid time series prediction model. The trained model accurately captures the degradation patterns of the equipment and provides a reliable basis for remaining life prediction.

[0048] A sliding time window method is used to extract features from the fault feature mapping network. The extracted features are then fed into a hybrid time series prediction model to obtain the remaining life prediction results for the equipment. The sliding window length is set to L (e.g., 30 time points) and the step size is set to S (e.g., 5 time points). A sliding window is applied to the time series data of the fault feature mapping network to extract the feature sequence within each window. The extracted feature sequence is then fed into the trained hybrid time series prediction model to obtain the corresponding remaining life prediction value. As the window slides, the prediction results are continuously updated, forming a dynamic remaining life prediction curve. This sliding window method enables continuous monitoring and prediction of equipment status, adapting to real-time changes in equipment operating conditions.

[0049] Based on the remaining life prediction results, a degradation trend curve is constructed. The predicted remaining life is plotted as a trend curve over time. The predicted curve is smoothed using exponential smoothing or local regression methods to eliminate the interference of short-term fluctuations. The degradation trend curve is compared and analyzed with the preset health baseline to obtain an equipment health trend map. Based on expert experience and historical data, a health baseline for normal equipment operation is set. The deviation between the actual degradation trend and the health baseline is calculated to quantify the degree of abnormality in the current state. The health level is divided according to the degree of deviation, including "healthy", "mild degradation", "moderate degradation", "severe degradation" and "imminent failure". The health level, degradation trend and remaining life prediction are combined to generate an equipment health trend map. The health trend map comprehensively reflects the current health status of the equipment and future development trends, providing visual support for subsequent health assessment and maintenance decisions.

[0050] Preferably, the process of executing step S4 may specifically include the following steps:

[0051] Establish a multi-dimensional health evaluation index system, including vibration intensity index, temperature anomaly index, current fluctuation index and acoustic emission energy index; quantify the equipment health trend map based on the multi-dimensional health evaluation index system to obtain equipment health indicators; normalize various equipment health indicators to obtain standardized health indicators; perform multi-source fusion of standardized health indicators to obtain a comprehensive health score; divide equipment health levels based on the comprehensive health score; perform index evaluation on standardized health indicators to obtain reliability indicators; generate an equipment health status report based on the equipment health level and reliability indicators.

[0052] Specifically, a multidimensional health evaluation index system was established to comprehensively assess the health status of the equipment. Vibration intensity indicators include the root mean square value, peak factor, and frequency band energy ratio, reflecting the vibration intensity and impact resistance of mechanical components. Temperature anomaly indicators include the temperature deviation rate, temperature rise rate, and temperature fluctuation coefficient, reflecting thermal and friction anomalies. Current fluctuation indicators include the root mean square current deviation, harmonic distortion, and current imbalance, reflecting electrical performance and load changes. Acoustic emission energy indicators include cumulative energy, event rate, and frequency center of gravity, reflecting material damage and microcrack development. This multidimensional health evaluation index system describes the health characteristics of the equipment from different perspectives, ensuring the comprehensiveness and reliability of the assessment.

[0053] Based on a multidimensional health evaluation index system, the equipment health trend map is quantified to obtain equipment health indicators. For each health evaluation indicator, a corresponding calculation method (e.g., vibration intensity uses the root mean square value of vibration) and a threshold range are set. Based on the data in the equipment health trend map, specific values ​​for each indicator are calculated. These values ​​are compared with normal baseline values ​​to quantify the degree of abnormality. The degradation rate is assessed by considering the indicator's changing trend over time. Ultimately, a set of quantitative equipment health indicators is obtained that objectively reflects the current status of the equipment. Equipment health indicators serve as the data foundation for subsequent health ratings and decision-making.

[0054] Normalize each device health indicator to obtain a standardized health indicator. Because different health indicators have different physical meanings and dimensions, standardization is necessary to make them comparable. Use the maximum-minimum normalization method to map each indicator value to the [0,1] interval or a standard normal distribution. Considering the directionality of the indicator, reverse transform the "smaller is better" indicator. Through normalization, standardized health indicators are obtained, facilitating subsequent multi-source fusion and comprehensive evaluation.

