An online monitoring method and system for motor lubrication

Abstract

The application discloses an online monitoring method and system for motor lubrication, and relates to the technical field of equipment state monitoring. The method comprises the following steps: acquiring multi-source historical data, intercepting multi-dimensional data sequences in multiple normal and abnormal lubrication states respectively, and obtaining a historical sample set through preprocessing; based on three-phase current sequences in the historical sample set, load fluctuation features are extracted to generate a window adjustment index, and an analysis window division scheme of each historical sample is dynamically determined; based on the analysis window, multi-dimensional data is analyzed, lubrication feature sequences corresponding to the analysis window are reconstructed and merged, and a historical lubrication sample set is obtained; a lubrication state evaluation model is constructed and trained based on the historical lubrication sample set; multi-dimensional data at a current time is collected in real time to construct a current lubrication sample, and the lubrication state is monitored in real time based on the trained lubrication state evaluation model. The application can realize precise online monitoring and intelligent early warning of the lubrication state of a closed gear reduction motor.

Description

Technical Field

[0001] This invention relates to the field of equipment condition monitoring technology, and in particular to an online monitoring method and system for motor lubrication. Background Technology

[0002] In industrial production, geared motors are key power equipment, and their stable operation is crucial to the reliability and efficiency of the entire production system. The lubrication condition of the gears and bearings inside the motor is a core factor affecting the equipment's lifespan and reliability. Poor lubrication directly leads to increased friction, abnormal temperature rise, and decreased efficiency, which in turn causes serious failures such as component wear, pitting, and even galling, resulting in unplanned downtime and significant economic losses.

[0003] In existing technologies, vibration analysis-based intelligent diagnostic methods are the mainstream means of monitoring equipment status. However, when traditional vibration analysis methods are applied to the lubrication status monitoring of geared motors, significant drawbacks emerge. For example, gearbox vibration signals are a mixed response of multiple excitation sources, including gear meshing, bearing rotation, and electromagnetic excitation, resulting in complex signal components and a low signal-to-noise ratio. In the early stages of lubrication degradation, the weak impacts and friction characteristics caused by abnormal friction are overwhelmed by the strong background of normal vibrations in terms of both energy and spectrum. Existing vibration analysis methods, such as spectral analysis and envelope analysis, struggle to stably and accurately separate and quantify the feature components solely related to lubrication status from this complex signal with strong noise and multi-component coupling. This leads to low signal-to-noise ratios, delayed warnings, and high false alarm rates in identifying early lubrication degradation, failing to meet the urgent need for early and accurate warnings of equipment status in high-reliability production.

[0004] Furthermore, the operating conditions of motors (such as load fluctuations) change dynamically, further increasing the uncertainty of vibration signals and making fixed-parameter analysis methods less adaptable. Therefore, the industry urgently needs a new method and system that can adapt to changes in operating conditions and directionally enhance features strongly correlated with lubrication physical processes from raw multi-source signals, thereby achieving high-sensitivity and high-reliability online monitoring. Summary of the Invention

[0005] To achieve accurate online monitoring and intelligent early warning of the lubrication status of enclosed geared motors, and to overcome the shortcomings of traditional vibration analysis methods such as feature submersion and low signal-to-noise ratio in the early deterioration stage of lubrication status, this invention provides an online monitoring method and system for motor lubrication, the technical solution of which is as follows:

[0006] In a first aspect, the present invention provides an online monitoring method for motor lubrication, comprising the following steps: acquiring multi-source historical data and extracting multiple segments of multidimensional data sequences under normal and abnormal lubrication conditions, preprocessing to obtain a historical sample set; extracting load fluctuation characteristics based on the three-phase current sequence in the historical samples to generate window adjustment indicators, and dynamically determining the analysis window division scheme for each historical sample; analyzing the multidimensional data based on the analysis windows divided from each historical sample, reconstructing the lubrication feature sequences corresponding to each analysis window of each historical sample and merging them to obtain a historical lubrication sample set; constructing a lubrication state evaluation model and training it based on the historical lubrication sample set; collecting multidimensional data at the current moment in real time to construct the current lubrication sample, and monitoring the lubrication state in real time based on the trained lubrication state evaluation model;

[0007] Each historical sample contains a vibration sequence, a three-phase current sequence, and a temperature sequence. Any historical sample is selected as the target historical sample, and any analysis window of the target historical sample is selected as the target analysis window. Wavelet packet decomposition is performed on the vibration sequence segment within the target analysis window, and the comprehensive weight of each sub-frequency band within the target analysis window is calculated. The wavelet coefficient sequences of each sub-frequency band are weighted using the normalized comprehensive weights and then superimposed and fused to obtain the composite wavelet coefficient sequence corresponding to the target analysis window. The inverse wavelet packet transform is performed on the composite wavelet coefficient sequence to obtain the lubrication feature sequence corresponding to the target analysis window. Similarly, the lubrication feature sequences corresponding to each analysis window within the target historical sample are obtained and merged in chronological order. The lubrication feature sequence corresponding to the entire target historical sample is taken as a historical lubrication sample.

