A method, device and medium suitable for mechanical equipment abnormal state multi-parameter monitoring

Abstract

The present application relates to the technical field of mechanical equipment state monitoring, and discloses a kind of method, equipment and medium suitable for mechanical equipment abnormal state multi-parameter monitoring, the present application aims at solving the problems of high false alarm rate, early abnormality not easy to identify and residual life difficult to quantitatively predict caused by single monitoring parameter and poor adaptability of static threshold under variable working conditions in prior art.The present application synchronously acquires vibration, temperature and working condition data; utilizes deep autoencoder to extract layer by layer and fuse information, generates a fusion feature vector; constructs a dynamic health baseline associated with working conditions, quantifies the health index by calculating the deviation degree of the feature vector from the baseline; based on long short-term memory network, analyzes the evolution law of health index, deduces the degradation trajectory to predict the residual service life.The present application realizes deep fusion of multi-source information and self-adaptive monitoring of working conditions, improves the accuracy of early fault warning, and provides a quantitative basis for life prediction.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment condition monitoring technology, and in particular to a method, equipment and medium for multi-parameter monitoring of abnormal conditions of mechanical equipment. Background Technology

[0002] With the continuous improvement of industrial automation, rotating mechanical equipment (such as wind turbines, aero engines, and precision machine tool spindles) serves as the core power source, and their operational reliability directly affects the safety and efficiency of the entire system. Currently, the industry widely adopts Predictive Fault and Health Management (PHM) technology for the maintenance of these devices. Existing PHM systems typically rely on various sensors (such as vibration and temperature sensors) installed on the equipment to collect operational data and identify and alarm on abnormal equipment conditions by monitoring whether these individual physical quantities exceed preset static alarm thresholds.

[0003] In complex industrial scenarios, the operating conditions of mechanical equipment (such as speed, load, and ambient temperature) often fluctuate frequently. This leads to significant non-stationarity and variability in the monitoring data characterizing the equipment's condition. Traditional monitoring methods based on single parameters and static thresholds struggle to distinguish whether data fluctuations are caused by normal changes in operating conditions or by early-stage fault degradation. This results in existing technologies often exhibiting low monitoring sensitivity and high false alarm rates when dealing with variable operating conditions. They struggle to accurately capture subtle early-stage fault characteristics and cannot reliably quantify and predict the remaining service life of equipment before a failure occurs. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, in order to solve the problems of high false alarm rate, difficulty in identifying early anomalies, and difficulty in quantifying and predicting remaining life caused by the single monitoring parameter and poor adaptability of static threshold in the existing technology under variable working conditions, the present invention provides a multi-parameter monitoring method, equipment and medium suitable for abnormal state monitoring of mechanical equipment.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a multi-parameter monitoring method for abnormal states of mechanical equipment, which includes the following steps:

[0008] S1. Simultaneously acquire multiple operating parameters including vibration data, temperature data, and operating condition data of mechanical equipment;

[0009] S2. Extract and fuse the multi-path operating parameters layer by layer to generate a single fusion feature vector that can characterize the overall state of the equipment.

[0010] S3. Compare the fused feature vector with the dynamic health baseline associated with the same operating condition data to calculate the health index of the current equipment health status;

[0011] S4. Analyze the evolution pattern of the time series of the health index, deduce the future degradation trajectory of the health index, and calculate the time required for the degradation trajectory to reach the preset failure threshold, thereby obtaining the remaining service life.

[0012] As a preferred embodiment of the multi-parameter monitoring method for abnormal states of mechanical equipment described in this invention, the specific steps for simultaneously acquiring multiple operating parameters including vibration data, temperature data, and operating condition data of the mechanical equipment are as follows:

[0013] Vibration sensors are installed at predetermined critical load-bearing locations on the mechanical equipment to collect vibration data, and temperature sensors are installed at predetermined critical heat-generating locations where heat easily accumulates to collect temperature data. Using a unified time synchronization protocol, each data point collected by the vibration sensors, temperature sensors, and the equipment control system used to acquire the operating condition data is appended with a record indicating the specific time of collection. The vibration data is collected at a higher first sampling frequency, and the temperature data and operating condition data are collected at a lower second sampling frequency.

[0014] As a preferred embodiment of the multi-parameter monitoring method for abnormal states of mechanical equipment described in this invention, the method involves: extracting and fusing the multiple operating parameters layer by layer to generate a single fusion feature vector that can characterize the overall state of the equipment. The specific steps are as follows:

[0015] A pre-training step is performed using a deep autoencoder structure containing an encoder and a decoder to add interference noise to the original operating parameters representing the device's health status. The operating parameters with added interference noise are then provided as input to the deep autoencoder, and the decoder is set to recover the original operating parameters without added interference noise as the training objective. During actual monitoring, only the encoder portion trained in the pre-training step is used. The multi-channel operating parameters acquired in real time are input into the encoder, and the final output is taken as a single fusion feature vector.