[0055] The standardized health indicators are multi-source fused to obtain a comprehensive health score. The DS evidence theory is used for multi-source information fusion. First, a basic probability distribution function is constructed to map each standardized health indicator to evidence support. The evidence support is the support degree of the standardized health indicator to support the device in a healthy state, and the value range is [0,1]. The conflict coefficient of each evidence support is calculated to evaluate the consistency between the information sources. For any two evidence support and , the conflict coefficient is calculated as: ;in, represents the conflict coefficient, Representative evidence support , Representative evidence support The credibility of each piece of evidence support is adjusted based on the conflict coefficient. For evidence support with a conflict coefficient greater than a preset conflict threshold, the weight of the corresponding health indicator is reduced by a fixed amount. The adjusted evidence support is fused using the DS combination rule, and the fused belief function and likelihood function are calculated. A weighted method is used to calculate the comprehensive score of the device health status based on the belief function and likelihood function. The score ranges from 0 to 100, with higher scores indicating healthier devices. DS evidence theory fusion can effectively handle the uncertainty and complementarity of multi-source information, improving the reliability of assessment results.

[0056] Based on a comprehensive health score, equipment health levels are assigned. Health level thresholds are set based on expert experience and historical data. Health status is typically categorized into five levels: excellent (score 90 to 100), good (score 75 to 90), moderate (score 60 to 75), poor (score 40 to 60), and critical (score 0 to 40). Different levels correspond to different maintenance strategies and response levels. As the score is dynamically updated, the equipment health level adjusts accordingly, reflecting changing health trends. Health level categorization simplifies complex assessment results, making it easier for maintenance personnel to quickly understand equipment status.

[0057] Standardized health indicators are evaluated to obtain reliability indicators. Based on standardized health indicators and historical failure data, equipment reliability models are established. The Weibull distribution or proportional hazard model is used to describe the probability distribution of equipment failure. The probability of normal equipment operation within a given timeframe, i.e., reliability, is calculated. Mean time between failures (MTBF) and remaining useful life (RSV) are predicted. Reliability indicators, combining reliability, MTBF, and RVS, are then generated to quantify the equipment's ability to operate reliably over a period of time. Reliability indicators provide a probabilistic basis for maintenance decisions and aid in developing risk management strategies.

[0058] Generate an equipment health report based on the equipment's health level and reliability indicators. This report includes basic equipment information, health score and level, detailed status of each health indicator, reliability prediction results, anomaly analysis, and maintenance recommendations. It uses charts to visually display equipment health trends and key indicator changes, identifies potential failure risks and anomalies requiring attention, provides preliminary maintenance recommendations and prioritization, and displays warning signs to clearly indicate equipment with serious anomalies or impending failure. The equipment health report comprehensively summarizes the equipment's health assessment results, providing information support for subsequent maintenance decisions.

[0059] Preferably, the process of executing step S5 may specifically include the following steps:

[0060] Based on the equipment health status report, identify abnormal degradation patterns, match the historical fault case library, and infer possible fault types; perform maintenance assessments based on the fault type and equipment remaining life prediction results, and generate maintenance priorities; formulate maintenance time window recommendations based on the maintenance priority; generate maintenance operation instructions based on the fault type; form maintenance decision recommendations based on the maintenance time window recommendations and maintenance operation instructions; monitor sensor data in real time through edge computing devices, preset health status thresholds, and when sensor data anomalies are detected, trigger the early warning mechanism and push maintenance decision recommendations.

[0061] Specifically, based on the equipment health report, abnormal degradation patterns are identified, matched against a historical fault case library, and the possible fault type is inferred. The abnormal indicators and degradation trends in the health report are analyzed to identify typical degradation pattern characteristics, such as linear degradation, accelerated degradation, or sudden degradation. The identified degradation pattern characteristics are then matched against the historical fault case library for similarity, identifying the historical fault case with the highest similarity. Based on the matching results, the possible fault type and development trend of the current equipment are inferred. For multiple possible fault types, their probability distribution is calculated to form a fault diagnosis result. The historical fault case library contains a large amount of labeled fault data and handling experience, providing a knowledge base for fault inference.

