Mechanical state monitoring method and system based on holographic time sequence probability distribution

Abstract

The invention provides a mechanical state monitoring method and system based on holographic time sequence probability distribution, and the method comprises the steps: collecting and preprocessing single-section multi-channel vibration data in a normal operation period, and obtaining the amplitudes and phases of different frequency multiplications; two-dimensional holographic spectrums are constructed respectively, corresponding holographic parameters are calculated, the holographic parameters are arranged to carry out windowing processing and average operation, and time sequence holographic parameter samples are obtained; respectively inputting each group of samples into an autoregression probability distribution operation module and a time sequence frequency spectrum self-coding reconstruction module of a pre-established mechanical state monitoring operation model for forward operation, and performing model optimization training on parameters of each operation module based on an operation result; and performing operation based on the vibration data in the to-be-detected period by using the trained model, and outputting early warning information when a set requirement is met. According to the scheme, the defects that in the prior art, the professional evaluation requirement is high, data acquisition and operation depend on fault data and the monitoring precision is insufficient can be overcome, and early-stage operation early warning of a large machine is reliably achieved.

Description

Technical Field

[0001] The present invention relates to the field of large - scale machinery state prediction and health management, and particularly to a mechanical state monitoring method and system based on holographic time - series probability distribution. Background Art

[0002] Rotating machinery is widely used in industries such as chemical engineering and electric power. With the development of science and technology, rotating machinery is developing towards high - speed, large - scale and automated directions. The size of equipment is getting larger and the structure is getting more complex. Correspondingly, a large number of problems such as strength, structure, vibration, reliability, etc. appear during the application process. Once a fault occurs during its operation, it will cause huge losses and very serious consequences. Therefore, researching on the fault diagnosis and treatment of rotating machinery to ensure the safe and reliable operation of the unit is of great significance.

[0003] In the prior art, there are some methods for processing the vibration information of rotating machinery by applying the holographic spectrum technology, which are mainly applied to the fault monitoring and diagnosis of large - scale rotating machinery such as air compressors and gas turbines. Its principle is to fuse key information such as vibration amplitude and phase in a cross - section in the frequency domain to fully display the overall picture of rotor vibration. For example, the patent document CN201610814507.X provides a model - based pseudo - subsynchronous fault holographic diagnosis method for rotating machinery. It constructs a pseudo - subsynchronous vibration fault model induced by axial vibration, proposes two assumptions in the model: setting a V - shaped groove defect on the surface of the vibration measuring band and applying axial excitation in the axial direction. Then, based on the model, the mechanism analysis of the generation of pseudo - subsynchronous vibration fault signals is realized. Finally, the two - dimensional holographic spectrum is used to analyze the fault signals, and the holographic spectrum characteristics at the subsynchronous fault frequency are constructed to realize the identification of pseudo - subsynchronous vibration faults of rotating machinery. It can be seen that in the existing research, the holographic spectrum and its related derivative technologies, although to a certain extent enrich the means of analyzing rotor vibration signals and improve the ability of fault diagnosis. However, its use process requires field engineers to have the ability of "recognizing pictures and interpreting spectra", and specific diagnostic conclusions are obtained by analyzing the holographic spectrum diagrams, which has relatively high requirements for the professionalism of operators and poor practicability. Moreover, the analysis process often needs to combine knowledge and prior information in multiple disciplinary fields such as mechanical structure and signal processing, relying too much on professional knowledge and expert experience, and unable to achieve efficient and extensive application and promotion.

[0004] Therefore, providing a reliable and efficient intelligent mechanical operation state analysis method based on the holographic spectrum technology is an urgently needed research direction in the field.