[0016] As a preferred embodiment of the multi-parameter monitoring method for abnormal states of mechanical equipment described in this invention, the fused feature vector is compared with a dynamic health baseline associated with data under the same operating conditions to calculate a health index of the current equipment health status. The specific steps are as follows:

[0017] A dynamic health baseline is pre-established. A statistical analysis is performed on a health sample library containing fused feature vectors corresponding to the healthy operation of the equipment under different operating conditions to obtain the population distribution characteristics of the fused feature vectors corresponding to the health status under each specific operating condition. Real-time comparison is performed to obtain the current real-time operating condition data, and the population distribution characteristics corresponding to the real-time operating condition data are retrieved from the dynamic health baseline. The current real-time fused feature vector is compared with the retrieved population distribution characteristics to quantify the degree to which the real-time fused feature vector deviates from the health status under its corresponding operating condition, and the health index is generated based on the degree of deviation.

[0018] As a preferred embodiment of the multi-parameter monitoring method for abnormal states of mechanical equipment described in this invention, the following steps are taken: The evolution pattern of the health index is analyzed over a time series to deduce its future degradation trajectory, and the time required for the degradation trajectory to reach a preset failure threshold is calculated, thereby obtaining the remaining service life.

[0019] The evolutionary pattern analysis of the health index time series is achieved through a Long Short-Term Memory (LSTM) network. The input of the LSM network includes both the time series of the health index and the corresponding time series of the operating condition data. The specific implementation method for deducing the future degradation trajectory of the health index and calculating the remaining service life includes: using the LSM network for iterative prediction to generate the future degradation trajectory, wherein each iteration uses the health index value predicted in the previous step as part of the input sequence for the next prediction; comparing the generated future degradation trajectory with a pre-set health index threshold representing equipment failure; determining the future time point when the predicted value of the future degradation trajectory first falls below the health index threshold; and calculating the time difference from the future time point to the current time to obtain the remaining service life.

[0020] Secondly, the present invention provides a multi-parameter monitoring device suitable for abnormal states of mechanical equipment, characterized in that it comprises:

[0021] The data acquisition unit includes a vibration sensor and a temperature sensor respectively installed on the mechanical equipment, and an interface for acquiring operating condition data from the control system of the mechanical equipment. The data acquisition unit is configured to synchronously acquire the vibration data, the temperature data and the operating condition data, and output these data together as multiple operating parameters.

[0022] A processing unit, connected to the data acquisition unit, is configured to receive the multiple operating parameters and perform the following operations: extracting and fusing the received multiple operating parameters layer by layer to generate a single fused feature vector that characterizes the overall state of the equipment; comparing the fused feature vector with a pre-stored dynamic health baseline associated with the operating data to calculate a health index of the current equipment health level; analyzing the evolution of the health index over time to deduce its future degradation trajectory, and calculating the time required for the degradation trajectory to reach a preset failure threshold, thereby obtaining the remaining service life.

[0023] The system also includes a human-computer interaction unit connected to the processing unit, which is used to display to the user the health index and remaining lifespan calculated by the processing unit.

[0024] As a preferred embodiment of the multi-parameter monitoring device for abnormal states of mechanical equipment described in this invention, the internal configuration of the processing unit specifically includes:

[0025] The processing unit is further configured to: perform the layer-by-layer extraction and information fusion processing through its internally embedded encoder program module; search for and call the corresponding population distribution features from the dynamic health baseline stored in its internal storage area based on real-time operating data; and generate the health index by comparing the real-time fused feature vector with the population distribution features to quantify the degree of deviation.

[0026] As a preferred embodiment of the multi-parameter monitoring device for abnormal states of mechanical equipment described in this invention, the processing unit further includes a long short-term memory network program module, and the processing unit is further configured to: perform iterative prediction through the long short-term memory network program module to generate the future degradation trajectory; compare the future degradation trajectory with a preset failure threshold, and calculate the remaining service life.

[0027] Thirdly, the present invention provides a multi-parameter monitoring medium suitable for abnormal states of mechanical equipment, characterized in that it comprises:

[0028] When the program is executed by the processor, it drives the processor to perform the following operations: receive multiple operating parameters including vibration data, temperature data, and operating condition data; convert the multiple operating parameters into a fused feature vector through layer-by-layer extraction and information fusion; call the corresponding dynamic health baseline based on the operating condition data, and generate a health index based on the comparison result between the fused feature vector and the dynamic health baseline; predict the future degradation trajectory based on the historical change pattern of the health index, and calculate the remaining service life based on the relationship between the degradation trajectory and the failure threshold.

[0029] The beneficial effects of this invention are:

[0030] This invention acquires multiple operating parameters, including vibration, temperature, and operating condition data, simultaneously. It then utilizes a deep autoencoder to extract and fuse information from the multi-source heterogeneous data layer by layer, generating a fused feature vector representing the overall equipment status. A dynamic health baseline, correlated with operating conditions, is constructed, and the health index is quantified by calculating the deviation between the fused feature vector and the baseline. Finally, based on a long short-term memory network, the historical evolution of the health index is analyzed to predict future degradation trajectories and remaining service life. This invention achieves deep fusion of multi-source information and adaptive monitoring of operating conditions, effectively improving the accuracy and robustness of early fault warnings for mechanical equipment and providing a quantitative basis for predictive life prediction. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is the overall flowchart of the method;