[0008] Preferably, sensing devices are deployed at key measurement points of the enclosed gear reducer motor to synchronously collect vibration data, three-phase current data, and temperature data at a fixed sampling frequency. This acquires historical data and maintenance records of multiple enclosed gear reducers of the same model operating over a long period. The sample length is set, and based on the maintenance records, multiple segments of multidimensional data sequences of equal length under normal and abnormal lubrication conditions are extracted from the historical data. The extracted historical data is cleaned and standardized to obtain multiple data matrices composed of vibration sequences, three-phase current sequences, and temperature sequences. Each data matrix is ​​used as a historical sample, with historical samples under normal lubrication conditions marked as 0 and historical samples under abnormal lubrication conditions marked as 1. The set of all historical samples and their corresponding labels is used as the historical sample set.

[0009] Preferably, the length of the initial analysis window, the length of the maximum analysis window, and the window adjustment step size are set based on the length of the historical samples. Any historical sample is selected as the target historical sample. The three-phase current sequence in the target historical sample is truncated for the first time based on the length of the initial analysis window to obtain the first initial analysis window of the target historical sample. The standard deviation and information entropy of each phase current in the three-phase current within the first initial analysis window are calculated. Then, the mean standard deviation and mean information entropy of the three-phase current are calculated. The weight coefficients of the two mean features are set based on human experience. The sum of the products of the mean standard deviation and the mean information entropy with the corresponding weight coefficients is used as the window adjustment index to obtain the window adjustment index of the first initial analysis window of the target historical sample.

[0010] Preferably, based on the length of the entire three-phase current sequence in the historical samples, the window adjustment index corresponding to each historical sample under normal lubrication conditions is calculated, and its average value is used as the adjustment threshold. For the window adjustment index of the first initial analysis window of the target historical sample, when it is less than the adjustment threshold, the sum of the initial analysis window and the window adjustment step size is used as the adaptive analysis window. The window adjustment index of the adaptive analysis window is calculated iteratively and compared with the adjustment threshold. The window length is dynamically adjusted until the maximum analysis window length is reached, or the window adjustment index is greater than or equal to the adjustment threshold. The length of the maximum analysis window or the length of the corresponding adaptive analysis window is used as the optimal partition length of the first analysis window of the target historical sample. Similarly, the partitioning scheme of the next analysis window of the target historical sample is obtained in turn. All historical samples are traversed to obtain the analysis window partitioning scheme of each historical sample.

[0011] Preferably, wavelet packet decomposition is performed on the vibration sequence segment within the target analysis window to obtain wavelet coefficient sequences of multiple sub-bands. Then, Hilbert transform is performed to obtain the energy envelope signal sequence of each sub-band. The square root of the sum of the squares of the three element values ​​at the same sampling time in the three-phase current sequence of the target historical sample is calculated sequentially to obtain the current guiding signal sequence corresponding to the target historical sample. The Pearson correlation coefficient between the energy envelope signal sequence of each sub-band within the target analysis window and the corresponding current guiding signal sequence segment, as well as the Spearman correlation coefficient between the energy envelope signal sequence of each sub-band within the target analysis window and the corresponding temperature sequence segment, are calculated to obtain the Pearson correlation coefficient characteristics and Spearman correlation coefficient characteristics of each sub-band within the target analysis window.

[0012] Preferably, a balanced weight is set for the two coefficient features based on the actual application scenario. The sum of the products of the absolute values ​​of the two coefficient features and the corresponding balanced weights is used as the comprehensive weight of the corresponding sub-band within the target analysis window. The comprehensive weights of each sub-band are summed and normalized to obtain the wavelet coefficient sequences of each sub-band by weighted fusion, thereby obtaining the composite wavelet coefficient sequence corresponding to the target analysis window. Then, based on the inverse wavelet packet transform, the lubrication feature sequences corresponding to each analysis window within the target historical samples are obtained. These sequences are then merged to obtain the historical lubrication samples corresponding to the target historical samples. All historical samples are traversed and combined with the corresponding labels to obtain the historical lubrication sample set.