[0005] The information disclosed in the background art part of the present invention is only intended to deepen the understanding of the general background art of the present invention, and should not be regarded as an admission or any form of suggestion that this information constitutes the prior art known to those skilled in the art. Summary of the Invention

[0006] To solve the above problems, the present invention provides a mechanical state monitoring method based on holographic time-series probability distribution. This method can effectively overcome the defects of high professional requirements for evaluation in the prior art, data acquisition and operation relying on fault data, and insufficient monitoring accuracy, and can efficiently and reliably realize early operation warning of large machinery; preferably, in one embodiment, the method includes:

[0007] Training data acquisition step: Select a set normal operation cycle as the training data cycle for the target mechanical to be measured, and collect vibration data of multiple channels of a single cross-section within the training data cycle to form a training source data set;

[0008] Data processing step: Preprocess each training source data according to the set principle, and obtain the amplitude and phase information of different frequency multiples;

[0009] Holographic parameter calculation step: Construct a two-dimensional holographic spectrum based on the amplitude and phase information of different frequency multiples respectively, and then calculate the corresponding holographic parameters. After arranging the holographic parameters of different frequency multiples in combination with time information, perform windowing processing and average operation to obtain a time-series holographic parameter matrix as a time-series holographic parameter sample;

[0010] Model training step: Construct a mechanical state monitoring operation model including an autoregressive probability distribution operation module and a time-series spectrum auto-encoding reconstruction module, input the holographic parameter sample to calculate the objective function value, and perform cyclic optimization of the parameters of different operation modules in combination with the set index by backpropagation until the objective function value meets the set requirements;

[0011] Operation detection execution step: Collect vibration data of multiple channels of a single cross-section within the measurement period for the target mechanical. After preprocessing, determine the corresponding time-series holographic parameters through the holographic parameter calculation step, input them into the trained mechanical state monitoring operation model, calculate the corresponding objective function value as the monitoring index, and analyze the operation state monitoring result of the mechanical to be measured in combination with the set index threshold. When the set requirements are met, output a warning message.

[0012] Optionally, in one embodiment, in the training data acquisition step, select a set normal operation cycle from historical operation data, or set the normal operation cycle after maintenance as the training data cycle, and collect vibration data of channels X and Y of a single cross-section within the training data cycle to form a training source data set.

[0013] In an alternative embodiment, in the data processing step, the process of preprocessing each training source data includes:

[0014] Perform mean removal processing on the vibration data of different channels respectively;

[0015] Furthermore, perform fast Fourier transform on the vibration data of different channels respectively to obtain the corresponding frequency domain distribution information;

[0016] Calculate the amplitude and phase information corresponding to different multiple frequencies of the vibration data in combination with the mechanical rotation speed information, where the different multiple frequencies include the rotation frequency, the second harmonic frequency, and the third harmonic frequency.

[0017] Further, in one embodiment, in the holographic parameter operation step, construct the corresponding two-dimensional holographic spectrum according to the following formula based on the amplitude and phase information of different multiple frequencies:

[0018]

[0019] Determine the sine coefficient and cosine coefficient in the two-dimensional holographic spectrum algorithm as holographic parameters;

[0020] In the formula, i represents the multiple frequency type in the vibration signal, i = 1 represents the rotation frequency, i = 2 represents the second harmonic frequency, i = 3 represents the third harmonic frequency, A is the vibration amplitude in the x direction, α- is the vibration phase in the x direction, B is the vibration amplitude in the y direction, β is the vibration phase in the y direction; ω represents the frequency, t represents the time; sx i , sy i respectively represent the sine holographic parameter corresponding to the i-th multiple frequency two-dimensional holographic spectrum of the X-channel data of the current source data and the sine holographic parameter corresponding to the i-th multiple frequency two-dimensional holographic spectrum of the Y-channel data, cx i , cy i respectively represent the cosine holographic parameter corresponding to the i-th multiple frequency two-dimensional holographic spectrum of the X-channel data of the current source data and the cosine holographic parameter corresponding to the i-th multiple frequency two-dimensional holographic spectrum of the Y-channel data.

[0021] Further, in one embodiment, in the holographic parameter operation step, the process of performing windowing processing and averaging operation includes:

[0022] After arranging the holographic parameters of the rotation frequency, the second harmonic frequency, and the third harmonic frequency at each moment to form a holographic parameter vector, perform windowing processing on the holographic parameters at multiple moments according to the set windowing parameters, and perform time-series averaging processing on the holographic parameters within a single time window along the time axis.