[0033] Figure 2 S1 data synchronization acquisition flowchart;

[0034] Figure 3 Flowchart for S2 deep feature fusion and SDAE training;

[0035] Figure 4 Flowchart for S3 Dynamic Health Baseline Construction and HI Calculation

[0036] Figure 5 This is a flowchart of the S4LSTM iterative RUL prediction process. Detailed Implementation

[0037] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0038] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0039] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0040] Example 1

[0041] Reference Figures 1-5 This is the first embodiment of the present invention, which provides a multi-parameter monitoring method for abnormal states of mechanical equipment, including the following steps:

[0042] S1. Simultaneously acquire multiple operating parameters including vibration data, temperature data, and operating condition data of mechanical equipment. The specific operation steps are as follows:

[0043] Vibration sensors are installed at predetermined critical load-bearing locations on the mechanical equipment to collect vibration data, and temperature sensors are installed at predetermined critical heat-generating locations where heat is easily accumulated to collect temperature data. Through a unified time synchronization protocol, each piece of data collected by the vibration sensors, temperature sensors, and the equipment control system used to acquire operating condition data is appended with a record indicating the specific time of collection. Vibration data is collected at a higher first sampling frequency, and temperature data and operating condition data are collected at a lower second sampling frequency.

[0044] It should be noted that sensor placement is necessary on the monitored machinery. Accelerometers are installed on critical load-bearing components to collect vibration data; temperature sensors are installed near components that readily reflect the equipment's thermal state (such as bearings and lubrication systems) to collect temperature data. Simultaneously, operating condition data reflecting the equipment's operational status, such as speed, load, and pressure, are acquired from the equipment's control system (such as a PLC or DCS) via standard industrial communication protocols. To ensure strict temporal comparability of data from different sources, all data is acquired by a data acquisition system with hardware synchronization capabilities. For each acquired data point, the system attaches a record indicating the specific time of acquisition with a unified benchmark. Considering the changing characteristics of different physical quantities, a higher first sampling frequency (e.g., above 20kHz) is set for rapidly changing vibration data, while a lower second sampling frequency (e.g., 1Hz) is set for slowly changing temperature and operating condition data.

[0045] It should also be noted that, to ensure strict temporal comparability of data from different sources, all data in this embodiment is collected by a data acquisition system with hardware synchronization capabilities. This system attaches a record with a unified benchmark to each piece of data, indicating the specific time it was collected. The specific implementation is as follows:

[0046] a) Establish a unified time reference: A high-precision crystal oscillator or an external GPS clock source is used as the master clock for the entire data acquisition system. The master clock generates a standard clock pulse signal at a fixed frequency, providing a unified time reference for all acquisition channels. In a distributed acquisition system, the time signal from the external time server is synchronously distributed to each acquisition node through a high-precision network time protocol (such as PTP / IEEE1588 protocol).

[0047] b) Synchronous Trigger Mechanism: All data acquisition cards are connected to the master clock via a hardware synchronization bus (such as the PXIe backplane bus or a dedicated synchronization trigger line). When the master clock sends a trigger pulse, all acquisition cards connected to this bus will simultaneously perform sampling actions under the drive of this pulse to complete the analog-to-digital conversion. This hardware-level synchronous trigger mechanism ensures the consistency of data point acquisition actions in time across different acquisition channels.

[0048] c) Hardware Timestamp Appending: Upon completion of each sampling, the hardware logic circuitry of the acquisition card reads the current master clock count and stores this count value as a timestamp for the data point, along with the data value. The timestamp records the exact moment the data point was acquired. Because this process is entirely completed at the hardware level and does not involve operating system scheduling, it avoids the uncertain delays that may arise from software timestamps.

[0049] d) Time reference calibration and maintenance: For long-term continuous operation acquisition systems, the master clock is periodically calibrated through an external time source (such as GPS, Beidou satellite signal or network time server) to compensate for the time drift that may be caused by the crystal oscillator and ensure that the attached time records maintain a consistent reference throughout the entire monitoring period.

[0050] S2. Perform layer-by-layer extraction and information fusion processing on multiple operating parameters to generate a single fusion feature vector that can characterize the overall status of the equipment. The specific operation steps are as follows:

[0051] A pre-training step is performed using a deep autoencoder structure containing an encoder and a decoder to add interference noise to the original operating parameters representing the device's health status. The operating parameters with added interference noise are then fed into the deep autoencoder, and the training objective is set to allow the decoder to recover the original operating parameters without added interference noise. During actual monitoring, only the encoder portion trained in the pre-training step is used. The multi-channel operating parameters acquired in real time are input into the encoder, and the final output is taken as a single fusion feature vector.

[0052] It should be noted that a pre-trained deep autoencoder model is used to extract and fuse features. Before inputting the data into the model, preprocessing is required: denoising the high-frequency vibration data; interpolating the low-frequency temperature and operating condition data to align their time points with the high-frequency data; and finally, standardizing all channel data to eliminate the influence of units. The data standardization can be expressed by the following formula:

[0053]

[0054] : Indicates the new data points obtained after standardization.