[0013] Preferably, a lubrication state assessment model is constructed based on a long short-term memory network. Historical lubrication samples are used as the input layer, and the label values ​​of the corresponding labels are used as the output layer. The Softmax function is used as the activation function of the model output layer, and the cross-entropy loss function is used as the loss function of the model. The lubrication state assessment model is trained based on the historical lubrication sample set until the loss function converges, thus obtaining the trained lubrication state assessment model.

[0014] Preferably, multidimensional data is synchronously collected starting from the moment the enclosed geared motor starts running. Based on the length of historical samples, data of equal length is collected and preprocessed to obtain the first current sample. The length of data collected within 1 second is used as the sliding step size based on the sampling frequency, and the current sample is updated through a sliding window. The load fluctuation characteristics corresponding to the current sample are extracted, and the analysis window partitioning scheme of the current sample is dynamically determined. The same calculation process as historical lubrication samples is used to obtain the current lubrication sample corresponding to the current sample. The current lubrication sample is input into the trained lubrication state evaluation model, and the probability that the lubrication state is abnormal within the time period corresponding to the current sample is output. A judgment threshold is set based on the accuracy requirements in the actual application scenario. When the output probability is greater than the judgment threshold, the corresponding early warning mechanism is triggered.

[0015] Secondly, the present invention provides an online monitoring system for motor lubrication, for implementing the above-mentioned online monitoring method for motor lubrication, comprising: a processor, a memory, a communication interface, a data acquisition device, and an alarm device. The processor stores computer program instructions for implementing the above-mentioned online monitoring method for motor lubrication, and the communication interface is communicatively connected to the data acquisition device and the alarm device.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] This invention utilizes three-phase current data to dynamically divide historical sample analysis windows, effectively matching the actual fluctuations of motor load and enhancing operational adaptability. Furthermore, by jointly guiding the analysis of current and temperature, and considering both the instantaneous modulation of load friction and the long-term trend of frictional heat accumulation, the vibration signal is reconstructed using a weighted frequency band. This allows for the targeted enhancement of weak features most closely related to lubrication degradation from complex vibration signals with strong noise and multi-component coupling, significantly improving the signal-to-noise ratio and specificity of lubrication characteristics. Finally, a lubrication state assessment model is constructed based on a long short-term memory network. Through deep learning of high signal-to-noise ratio lubrication feature sequences, the complex nonlinear temporal mapping relationship of lubrication state changes is accurately captured. This overcomes the problems of feature overload and low signal-to-noise ratio in early lubrication monitoring caused by traditional vibration analysis methods, achieving more sensitive and reliable early online monitoring and intelligent early warning of the lubrication state of enclosed geared motors. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an implementation of an online monitoring method for motor lubrication according to an embodiment of the present invention.

[0019] Figure 2 This is a structural block diagram of an online monitoring system for motor lubrication according to an embodiment of the present invention. Detailed Implementation

[0020] The technical features of the present invention will be further described in detail below with reference to the accompanying drawings so that those skilled in the art can understand them.

[0021] An online monitoring method for motor lubrication, the implementation process is as follows: Figure 1 As shown, the specific implementation steps are as follows:

[0022] Step S1: Obtain multi-source historical data and extract multiple segments of multi-dimensional data sequences under normal and abnormal lubrication conditions, and preprocess them to obtain a historical sample set.

[0023] This step aims to collect synchronous data recorded during historical operating cycles by deploying various sensors on the enclosed geared motor, thereby building a sample library covering normal operating conditions and lubrication failure conditions for subsequent model training.

[0024] Specifically, sensing devices are deployed at key measurement points of the enclosed geared motor to synchronously collect vibration data, three-phase current data, and temperature data at a fixed sampling frequency. This process acquires historical data and maintenance records from the long-term operation of multiple enclosed geared motors of the same model. A sample length is set, and based on the maintenance records, multiple segments of equal-length multidimensional data sequences under normal and abnormal lubrication conditions are extracted from the historical data. The extracted historical data is cleaned and standardized to obtain multiple data matrices composed of vibration sequences, three-phase current sequences, and temperature sequences. Each data matrix is ​​considered a historical sample, with historical samples under normal lubrication conditions marked as 0 and historical samples under abnormal lubrication conditions marked as 1. The set of all historical samples and their corresponding labels is considered the historical sample set.