[0023] In a preferred embodiment, in the holographic parameter operation step, perform time-series averaging processing on the holographic parameters according to the following formula:

[0024]

[0025] In the formula, H T represents the result of holographic parameter averaging processing, T ω represents the window length, t represents the time, represents the holographic parameter at the t - T ω moment of the n-th multiple frequency data.

[0026] Optionally, in one embodiment, in the model training step, based on a mechanical state monitoring operation model including an autoregressive probability distribution operation module and a time-frequency spectrum autoencoding reconstruction module, an objective function is designed by combining the loss terms of different operation modules. Each group of time-series holographic parameter samples obtained is respectively input into the autoregressive probability distribution operation module and the time-frequency spectrum autoencoding reconstruction module for forward operation to calculate the objective function value.

[0027] Further, in one embodiment, in the model training step, in the mechanical state monitoring operation model, the autoregressive probability distribution operation module constructs a masked fully connected layer on the basis of a classical fully connected layer, and through a sequential masking strategy, each output of the parameterized probability density estimator only depends on the input of its previous holographic parameters.

[0028] In one embodiment, in the model training step, in the mechanical state monitoring operation model, the time-frequency spectrum autoencoding module is responsible for the frequency-domain autoencoding reconstruction of vibration data, performs a fast Fourier transform on the vibration signals of different channels of the target machine, and performs a time windowing process on the frequency-domain distribution to calculate the reconstructed time-frequency spectrum. Among them, both the encoder and the decoder adopt a long short-term memory network structure.

[0029] In an alternative embodiment, in the model training step, the objective function is designed as follows:

[0030] l = α holo l holo + α rec l rec

[0031] Where

[0032] In the formula, l represents the target operation output, α represents the weight coefficient of different penalty terms, F T represents the windowed frequency-domain distribution, H represents the holographic parameter, θ g represents the parameters of the parameterized probability density estimator, and g(·) represents the parameterized probability density estimator.

[0033] Based on other aspects of the method described in any one or more of the above embodiments, the present invention also provides a storage medium, on which program code for implementing the method described in any one or more of the above embodiments is stored.

[0034] Based on the application aspect of the method described in any one or more of the above embodiments, the present invention also provides a mechanical state monitoring system based on holographic time-series probability distribution, and this system executes the method described in any one or more of the above embodiments.

[0035] Compared with the closest prior art, the present invention also has the following beneficial effects:

[0036] A mechanical state monitoring method and system based on holographic time - series probability distribution provided by the present invention collect vibration data of multiple channels of a single cross - section during a normal operation cycle and perform pre - processing to obtain the amplitudes and phases of different harmonics. Before subsequent operations, pre - processing the collected source data can effectively avoid abnormal data from participating in the operations, and perform operations based on high - quality source data, thereby fundamentally improving the accuracy of the operation results at each stage. Moreover, the source data comprehensively considers the spatio - temporal distribution characteristics of vibration data of multiple channels of rotating machinery;

[0037] Construct two - dimensional holographic spectra respectively and calculate the corresponding holographic parameters. Carry out windowing processing and average operation on the arranged holographic parameters to obtain time - series holographic parameter samples; Use holographic spectrum analysis technology to intelligently fuse the amplitude, phase, and frequency information of multiple channels of a single cross - section of rotating machinery, provide unique sample data for the mechanical state monitoring operation model, conduct intelligent research on holographic spectrum technology from the perspective of data - driven, realize the effective extended application of holographic spectrum technology, and expand the application scope of classical holographic theory from the dimension of time - series data - driven;

[0038] Input each group of samples into the forward operations of the autoregressive probability distribution operation module and the time - series spectrum auto - encoder reconstruction module of the pre - built mechanical state monitoring operation model respectively, and optimize and train the parameters of each operation module based on the operation results; Furthermore, use the trained model to perform operations based on the vibration data during the period to be measured, and output a warning message when the set requirements are met. Train the constructed model with normal data. When training the model, only normal data of the equipment is required, getting rid of the limitations on fault data samples and labels; In addition, through a parametric probability density estimator, the tracking and processing of the dynamic evolution law of the holographic parameter hidden space in the whole life cycle of rotating machinery are realized. It can realize early warning of the operation state of large units in industries such as petrochemical industry, judge whether key components such as rotors are in a normal or abnormal state, and has higher accuracy and operation robustness than traditional solutions; Provide guidance for the predictive maintenance of large - scale equipment, ensure the long - term stable operation of the device and the safe progress of industrial production, and has important engineering application value and relatively broad application prospects.