[0055] This represents any data point in the raw data stream. The range of values ​​depends on the specific physical quantity. For example, vibration data (unit: m / s²) may vary between -50 and +50; temperature data (unit: °C) may vary between 20 and 80.

[0056] : Indicates the average value of the data over a period of time (e.g., a collection cycle or a training batch).

[0057] : This represents the standard deviation of the data over the same time period, which measures the degree of fluctuation in the data from the mean.

[0058] The deep autoencoder used in this embodiment comprises an encoder and a decoder. This model requires a pre-training step: a large amount of operational data from the monitored device under confirmed health conditions is collected as a training set. During training, noise is added to each input healthy sample. This noise-added operational parameters are then provided as input to the deep autoencoder, and the decoder is set to recover the original operational parameters without added noise as the training objective. Through this noise reduction training, the encoder is forced to learn a more fundamental and robust intrinsic structure within the data.

[0059] In actual monitoring, only the encoder portion trained in the aforementioned pre-training steps is used. Multiple operating parameters acquired in real-time and preprocessed are input into this encoder, and its final output is a low-dimensional vector, which is the single fused feature vector generated in this step that characterizes the overall state of the equipment at the current moment. For example, a raw data segment containing multiple vibration, temperature, and operating conditions data, totaling thousands of data points, can be converted into a fused feature vector with tens or hundreds of dimensions.

[0060] S3. Compare the fused feature vector with the dynamic health baseline associated with data under the same operating conditions to calculate the health index of the current equipment health status. The specific operation steps are as follows:

[0061] A dynamic health baseline is pre-established. Statistical analysis is performed on a health sample library containing fused feature vectors corresponding to the healthy operation of equipment under different operating conditions to obtain the population distribution characteristics of the fused feature vectors corresponding to the health status under each specific operating condition. Real-time comparison is performed to obtain the current real-time operating condition data and to call the population distribution characteristics corresponding to the real-time operating condition data from the dynamic health baseline. The current real-time fused feature vector is compared with the called population distribution characteristics to quantify the degree to which the real-time fused feature vector deviates from the health status under its corresponding operating condition, and a health index is generated based on the degree of deviation.

[0062] It should be noted that the process of establishing a dynamic health baseline is as follows:

[0063] During periods when the equipment is in a confirmed health state (e.g., immediately after a major overhaul), operational data is collected under various typical operating conditions. This data is converted into fused feature vectors by an encoder, while the corresponding operating condition data for each vector is recorded (e.g., speed of 1200 rpm, load of 80%). The multidimensional operating condition space is divided into grids, for example, dividing the speed into multiple intervals and the load into multiple intervals, forming a two-dimensional operating condition grid. For all health fused feature vectors falling within the same operating condition grid, statistical analysis methods (e.g., Gaussian mixture model) are used to fit the group distribution characteristics, i.e., describing where these vectors cluster and what their shape is. These group distribution characteristics (e.g., the mean, covariance, and weights of multiple Gaussian components) are recorded and stored in a lookup table indexed by the operating condition grid. This lookup table is what this invention refers to as the dynamic health baseline. Essentially, it is a table mapping operating condition health states, recording what the fused feature vectors should look like when the equipment is healthy under different operating conditions.

[0064] It should also be noted that the real-time calculation process for the health index is as follows:

[0065] During equipment operation, whenever a new fused feature vector is generated, the system first acquires the corresponding real-time operating condition data. Based on this data, it then finds the best-matching operating condition grid in the dynamic health baseline lookup table and retrieves the corresponding population distribution features. The current real-time fused feature vector is compared with these features to quantify its conformity. This quantification is typically achieved by calculating a statistical distance or a negative log-likelihood (NLL). A smaller NLL value indicates that the current feature vector closely matches the health distribution under that operating condition; a larger NLL value indicates a greater deviation. This NLL value is then converted into a health index (HI) between 0 and 1 using an exponential function. An HI value close to 1 indicates that the equipment is very healthy, while a decreasing HI value indicates that the equipment's condition is beginning to deteriorate. This exponential function design makes the health index curve very smooth and intuitive, facilitating subsequent trend analysis. The formula for calculating the health index HI is as follows;

[0066]

[0067] : Represents the final calculated health index, ranging from 0 to 1.

[0068] : Represents the natural exponential function, that is, the exponential operation with the natural constant e as the base.

[0069] : Represents a scaling factor, a positive number used to adjust the sensitivity of the health index to deviations. This is a hyperparameter set by the user based on experience; it has no fixed range but is usually a small positive number, such as between 0.01 and 1. The larger, right The more sensitive the change, the faster it decreases. In this embodiment, a selection can be made. =0.1.

[0070] : Represents the negative log-likelihood. It represents the degree to which the current fused feature vector deviates from its corresponding healthy state under the operating conditions. This value is non-negative, and the greater the deviation, the larger the NLL value. The value range is [0, +∞). In practical applications, its value is close to 0 under healthy conditions, and may reach several hundred or even several thousand under severe faults.