[0025] The sampling frequency can be set to 1kHz, and the sample length can be set to 10,000 sampling points, i.e., the sampling duration is 10 seconds. The selection of the sample length should ensure that the historical samples can cover a sufficient number of mechanical operation cycles and current cycles for frequency domain analysis, and also reflect the cumulative change of temperature over a meaningful period of time, thus forming an effective state observation unit. The specific deployment of the sensing devices is as follows: vibration data is collected by an industrial piezoelectric accelerometer installed radially on the bearing housing at the motor drive end; three-phase current data are collected by a current transformer installed on the three-phase power supply line at the motor drive end; and temperature data is collected by a platinum resistance temperature sensor mounted on the outer ring mounting surface of the bearing at the drive end. The specific preprocessing method is as follows: short-time missing values ​​are filled by linear interpolation; transient outliers are identified and removed by a method based on amplitude thresholds; and the three-phase current data are filtered by a 50Hz power frequency notch filter. Each historical sample is a five-dimensional time-series data matrix.

[0026] In addition, to construct the historical sample set required for supervised learning, each historical sample is assigned a lubrication state label. The specific operation can be carried out by experts in the relevant fields, who can manually analyze the historical operating data, maintenance records and performance degradation reports of the corresponding equipment and manually add the corresponding labels. Moreover, when constructing the historical sample set, it should be ensured that the number of historical samples under normal and abnormal lubrication states extracted from the historical data is roughly balanced, so as to avoid performance deviation caused by class imbalance during subsequent model training.

[0027] Step S2: Based on the three-phase current sequence in the historical samples, extract the load fluctuation characteristics to generate window adjustment index, and dynamically determine the analysis window division scheme for each historical sample.

[0028] Since current signals can reflect load conditions, when current data is stable, it means that the load conditions are stable, and a longer analysis window is suitable for obtaining the statistical characteristics of the vibration signal. When current data is unstable, it means that the load conditions fluctuate and change in a complex manner, so the length of the analysis window needs to be shortened to focus on transient characteristics and avoid confusion between data of different states. Therefore, this step aims to dynamically determine the optimal window length based on the fluctuation characteristics of the current signal, and to adaptively change the length of the analysis window according to the load conditions to eliminate the influence of load condition differences on the lubrication state, thus providing an adaptive analysis window for the subsequent wavelet transform of the vibration signal.

[0029] Specifically, based on the length of the historical samples, the length of the initial analysis window, the length of the maximum analysis window, and the window adjustment step size are set. Any historical sample is selected as the target historical sample. Based on the length of the initial analysis window, the three-phase current sequence in the target historical sample is truncated for the first time to obtain the first initial analysis window of the target historical sample. The standard deviation and information entropy of each phase current in the three-phase current within the first initial analysis window are calculated. Then, the mean standard deviation and mean information entropy of the three-phase current are calculated. Based on human experience, the weight coefficients of the two mean features are set. The sum of the products of the mean standard deviation and the mean information entropy with the corresponding weight coefficients is used as the window adjustment index to obtain the window adjustment index of the first initial analysis window of the target historical sample.

[0030] The initial analysis window length can be set to 100 sampling points, the maximum analysis window length can be set to 1000 sampling points, and the window adjustment step size can be set to 50 sampling points. The change in the standard deviation of the three-phase current can directly reflect the degree of fluctuation of the load torque. Its increase indicates that the mechanical transmission resistance fluctuation is severe and the non-stationarity is enhanced due to poor lubrication. The information entropy of the three-phase current can be used as a quantitative indicator of the complexity and randomness of the three-phase current waveform. Its increase indicates that the ordered components in the current signal are reduced and the randomness is enhanced. It can reflect the intensification of nonlinearity in the transmission process and the dispersion of spectral energy caused by lubrication deterioration.

[0031] Furthermore, abnormal lubrication conditions can lead to mechanical transmission jamming, load disturbances, and a more chaotic distribution of current data with a higher information entropy value. Conversely, when lubrication is good, mechanical operation and transmission are smoother, load conditions are more stable, current data distribution is more regular, and the information entropy value is lower. Therefore, the information entropy value is more sensitive to lubrication conditions. Thus, the weighting coefficient for the mean standard deviation of three-phase current can be set to 0.3, and the weighting coefficient for the mean information entropy of three-phase current can be set to 0.7, with the sum of the weighting coefficients of the two mean features being 1. The window adjustment index comprehensively reflects the fluctuation intensity of the motor load torque and the disorder of the current signal within the initial analysis window. A higher value indicates a more unstable operating condition and a stronger correlation with potential lubrication anomalies. Therefore, a shorter analysis window should be used for subsequent analysis to more accurately capture the changing characteristics of the lubrication condition.