[0062] Maintenance assessments are conducted based on fault type and equipment remaining life prediction results to generate maintenance priorities. The risk level of the fault is assessed based on the severity and impact of the fault type. The urgency of the fault's development is assessed in conjunction with the remaining life prediction results. The criticality of the fault's impact is assessed, taking into account the equipment's importance and substitutability within the production system. A maintenance priority score is calculated, combining risk level, urgency, and criticality. Equipment maintenance needs are ranked based on the score, determining maintenance priorities. This maintenance priority mechanism ensures that limited maintenance resources are prioritized for the most critical and urgent equipment, achieving optimal resource allocation.

[0063] Develop maintenance time window recommendations based on maintenance priorities. Estimate the safe operating time of equipment based on remaining life predictions and degradation trends. Combined with production plans and downtime schedules, identify the optimal maintenance opportunity. Set maintenance time windows, including the earliest, optimal, and latest maintenance times. Adopt different time window strategies for equipment of different priorities, with a conservative strategy for high-priority equipment and a more flexible strategy for low-priority equipment. Generate maintenance time window recommendations to inform maintenance scheduling decisions. A reasonable time window ensures safe equipment operation while minimizing disruption to production.

[0064] Generate maintenance instructions based on the fault type. Based on the inferred fault type, the maintenance knowledge base is queried to extract the corresponding maintenance plan. The maintenance plan includes spare parts requirements, tool preparation, safety measures, operating procedures, and quality inspection. The maintenance instructions are customized based on the specific equipment conditions. For complex faults, a multi-level maintenance plan is provided, including temporary repairs, standard repairs, and complete refurbishment options. For emergency situations, an emergency response guide is provided. Detailed maintenance instructions improve maintenance efficiency and quality, reducing human errors and maintenance time.

[0065] Maintenance decision recommendations are constructed based on maintenance time window recommendations and maintenance operation guidance. These recommendations are integrated into a complete maintenance decision recommendation. The recommendations include maintenance priority, recommended maintenance time, estimated maintenance duration, maintenance type, spare parts requirements, staffing, and skill requirements. Decision information at different levels and with different focuses is provided for maintenance personnel in different positions. Decision recommendations are presented in a concise and intuitive format for quick understanding and implementation. Decision support tools are provided to assist maintenance managers in resource scheduling and planning. Maintenance decision recommendations provide comprehensive guidance for maintenance activities and optimize the utilization efficiency of maintenance resources.

[0066] Edge computing devices monitor sensor data in real time, presetting health thresholds. When sensor data anomalies are detected, early warning mechanisms are triggered and maintenance decision recommendations are pushed. Edge computing nodes are deployed on-site to enable local processing and analysis of sensor data. Health thresholds for various types of sensor data are pre-set, including normal, warning, and danger thresholds. Edge computing nodes continuously monitor sensor data streams and detect anomalies in real time. When data exceeds pre-set thresholds or exhibits abnormal patterns, early warning mechanisms at the corresponding level are triggered. This anomaly information is combined with generated maintenance decision recommendations, and alerts and action suggestions are pushed to maintenance personnel via mobile terminals or control center interfaces. Edge computing enables real-time and autonomous anomaly detection, reduces communication burdens and response delays, and improves emergency response capabilities.

[0067] This embodiment uses frequency domain conversion and modulation analysis to form a multidimensional feature spectrum and fault feature mapping network. This fully captures subtle changes in the characteristics of equipment during different operating stages, effectively identifying early signs of failure and significantly improving the sensitivity and accuracy of fault diagnosis. By constructing a hybrid time series prediction model that combines the advantages of a bidirectional long-short-term memory network and a self-attention mechanism, it can accurately predict the remaining service life and degradation trend of equipment, providing a scientific basis for maintenance decisions and effectively avoiding the unplanned production line downtime and economic losses caused by traditional passive maintenance models. By establishing a multidimensional health evaluation indicator system and integrating multi-source data with DS evidence theory, it can objectively assess the health status of equipment, address complex failure modes under different operating conditions, and improve the robustness and adaptability of equipment health management. Edge computing technology enables real-time anomaly detection and maintenance recommendation push, providing timely warnings and accurate maintenance decision recommendations when equipment abnormalities occur. This shifts equipment maintenance from traditional passive response to proactive prevention, significantly reducing maintenance costs and improving maintenance efficiency.