[0071] S4. Analyze the evolution pattern of the health index over time, deduce the future degradation trajectory of the health index, and calculate the time required for the degradation trajectory to reach the preset failure threshold, thereby obtaining the remaining service life. The specific operation steps are as follows:

[0072] The evolutionary pattern analysis of the health index over time is achieved through a Long Short-Term Memory (LSTM) network. The input to the LSM network includes both the time series of the health index and the corresponding time series of operating condition data. The specific implementation method for deducing the future degradation trajectory of the health index and calculating the remaining service life includes: using the LSM network for iterative prediction to generate the future degradation trajectory, where each iteration uses the predicted health index value from the previous step as part of the input sequence for the next prediction; comparing the generated future degradation trajectory with a pre-set health index threshold representing equipment failure; determining the future time point when the predicted value of the future degradation trajectory first falls below the health index threshold; and calculating the time difference between the future time point and the current time to obtain the remaining service life.

[0073] It should be noted that the process of evolution pattern analysis and trend prediction is as follows:

[0074] This embodiment employs a Long Short-Term Memory (LSTM) network to learn the evolution pattern of the health index. LSTM is a deep learning model for processing time-series data. During LSTM training, historical operating data of the equipment (including healthy and deteriorating phases) is used. The health index and operating condition data are taken as the input sequence, and the future health index is taken as the output target. Through learning from a large amount of historical data, LSTM masters the complex evolution pattern of the health index's decline rate under what operating conditions.

[0075] In actual prediction, the prediction function is triggered when the system detects a continuous and irreversible downward trend in the health index. The prediction employs an iterative method: a sequence of health index and operating condition data from a past period (e.g., the past 2 hours) is fed into a trained LSTM model to predict the health index for the first future time point. This predicted value, along with the assumed future operating condition (e.g., maintaining the current average operating condition), is appended to the end of the historical sequence to predict the second future time point. This process is repeated until the predicted health index falls below a pre-set failure threshold (e.g., 0.2). This threshold is determined based on a comprehensive consideration of equipment importance, maintenance costs, and safety requirements.

[0076] It should also be noted that the remaining useful life is calculated as follows:

[0077] Suppose that at the current moment, after N iterations of prediction, the health index falls below the failure threshold for the first time. Then, the time corresponding to these N iterations is the remaining safe operating time of the device from the current moment, i.e., the remaining useful life (RUL). This RUL value is displayed in real-time on the monitoring interface, and different levels of maintenance recommendations are triggered based on its length (e.g., recommending an inspection within 24 hours or recommending immediate shutdown for maintenance). The formula for calculating the remaining useful life (RUL) is as follows:

[0078]

[0079] : Indicates the final calculated remaining useful life, in time (e.g., hours, days).

[0080] : This indicates the predicted failure point, determined through iterative prediction, when the future health index first falls below the failure threshold.

[0081] : Indicates the current time when this RUL calculation is being performed.

[0082] Example 2

[0083] This embodiment provides a multi-parameter monitoring device suitable for abnormal conditions of mechanical equipment, including:

[0084] The data acquisition unit includes vibration sensors and temperature sensors installed on the mechanical equipment, as well as an interface for acquiring operating condition data from the control system of the mechanical equipment. The data acquisition unit is configured to synchronously acquire vibration data, temperature data, and operating condition data, and output these data together as multiple operating parameters.

[0085] The processing unit, connected to the data acquisition unit, receives multiple operating parameters and is configured to perform the following operations: extract and fuse the received multiple operating parameters layer by layer to generate a single fused feature vector that characterizes the overall state of the equipment; compare the fused feature vector with a pre-stored dynamic health baseline associated with the same operating condition data to calculate the health index of the current equipment health level; analyze the evolution pattern of the health index time series to deduce the future degradation trajectory of the health index, and calculate the time required for the degradation trajectory to reach a preset failure threshold, thereby obtaining the remaining service life.

[0086] It also includes a human-computer interaction unit, which is connected to the processing unit and is used to display the health index and remaining lifespan calculated by the processing unit to the user.

[0087] It should be noted that this device is an independent system integrating data acquisition, processing, and human-computer interaction functions. The composition and collaborative working method of its internal units are as follows:

[0088] 1. Data Acquisition Unit

[0089] The data acquisition unit in this embodiment serves as the interface for the physical connection between the device and the monitored mechanical equipment. Its core task is to convert various physical signals into a unified digital data stream with precise time recording. This unit specifically includes:

[0090] Sensor Interface and Signal Conditioning Circuitry: The unit's external interface provides multiple standardized sensor connection ports, such as a BNC connector for connecting an accelerometer and a terminal block for connecting a resistance temperature detector (RTD). The internal signal conditioning circuitry amplifies, filters, and performs anti-aliasing processing on the incoming raw analog signal to optimize signal quality.

[0091] Multifunctional Synchronous Analog-to-Digital Converter (ADC) Module: Conditioned multi-channel analog signals, along with operating condition data acquired from the control system of the monitored equipment (such as a PLC) via digital communication interfaces (such as RS-485 or Ethernet), are all aggregated into a single multifunctional synchronous ADC module. The core of the module is a high-speed, high-precision ADC array controlled by an FPGA.