[0032] Specifically, the calculation process of the three-phase current information entropy is as follows: extract the maximum and minimum values ​​of a certain phase current sequence segment within the initial analysis window, divide the value interval formed by the maximum and minimum values ​​into multiple sub-intervals of equal width, for example, 10 sub-intervals, count the number of sampling points in each sub-interval, and use the ratio between the number of sampling points in each sub-interval and the length of the initial analysis window as the probability of the corresponding sub-interval. Based on the existing information entropy calculation formula, the information entropy value corresponding to a certain phase current within the initial analysis window is obtained. Similarly, the information entropy values ​​corresponding to the three phase currents within the initial analysis window are obtained.

[0033] Furthermore, based on the length of the entire three-phase current sequence in the historical samples, the window adjustment index corresponding to each historical sample under normal lubrication conditions is calculated, and its average value is used as the adjustment threshold. For the window adjustment index of the first initial analysis window of the target historical sample, when it is less than the adjustment threshold, the sum of the initial analysis window and the window adjustment step size is used as the adaptive analysis window. The window adjustment index of the adaptive analysis window is calculated iteratively and compared with the adjustment threshold. The window length is dynamically adjusted until the maximum analysis window length is reached, or the window adjustment index is greater than or equal to the adjustment threshold. The length of the maximum analysis window or the length of the corresponding adaptive analysis window is used as the optimal partition length of the first analysis window of the target historical sample. Similarly, the partitioning scheme of the next analysis window of the target historical sample is obtained in turn. By traversing all historical samples, the analysis window partitioning scheme of each historical sample is obtained.

[0034] Specifically, when dividing the last analysis window of the target historical sample, if the remaining length of the three-phase current sequence is less than or equal to the sum of the initial analysis window and the window adjustment step size, the remaining length is directly used as the division length of the last analysis window. That is, when the remaining length of the three-phase current sequence is 50, 100, or 150, the remaining sampling points are directly used as the last analysis window.

[0035] Step S3: Analyze the multidimensional data based on the analysis windows divided from each historical sample, reconstruct the lubrication feature sequence corresponding to each analysis window of each historical sample and merge them to obtain a historical lubrication sample set.

[0036] This step aims to extract and enhance the feature components directly related to the lubrication state from the complex raw vibration data. By introducing current features that characterize load fluctuations and temperature features that characterize heat accumulation as physical guides, within the adaptive analysis window defined in step S2, the frequency bands in the vibration data that are sensitive to the lubrication state are dynamically identified and reconstructed into a continuous feature signal, thereby providing a high signal-to-noise ratio input for the subsequent lubrication state assessment model.

[0037] Specifically, any historical sample is selected as the target historical sample, and any analysis window of the target historical sample is selected as the target analysis window. Wavelet packet decomposition is performed on the vibration sequence segment within the target analysis window to obtain wavelet coefficient sequences of multiple sub-frequency bands. Then, Hilbert transform is performed to obtain the energy envelope signal sequence of each sub-frequency band. The square root of the sum of the squares of the three element values ​​at the same sampling time in the three-phase current sequence of the target historical sample is calculated sequentially to obtain the current guiding signal sequence corresponding to the target historical sample. The Pearson correlation coefficient between the energy envelope signal sequence of each sub-frequency band within the target analysis window and the corresponding current guiding signal sequence segment, as well as the Spearman correlation coefficient between the energy envelope signal sequence of each sub-frequency band within the target analysis window and the corresponding temperature sequence segment, are calculated to obtain the Pearson correlation coefficient characteristics and Spearman correlation coefficient characteristics of each sub-frequency band within the target analysis window.

[0038] Among them, the current-guided signal sequence is generated by synthesizing the instantaneous amplitude of the spatial vector based on the three-phase current sequence, which can effectively suppress the interference caused by the three-phase current imbalance, thus more purely representing the instantaneous change of the load torque; the specific operation of wavelet packet decomposition can be carried out by performing 4-level wavelet packet decomposition using the db4 wavelet basis function to obtain the wavelet coefficient sequence of 16 sub-bands; the calculation process of Pearson correlation coefficient and Spearman correlation coefficient is existing technology. The value of Pearson correlation coefficient can characterize the correlation strength between the vibration energy of the sub-band and the instantaneous fluctuation of the load, quantifying the linear correlation and synchronicity between the instantaneous amplitude changes of the vibration signal and the current signal. The larger its absolute value, the deeper the vibration energy of the corresponding sub-band is modulated by the instantaneous fluctuation of the load; the value of Spearman correlation coefficient can quantify the consistency of the changing trend between the vibration energy of the sub-band and the temperature signal. The larger its absolute value, the stronger the correlation between the long-term trend of the vibration energy of the corresponding sub-band and the temperature rise trend of the equipment.