[0068] Example 2

[0069] See also Figure 2As shown, for parts not described in detail in this embodiment, please refer to the description of Example 1. The industrial equipment fault prediction and health management system based on multi-sensor fusion is provided, including:

[0070] Data decoupling analysis module: acquires multi-source sensor data from industrial equipment, performs signal decoupling analysis on the multi-source sensor data, and obtains decoupling characteristic spectra;

[0071] Feature analysis module: performs frequency domain conversion and modulation analysis on the decoupled feature spectrum to form a multi-dimensional feature spectrum, analyzes the modal correlation of the multi-dimensional feature spectrum, and obtains the fault feature mapping network;

[0072] Life prediction module: Builds a hybrid time series prediction model, performs remaining life prediction and degradation trend assessment on the fault feature mapping network based on the hybrid time series prediction model, and thus obtains a health trend map of the equipment;

[0073] Index evaluation module: Establish a health status evaluation index system, conduct reliability assessment on equipment health trend maps, and obtain equipment health status reports;

[0074] Maintenance decision module: Generates maintenance decision recommendations based on equipment health status reports, and implements real-time anomaly detection and maintenance recommendation push through edge computing.

[0075] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art will be able to modify the technical solutions described in the foregoing embodiments or to substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

[0076] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0077] In the description of the present invention, it should be understood that the terms "first", "second", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0078] In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0079] In the description of the present invention, “several” means one or more, and “a large number” means two or more.

[0080] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0081] The formulas in this manual are all dimensionless and calculated using numerical values. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field based on actual conditions.

[0082] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. The method of industrial equipment fault prediction and health management based on multi-sensor fusion is characterized by: include: Step S1: Acquire multi-source sensor data of industrial equipment, perform signal decoupling analysis on the multi-source sensor data, and obtain a decoupling characteristic spectrum; Step S2: performing frequency domain conversion and modulation analysis on the decoupled characteristic spectrum to form a multi-dimensional characteristic spectrum, analyzing the modal correlation of the multi-dimensional characteristic spectrum, and obtaining a fault feature mapping network; Step S3: Construct a hybrid time series prediction model, and perform remaining life prediction and degradation trend assessment on the fault feature mapping network based on the hybrid time series prediction model, thereby obtaining a device health trend map; Step S4: Establish a health status evaluation indicator system, conduct reliability assessment on the equipment health trend map, and obtain an equipment health status report; Step S5: Generate maintenance decision recommendations based on the equipment health status report, and implement real-time anomaly detection and maintenance recommendation push through edge computing; Step S1 includes: Acquire multi-source sensor data from industrial equipment, including vibration signals, temperature signals, current signals, and acoustic emission signals; perform signal decomposition on vibration signals to obtain vibration signal characteristics; perform time-series segmentation processing on temperature signals to obtain temperature gradient change characteristics; perform envelope demodulation on current signals to obtain current spectrum characteristics; perform spectrum analysis on acoustic emission signals to obtain acoustic emission energy distribution characteristics; modulate and analyze vibration signal characteristics and current spectrum characteristics to obtain load fluctuation signals; perform adaptive separation processing based on vibration signal characteristics and load fluctuation signals to obtain intrinsic vibration signals; and construct a decoupled characteristic spectrum through cross-validation based on intrinsic vibration signals, temperature gradient change characteristics, current spectrum characteristics, and acoustic emission energy distribution characteristics. Step S2 includes: Perform Hilbert-Huang transform on the decoupled characteristic spectrum to obtain the time-frequency domain feature representation; based on the time-frequency domain feature representation, extract the instantaneous frequency and instantaneous amplitude features to form the instantaneous characteristic spectrum; perform frequency domain conversion on the instantaneous characteristic spectrum to obtain the spectrum feature map; perform fault mode identification based on the spectrum feature map and establish a fault feature library; perform tensor decomposition on the vibration signal features, temperature gradient change features, current spectrum features and acoustic emission energy distribution features to obtain cross-modal feature representation; based on the cross-modal feature representation, calculate the correlation matrix between each mode and construct a modal correlation network; perform key feature screening based on the modal correlation network to obtain a fault diagnosis feature set; perform feature matching based on the fault diagnosis feature set and the fault feature library to construct a fault feature mapping network; Step S3 includes: Construct a bidirectional long short-term memory network model and a self-attention mechanism model; integrate the bidirectional long short-term memory network model with the self-attention mechanism model to construct an initial time series prediction model; use historical fault data to train the initial time series prediction model to obtain a hybrid time series prediction model; use the sliding time window method to extract features from the fault feature mapping network, and input the extracted features into the hybrid time series prediction model to obtain the equipment remaining life prediction results; based on the remaining life prediction results, construct a degradation trend curve; compare and analyze the degradation trend curve with the preset health benchmark to obtain an equipment health trend map.