[0092] Synchronization and Timestamp Attachment Mechanism: The FPGA internally operates a unified clock driven by a high-stability crystal oscillator. Under the control of this clock, the FPGA simultaneously sends sampling trigger signals to all ADC channels, ensuring that all channels are physically synchronized for sampling. At each sampling instant, the FPGA immediately reads the value of its internal counter and uses this value as a record indicating the specific moment of acquisition, packaging it together with the data samples from all acquired channels to form a complete data frame.

[0093] The data acquisition unit ultimately outputs multiple operating parameters containing vibration, temperature, and operating condition information to the processing unit via a high-speed internal bus in a predefined data frame format.

[0094] 2. Processing Unit

[0095] The processing unit in this embodiment is typically mounted on an embedded computing motherboard. The unit receives data frames from the data acquisition unit, performs a series of calculations on them, and ultimately outputs key results regarding the device's health status and future trends. Its internal configuration and workflow are as follows:

[0096] Hardware configuration: Employs a multi-core ARM or x86 architecture central processing unit (CPU), coupled with a dedicated graphics processing unit (GPU) for parallel computing or a dedicated artificial intelligence (AI) acceleration chip. The unit also includes high-speed RAM for runtime data caching and a non-volatile solid-state drive (SSD) for storing the operating system, core software programs, and critical data models.

[0097] Software Configuration and Data Flow: The core monitoring and prediction software program of this invention is pre-installed in the SSD of the processing unit. When the device is powered on, the processing unit first reads data frames from the output bus of the data acquisition unit.

[0098] The first stage (feature fusion): The software program first calls the encoder program module embedded in the processing unit. The CPU is responsible for data preprocessing (such as denoising, interpolation, and standardization), and then sends the preprocessed data to the GPU. The GPU executes the encoder program module's operations in parallel, processing the high-dimensional multi-path parameters into a low-dimensional single fused feature vector.

[0099] Phase Two (Health Assessment): The software program sends the fused feature vector calculated by the GPU, along with the operating condition data parsed from the data frame, back to the CPU. The CPU loads a dynamic health baseline data table pre-stored in the SSD into memory. Based on the operating condition data, the program quickly looks up and retrieves the corresponding population distribution characteristics from this table. The CPU performs statistical calculations, comparing the fused feature vector with the distribution characteristics to quantify the degree of deviation and ultimately calculate the Health Index (HI).

[0100] The third stage (lifetime prediction): The software program continuously monitors the trend of HI value changes. When a preset degradation condition is detected, another Long Short-Term Memory (LSTM) network program module embedded in the processing unit is activated. The CPU sends the historical HI and operating condition sequence to the GPU, and the GPU performs iterative prediction operations of the LSTM module to deduce the future degradation trajectory. The CPU then calculates the final remaining lifetime (RUL) based on this trajectory and a preset failure threshold.

[0101] The two key results, HI and RUL, calculated by the processing unit are sent to the human-machine interaction unit via the internal data bus.

[0102] 3. Human-Computer Interaction Unit

[0103] The human-computer interaction unit in this embodiment serves as the window for information exchange between the device and the external world. Its core task is to transform the abstract data output by the processing unit into visual information that human users can understand and use for decision-making. This unit provides two interaction methods:

[0104] Local display interface: The unit includes a display controller that receives display commands and data from the processing unit and drives a local display screen (such as an LCD touchscreen) integrated with the device. The screen displays the current health index and remaining lifespan in real time in a graphical manner (such as a dashboard or graph), and can also display alarm information.

[0105] Remote Communication Interface: The unit also includes a separate network communication module with a built-in web server. The processing unit continuously pushes key data such as HI and RUL to this web server. External users (such as engineers in a remote monitoring center) can access the device's IP address through a standard web browser. The web server sends a pre-designed web page front-end to the browser. This front-end communicates with the web server in real time via technologies such as AJAX or WebSocket, thereby dynamically and graphically displaying all the device's operating statuses and predicted results on a remote computer screen.

[0106] Example 3

[0107] This embodiment provides a multi-parameter monitoring medium suitable for abnormal conditions of mechanical equipment, including:

[0108] It stores a computer program, which, when executed by the processor, causes the processor to perform the following operations: receive multiple operating parameters containing vibration data, temperature data, and operating condition data; convert the multiple operating parameters into a fused feature vector through layer-by-layer extraction and information fusion; call the corresponding dynamic health baseline based on the operating condition data, and generate a health index based on the comparison result between the fused feature vector and the dynamic health baseline; predict the future degradation trajectory based on the historical change pattern of the health index, and calculate the remaining service life based on the relationship between the degradation trajectory and the failure threshold.

[0109] Example 4

[0110] To further illustrate the present invention, the following description uses the maintenance of the main bearing of a 5MW wind turbine generator set in a large offshore wind farm as an example to provide a more detailed explanation of the technical solution of the present invention.