[0039] Furthermore, based on practical application scenarios, a balanced weight is set for the two coefficient features. The sum of the products of the absolute values ​​of the two coefficient features and their corresponding balanced weights is used as the comprehensive weight of the corresponding sub-band within the target analysis window. The comprehensive weights of each sub-band are summed and normalized. The normalized comprehensive weights are then used to weight the wavelet coefficient sequences of the corresponding sub-bands. The weighted wavelet coefficient sequences of each sub-band are superimposed and fused to obtain the composite wavelet coefficient sequence corresponding to the target analysis window. The composite wavelet coefficient sequence is subjected to inverse wavelet packet transform to obtain the lubrication feature sequence corresponding to the target analysis window. Similarly, the lubrication feature sequences corresponding to each analysis window within the target historical samples are obtained and merged in chronological order. The lubrication feature sequence corresponding to the entire target historical sample is taken as a historical lubrication sample. All historical samples are traversed and combined with the corresponding labels to obtain the historical lubrication sample set.

[0040] The balancing weights are used to adjust the contribution ratios of current correlation and temperature correlation, and can be set to 0.6 and 0.4 respectively, with the sum of the two balancing weights being 1. The larger the value of the comprehensive weight, the stronger the dual physical correlation between the corresponding sub-band and the lubrication state, that is, the deeper the corresponding sub-band is affected by both load fluctuation modulation and heat accumulation, and should be given a higher weight in subsequent reconstruction. The purpose of normalizing the comprehensive weight is to ensure that the comprehensive weights of each sub-band can be compared and weighted on the same benchmark. The composite wavelet coefficient sequence represents the vibration components most relevant to the lubrication state within the corresponding analysis window. The lubrication feature sequence is a reconstruction from multi-source original signals to a single-dimensional time-series feature signal focused on the lubrication state, which significantly improves the signal-to-noise ratio of lubrication-related feature components and provides a direct and efficient data foundation for subsequent training of accurate lubrication state assessment based on deep learning.

[0041] Step S4: Construct a lubrication condition assessment model and train it based on a historical lubrication sample set.

[0042] This step aims to train a long short-term memory network model using a historical lubrication sample set to establish an accurate and intelligent evaluation mapping relationship from lubrication characteristics to lubrication status, providing a core decision algorithm for the final lubrication status diagnosis.

[0043] Specifically, a lubrication state assessment model is constructed based on a long short-term memory network. Historical lubrication samples are used as the input layer, and the corresponding label values ​​are used as the output layer. The Softmax function is used as the activation function of the model output layer, and the cross-entropy loss function is used as the loss function of the model. The lubrication state assessment model is trained based on the historical lubrication sample set until the loss function converges, thus obtaining the trained lubrication state assessment model.

[0044] The core of the lubrication condition assessment model is a two-layer long short-term memory network layer used to extract temporal features, followed by a Dropout layer to prevent overfitting.

[0045] Step S5: Collect multidimensional data at the current moment in real time to construct the current lubrication sample, and monitor the lubrication status in real time based on the trained lubrication status evaluation model.

[0046] This step aims to process the real-time acquired data into a current lubrication sample according to the aforementioned processing chain, and input it into the trained lubrication status assessment model to achieve real-time, automatic diagnosis and early warning of the lubrication status of the gear reducer motor.

[0047] Specifically, multidimensional data is synchronously collected starting from the moment the enclosed geared motor starts running. Based on the length of historical samples, data of equal length is collected and preprocessed to obtain the first current sample. The length of data collected within 1 second is used as the sliding step size based on the sampling frequency, and the current sample is updated through a sliding window. The load fluctuation characteristics corresponding to the current sample are extracted, and the analysis window partitioning scheme of the current sample is dynamically determined. The same calculation process as historical lubrication samples is used to obtain the current lubrication sample corresponding to the current sample. The current lubrication sample is input into the trained lubrication state evaluation model, and the probability that the lubrication state is abnormal within the time period corresponding to the current sample is output. A judgment threshold is set based on the accuracy requirements in the actual application scenario. The judgment threshold can be set to 0.6. When the output probability is greater than the judgment threshold, the corresponding early warning mechanism is triggered.

[0048] When the sampling frequency is set to 1kHz and the length of the historical sample is 10,000 sampling points, 10 seconds of time-series data is collected first as the first current sample, and the current sample is updated every 1 second after the first current sample.