2. The method for industrial equipment fault prediction and health management based on multi-sensor fusion according to claim 1 is characterized in that: Decomposing the vibration signal to obtain vibration signal features includes: A wavelet basis function suitable for bearing fault characteristics is selected, and the vibration signal is decomposed by wavelet packets to obtain sub-signals of different frequency bands. The energy of the sub-signals of different frequency bands is calculated to obtain a frequency band energy distribution diagram. Frequency band selection is performed based on the frequency band energy distribution diagram to screen out the key frequency bands containing fault information. The key frequency band signal is reconstructed to obtain the vibration sub-band signal. The kurtosis, skewness and root mean square value of the vibration sub-band signal are calculated to form the vibration signal characteristics.

3. The method for industrial equipment fault prediction and health management based on multi-sensor fusion according to claim 1 is characterized in that: The performing frequency domain conversion on the instantaneous characteristic spectrum to obtain a frequency spectrum characteristic graph includes: Fast Fourier transform is used to convert the instantaneous characteristic spectrum into frequency domain to obtain the initial spectrum; the initial spectrum is subjected to second-order spectrum analysis to obtain a bispectrum; the bispectrum is subjected to adaptive threshold segmentation to extract significant characteristic frequency points; and a spectrum characteristic map is constructed based on the significant characteristic frequency points.

4. The method for industrial equipment fault prediction and health management based on multi-sensor fusion according to claim 1 is characterized in that: Step S4 includes: Establish a multi-dimensional health evaluation index system, including vibration intensity index, temperature anomaly index, current fluctuation index and acoustic emission energy index; quantify the equipment health trend map based on the multi-dimensional health evaluation index system to obtain equipment health indicators; normalize various equipment health indicators to obtain standardized health indicators; perform multi-source fusion of standardized health indicators to obtain a comprehensive health score; divide equipment health levels based on the comprehensive health score; perform index evaluation on standardized health indicators to obtain reliability indicators; generate an equipment health status report based on the equipment health level and reliability indicators.

5. The method for industrial equipment fault prediction and health management based on multi-sensor fusion according to claim 4 is characterized in that: The multi-source fusion of standardized health indicators to obtain a comprehensive health score includes: A basic probability distribution function was constructed to map each standardized health indicator into evidence support. The conflict coefficient of each evidence support was calculated, and the credibility of each standardized health indicator was adjusted based on the conflict coefficient to obtain the adjusted evidence support. The adjusted evidence support was fused using the DS evidence theory, and the fused belief function and likelihood function were calculated. Based on the belief function and likelihood function, a comprehensive health score was calculated.

6. The method for industrial equipment fault prediction and health management based on multi-sensor fusion according to claim 1 is characterized in that: Generating maintenance decision recommendations based on the equipment health status report includes: Based on the equipment health status report, abnormal degradation patterns are identified, matched with the historical fault case library, and possible fault types are inferred; maintenance assessments are conducted based on the fault type and the equipment's remaining life prediction results, and maintenance priorities are generated; based on the maintenance priorities, maintenance time window recommendations are formulated; based on the fault type, maintenance operation instructions are generated; and maintenance decision recommendations are formed based on the maintenance time window recommendations and maintenance operation instructions.

7. The method for industrial equipment fault prediction and health management based on multi-sensor fusion according to claim 1 is characterized in that: The implementation of real-time anomaly detection and maintenance suggestion push through edge computing includes: Sensor data is monitored in real time through edge computing devices, and health status thresholds are preset. When abnormal sensor data is detected, an early warning mechanism is triggered and maintenance decision recommendations are pushed.

Images