[0111] In this embodiment, the wind farm's operation and maintenance team faces a problem: the maintenance cost of offshore wind turbines is extremely high, and each offshore maintenance trip requires renting expensive engineering vessels. Therefore, it is necessary to know which turbine's main bearing is likely to fail in the coming months in order to plan a concentrated offshore maintenance trip in advance, avoiding emergency shutdowns and additional offshore costs caused by a single turbine's sudden failure. The monitoring equipment of this invention is deployed on this wind turbine, numbered WF-07, and its specific operating process is as follows:

[0112] S1: Constructing a comprehensive physical sensing network:

[0113] The monitoring equipment of this invention was deployed on the main bearing housing of the WF-07 wind turbine. A high-frequency vibration acceleration sensor was installed at the horizontal radial position of the bearing housing to capture minute impact signals during bearing rotation. Simultaneously, a PT100 temperature sensor was installed in close contact with the outer shell of the bearing housing to monitor the real-time temperature of the bearing. The monitoring equipment was also connected to the main control cabinet of the wind turbine via a network cable, and read the two key operating parameters—real-time wind speed and current power generation—using the Modbus TCP protocol.

[0114] After all sensors were installed, the internal synchronization acquisition card of the monitoring equipment began operation. It acquired vibration signals at a high speed of 25.6kHz, and simultaneously acquired temperature and operating condition data at a frequency of 1Hz. The equipment assigned a unified timestamp to these three types of data from different sources and with different frequencies to ensure that they could be aligned during subsequent analysis.

[0115] S2: Extracting digital feature profiles of equipment from massive amounts of data:

[0116] The monitoring equipment continuously collects raw data. For complex waveforms and numbers, the internal processing unit activates its internally integrated deep feature fusion module. This module utilizes deep learning technology to process the data: first, it filters out wave noise and electromagnetic interference from the vibration signal; then, it aligns and standardizes the four heterogeneous data points—vibration, temperature, rotational speed, and power—before inputting them into a pre-trained deep autoencoder. Through layers of extraction and information compression by the encoder, the tens of thousands of raw data points per second are abstracted and condensed into a high-dimensional feature vector containing 128 values. This vector is the digital feature profile (i.e., the fused feature vector) of the main bearing at the current moment. While it lacks intuitive physical meaning, it highly summarizes and contains the most essential, multi-parameter coupled operating status information of the bearing, serving as the sole data basis for subsequent health measurements and trend predictions.

[0117] S3: Dynamic assessment of health status:

[0118] With the equipment's digital profile established, the next step is to determine its operational status. Let's assume the wind turbine is currently operating at a high load of 12 rpm and 5 MW. The monitoring equipment will immediately search the dynamic health baseline: Under the same 12 rpm and 5 MW conditions, what distribution characteristics should the digital profile of a fully healthy main bearing exhibit?

[0119] After locating the corresponding standard, the equipment compared the currently measured digital feature profile with the standard. The results showed that although the temperature and rotational speed appeared to be within the normal range, certain high-frequency components in the vibration signal had caused a slight deviation between the measured profile and the standard. Calculations determined that the current deviation (NLL value) was slightly high, with a converted Health Index (HI) of 0.85. This indicates that while the bearing is not yet damaged, it is no longer in perfect health.

[0120] S4: Predict the future and guide decision-making:

[0121] As time went on, the monitoring equipment found that the Health Index (HI) of wind turbine WF-07 began to show a slow but continuous downward trend: from 0.95 last month to 0.85 this week, and then to 0.82 today.

[0122] Upon detecting this trend, the device activated its lifetime prediction module. This module uses a Long Short-Term Memory (LSTM) network, combined with historical HI decline curves and the expected wind speeds in the area over the next few weeks (i.e., future operating conditions), to project the future trajectory of HI.

[0123] Simulation results show that, based on the current rate of degradation and future wind conditions, the health index of this main bearing will fall below the failure threshold of 0.2 in 45 days, and the bearing will face an extremely high risk of seizure. The monitoring equipment displayed a yellow warning on the screen at the operations and maintenance center: the main bearing of wind turbine WF-07 is showing early degradation, with a current estimated remaining useful life (RUL) of 45 days (±3 days).

[0124] Upon receiving this warning message, the operations and maintenance team decided not to shut down the system immediately, but instead to take advantage of the 45-day window to procure spare parts in advance and include them in the routine overseas maintenance plan for the following month.

[0125] In summary, this invention simultaneously acquires multiple operating parameters including vibration, temperature, and operating condition data. It utilizes a deep autoencoder to extract and fuse information from multi-source heterogeneous data layer by layer, generating a fused feature vector representing the overall equipment status. A dynamic health baseline correlated with operating conditions is constructed, and the health index is quantified by calculating the deviation between the fused feature vector and the baseline. Furthermore, based on a long short-term memory network, the historical evolution of the health index is analyzed to predict future degradation trajectories and thus the remaining service life. This invention achieves deep fusion of multi-source information and adaptive monitoring of operating conditions, effectively improving the accuracy and robustness of early fault warnings for mechanical equipment and providing a quantitative basis for predictive life prediction of equipment.