[0049] This invention also discloses an online monitoring system for motor lubrication, used to implement the aforementioned online monitoring method for motor lubrication, the system structure of which is as follows: Figure 2 As shown, it includes: a processor, a memory, a communication interface, a data acquisition device, and an alarm device. The processor stores computer program instructions for implementing the above-mentioned online monitoring method for motor lubrication. The communication interface is communicatively connected to the data acquisition device and the alarm device. The data acquisition device includes an acceleration sensor installed on the bearing housing at the motor drive end, a current transformer installed on the three-phase power supply line at the motor drive end, and a platinum resistance temperature sensor installed on the mounting surface of the outer ring of the bearing at the drive end. The alarm device includes an audible and visual alarm that issues on-site prompts and a WiFi module for sending alarm notification information to users.

[0050] The embodiments included in this invention are merely descriptions of preferred embodiments of the invention and are not limited to the precise structures described above and shown in the accompanying drawings. Various modifications and changes can be made without departing from the scope of protection. Any variations and improvements made by those skilled in the art to the technical solutions of this invention without departing from the design concept of this invention should fall within the scope of protection of this invention.

Claims

1. An online monitoring method for motor lubrication, characterized in that: Multi-source historical data is acquired and multiple segments of multi-dimensional data sequences under normal and abnormal lubrication conditions are extracted and preprocessed to obtain a historical sample set. Based on the three-phase current sequence in the historical samples, load fluctuation characteristics are extracted to generate window adjustment indexes, and the analysis window division scheme of each historical sample is dynamically determined. The multidimensional data is analyzed based on the analysis windows divided from each historical sample. The lubrication feature sequences corresponding to each analysis window of each historical sample are reconstructed and merged to obtain a historical lubrication sample set. A lubrication status assessment model is constructed and trained based on the historical lubrication sample set. The system collects multidimensional data in real time to construct the current lubrication sample and monitors the lubrication status in real time based on the trained lubrication status evaluation model. Each historical sample contains a vibration sequence, a three-phase current sequence, and a temperature sequence. Any historical sample is selected as the target historical sample. Based on the length of the historical samples, the length of the initial analysis window, the length of the maximum analysis window, and the window adjustment step size are set. Based on the length of the initial analysis window, the three-phase current sequence in the target historical sample is truncated for the first time to obtain the first initial analysis window of the target historical sample. The standard deviation and information entropy of each phase current in the three-phase current within the first initial analysis window are calculated. Then, the mean standard deviation and mean information entropy of the three-phase current are calculated. Based on human experience, the weight coefficients of the two mean features are set. The sum of the products of the mean standard deviation and the mean information entropy with the corresponding weight coefficients is used as the window adjustment index to obtain the window adjustment index of the first initial analysis window of the target historical sample. Based on the length of the entire three-phase current sequence in the historical samples, the window adjustment index corresponding to each historical sample under normal lubrication conditions is calculated, and its average value is used as the adjustment threshold. For the window adjustment index of the first initial analysis window of the target historical sample, when it is less than the adjustment threshold, the sum of the initial analysis window and the window adjustment step size is used as the adaptive analysis window. The window adjustment index of the adaptive analysis window is calculated iteratively and compared with the adjustment threshold. The window length is dynamically adjusted until the maximum analysis window length is reached, or the window adjustment index is greater than or equal to the adjustment threshold. The length of the maximum analysis window or the length of the corresponding adaptive analysis window is used as the optimal partition length of the first analysis window of the target historical sample. Similarly, the partitioning scheme of the next analysis window of the target historical sample is obtained in turn. All historical samples are traversed to obtain the analysis window partitioning scheme of each historical sample. Any analysis window of the target historical sample is selected as the target analysis window. Wavelet packet decomposition is performed on the vibration sequence segment within the target analysis window. The comprehensive weight of each sub-frequency band within the target analysis window is calculated. The wavelet coefficient sequences of each sub-frequency band are weighted using the normalized comprehensive weights and then superimposed and fused to obtain the composite wavelet coefficient sequence corresponding to the target analysis window. The composite wavelet coefficient sequence is subjected to inverse wavelet packet transform to obtain the lubrication feature sequence corresponding to the target analysis window. Similarly, the lubrication feature sequences corresponding to each analysis window within the target historical sample are obtained and merged in chronological order. The lubrication feature sequence corresponding to the entire target historical sample is taken as a historical lubrication sample.