[0126] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for multi-parameter monitoring of abnormal states of mechanical equipment, characterized in that, Includes the following steps: S1. Simultaneously acquire multiple operating parameters including vibration data, temperature data, and operating condition data of the mechanical equipment; install vibration sensors at predetermined key load-bearing locations of the mechanical equipment that reflect operational impacts to collect the vibration data, and install temperature sensors at predetermined key heat-generating locations that easily accumulate heat to collect the temperature data; through a unified time synchronization protocol, attach a record indicating the specific time of collection to each piece of data collected by the vibration sensors, the temperature sensors, and the equipment control system used to acquire the operating condition data; collect the vibration data at a higher first sampling frequency, and collect the temperature data and the operating condition data at a lower second sampling frequency than the first sampling frequency; S2. Extract and fuse the multi-path operating parameters layer by layer to generate a single fusion feature vector that can characterize the overall state of the equipment. conduct The pre-training step involves using a deep autoencoder structure that includes an encoder and a decoder to add interference noise to the original operating parameters that represent the health status of the device; the operating parameters with added interference noise are then provided as input to the deep autoencoder, and the training objective is set to enable the decoder to recover the original operating parameters without added interference noise. In actual monitoring, only the encoder portion trained in the pre-training step is used; the multi-path running parameters acquired in real time are input into the encoder, and the final output is taken as a single fusion feature vector; S3. Compare the fused feature vector with a dynamic health baseline associated with the same operating condition data to calculate a health index of the current equipment health level. Specifically, the dynamic health baseline is pre-established, and a health sample library containing fused feature vectors corresponding to the healthy operation of the equipment under different operating conditions is statistically analyzed to obtain the group distribution characteristics of the fused feature vectors corresponding to the health status under each specific operating condition. Real-time comparison is performed to obtain the current real-time operating condition data, and the group distribution characteristics corresponding to the real-time operating condition data are called from the dynamic health baseline. The current real-time fused feature vector is compared with the called group distribution characteristics to quantify the degree to which the real-time fused feature vector deviates from the health status under its corresponding operating condition, and the health index is generated based on the degree of deviation. S4. Analyze the evolution pattern of the time series of the health index, deduce the future degradation trajectory of the health index, and calculate the time required for the degradation trajectory to reach the preset failure threshold, thereby obtaining the remaining service life. Specifically, the evolution pattern analysis of the time series of the health index is achieved through a long short-term memory network. The input of the long short-term memory network includes both the time series of the health index and the time series of the corresponding working condition data.

2. The multi-parameter monitoring method for abnormal states of mechanical equipment as described in claim 1, characterized in that, The specific implementation method of deducing the future degradation trajectory of the health index and calculating the remaining service life in step S4 includes: using the long short-term memory network for iterative prediction to generate the future degradation trajectory, wherein each iteration uses the health index value predicted in the previous step as part of the input sequence for the next prediction; comparing the generated future degradation trajectory with a pre-set health index threshold representing device failure; determining the future time point when the predicted value of the future degradation trajectory first falls below the health index threshold; and calculating the time difference between the future time point and the current time to obtain the remaining service life.

3. A multi-parameter monitoring device for abnormal states of mechanical equipment, used to implement the method described in claim 1, characterized in that, include: The data acquisition unit includes a vibration sensor and a temperature sensor respectively installed on the mechanical equipment, and an interface for acquiring operating condition data from the control system of the mechanical equipment. The data acquisition unit is configured to synchronously acquire the vibration data, the temperature data and the operating condition data, and output these data together as multiple operating parameters. A processing unit, connected to the data acquisition unit, is configured to receive the multiple operating parameters and perform the following operations: extracting and fusing the received multiple operating parameters layer by layer to generate a single fused feature vector that characterizes the overall state of the equipment; comparing the fused feature vector with a pre-stored dynamic health baseline associated with the operating data to calculate a health index of the current equipment health level; analyzing the evolution of the health index over time to deduce its future degradation trajectory, and calculating the time required for the degradation trajectory to reach a preset failure threshold, thereby obtaining the remaining service life. The system also includes a human-computer interaction unit connected to the processing unit, which is used to display to the user the health index and remaining lifespan calculated by the processing unit.

4. The multi-parameter monitoring device for abnormal states of mechanical equipment as described in claim 3, characterized in that, The processing unit is further configured to: perform the layer-by-layer extraction and information fusion processing through its internally embedded encoder program module; search for and call the corresponding population distribution features from the dynamic health baseline stored in its internal storage area based on real-time operating data; and generate the health index by comparing the real-time fused feature vector with the population distribution features to quantify the degree of deviation.

5. The multi-parameter monitoring device for abnormal states of mechanical equipment as described in claim 3, characterized in that, The processing unit also has a long short-term memory network program module embedded in it, and the processing unit is further configured to: perform iterative prediction through the long short-term memory network program module to generate the future degradation trajectory; compare the future degradation trajectory with a preset failure threshold, and calculate the remaining service life.

6. A multi-parameter monitoring medium suitable for abnormal conditions of mechanical equipment, used to implement the method of claim 1, wherein a computer program is stored thereon, characterized in that, include: When the program is executed by the processor, it drives the processor to perform the following operations: receive multiple operating parameters containing vibration data, temperature data, and operating condition data; and convert the multiple operating parameters into a fused feature vector through layer-by-layer extraction and information fusion. The corresponding dynamic health baseline is invoked based on the operating condition data, and a health index is generated based on the comparison result between the fused feature vector and the dynamic health baseline; the future degradation trajectory is predicted based on the historical change pattern of the health index, and the remaining service life is calculated based on the relationship between the degradation trajectory and the failure threshold.

Images