2. The online monitoring method for motor lubrication according to claim 1, characterized in that, The process of acquiring multi-source historical data and extracting multiple segments of multidimensional data sequences under normal and abnormal lubrication conditions, followed by preprocessing to obtain a historical sample set, includes: deploying sensing devices at key measurement points of the enclosed gear reducer motor to synchronously collect vibration data, three-phase current data, and temperature data at a fixed sampling frequency; acquiring long-term historical data and maintenance records of multiple enclosed gear reducers of the same model; setting the sample length and, based on the maintenance records, extracting multiple segments of multidimensional data sequences of equal length under normal and abnormal lubrication conditions from the historical data; cleaning the extracted historical data and standardizing the cleaned historical data to obtain multiple data matrices composed of vibration sequences, three-phase current sequences, and temperature sequences; using each data matrix as a historical sample; marking historical samples under normal lubrication conditions as 0 and historical samples under abnormal lubrication conditions as 1; and using the set of all historical samples and their corresponding labels as the historical sample set.

3. The online monitoring method for motor lubrication according to claim 2, characterized in that, The process of reconstructing and merging the lubrication feature sequences corresponding to each analysis window of each historical sample includes: performing wavelet packet decomposition on the vibration sequence segment within the target analysis window to obtain wavelet coefficient sequences of multiple sub-frequency bands, and then performing Hilbert transform to obtain the energy envelope signal sequence of each sub-frequency band; sequentially calculating the square root of the sum of the squares of the three element values ​​at the same sampling time in the three-phase current sequence of the target historical sample to obtain the current guiding signal sequence corresponding to the target historical sample; calculating the Pearson correlation coefficient between the energy envelope signal sequence of each sub-frequency band within the target analysis window and the corresponding current guiding signal sequence segment, as well as the Spearman correlation coefficient between the energy envelope signal sequence of each sub-frequency band within the target analysis window and the corresponding temperature sequence segment, to obtain the Pearson correlation coefficient characteristics and Spearman correlation coefficient characteristics of each sub-frequency band within the target analysis window.

4. The online monitoring method for motor lubrication according to claim 3, characterized in that, The process of reconstructing and merging the lubrication feature sequences corresponding to each analysis window of each historical sample further includes: setting a balanced weight for two coefficient features based on the actual application scenario; using the sum of the products of the absolute values ​​of the two coefficient features and their corresponding balanced weights as the comprehensive weight of the corresponding sub-band within the target analysis window; summing and normalizing the comprehensive weights of each sub-band to weightedly fuse the wavelet coefficient sequences of each sub-band to obtain the composite wavelet coefficient sequence corresponding to the target analysis window; then, based on the inverse wavelet packet transform, obtaining the lubrication feature sequences corresponding to each analysis window within the target historical sample; and finally merging them to obtain the historical lubrication samples corresponding to the target historical sample; and traversing all historical samples and combining them with the corresponding labels to obtain the historical lubrication sample set.

5. The online monitoring method for motor lubrication according to any one of claims 2 to 4, characterized in that, The process of constructing a lubrication state assessment model and training it based on a historical lubrication sample set includes: constructing a lubrication state assessment model based on a long short-term memory network, using the Softmax function as the activation function of the model's output layer, using the cross-entropy loss function as the model's loss function, and training the lubrication state assessment model based on the historical lubrication sample set until the loss function converges, thereby obtaining the trained lubrication state assessment model.

6. The online monitoring method for motor lubrication according to any one of claims 2 to 4, characterized in that, The process of collecting multidimensional data at the current moment to construct the current lubrication sample and monitoring the lubrication status in real time based on the trained lubrication status evaluation model includes: synchronously collecting multidimensional data from the start-up time of the enclosed gear reducer motor; collecting data of equal length based on the length of historical samples and preprocessing it to obtain the first current sample; using the length of data collected within 1 second as the sliding step size based on the sampling frequency and updating the current sample through a sliding window; extracting the load fluctuation characteristics corresponding to the current sample; dynamically determining the analysis window partitioning scheme for the current sample; using the same calculation process as historical lubrication samples to obtain the current lubrication sample corresponding to the current sample; inputting the current lubrication sample into the trained lubrication status evaluation model; outputting the probability that the lubrication status is abnormal within the time period corresponding to the current sample; setting a judgment threshold based on the accuracy requirements in the actual application scenario; and triggering a corresponding early warning mechanism when the output probability is greater than the judgment threshold.

7. An online monitoring system for motor lubrication, characterized in that: It includes a processor, a memory, a communication interface, a data acquisition device, and an alarm device. The processor stores computer program instructions for implementing the online monitoring method for motor lubrication as described in any one of claims 1 to 6. The communication interface is communicatively connected to the data acquisition device and the alarm device.

Images