[0039] Other features and advantages of the present invention will be described in the subsequent specification, and, in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in the specification, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0041] Figure 1 It is a schematic flow chart of the mechanical state monitoring method based on holographic time series probability distribution provided by an embodiment of the present invention;

[0042] Figure 2 It is a schematic diagram of the execution principle of the mechanical state monitoring method based on holographic time series probability distribution provided by an embodiment of the present invention;

[0043] Figure 3 It is a schematic diagram of the data distribution of different channels in the whole life cycle of a synthesis gas unit in the mechanical state monitoring method based on holographic time series probability distribution provided by an embodiment of the present invention;

[0044] Figure 4 It is a schematic diagram of the detection data distribution of the traditional peak-to-peak analysis method in the mechanical state monitoring method based on holographic time series probability distribution provided by an embodiment of the present invention;

[0045] Figure 5 It is a schematic diagram of the monitoring operation result distribution of the mechanical state monitoring method based on holographic time series probability distribution provided by an embodiment of the present invention;

[0046] Figure 6 It is a schematic diagram of the structure of the mechanical state monitoring system based on holographic time series probability distribution provided by an embodiment of the present invention. Detailed implementation manners

[0047] The following will combine the drawings and embodiments to detail the implementation manners of the present invention, so that the implementers of the present invention can fully understand how to apply technical means to solve technical problems and achieve the implementation process of technical effects, and implement the present invention specifically according to the above implementation process. It should be noted that as long as there is no conflict, each embodiment in the present invention and each feature of each embodiment can be combined with each other, and the formed technical solutions are all within the protection scope of the present invention.

[0048] Although the flow chart describes the operations as sequential processing, many of the operations can be performed in parallel, concurrently, or simultaneously. The order of the operations can be rearranged. The processing can be terminated when its operations are completed, but there can also be additional steps not included in the drawings. The processing can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0049] Computer devices include user devices and network devices. Among them, user devices or clients include, but are not limited to, computers, smartphones, PDAs (Personal Digital Assistants), etc.; network devices include, but are not limited to, a single network server, a server group composed of multiple network servers, or a cloud composed of a large number of computers or network servers based on cloud computing. The computer device can run independently to implement the present invention, or can be connected to a network and implement the present invention through interaction with other computer devices in the network. The network where the computer device is located includes, but is not limited to, the Internet, wide area network, metropolitan area network, local area network, VPN network, etc.

[0050] Here, terms such as "first", "second", etc. may be used to describe various units, but these units should not be limited by these terms. These terms are only used to distinguish one unit from another. The term "and / or" used herein includes any and all combinations of one or more of the listed related items. When a unit is referred to as being "connected" or "coupled" to another unit, it can be directly connected or coupled to the other unit, or there may be an intermediate unit.

[0051] The terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments. Unless the context clearly indicates otherwise, the singular forms "a", "an" used herein are also intended to include the plural. It should also be understood that the terms "comprising" and / or "including" used herein specify the presence of the stated features, integers, steps, operations, units and / or components, and do not exclude the presence or addition of one or more other features, integers, steps, operations, units, components and / or their combinations.

[0052] Rotating machinery is one of the most commonly used machinery in industrial production and has a wide range of applications in industries such as chemical engineering and power. With the development of science and technology, rotating machinery is developing towards high speed, large scale and automation. The size of the equipment is getting larger and the structure is getting more and more complex. Correspondingly, a large number of problems such as strength, structure, vibration, reliability, etc. have emerged in the application process. Once a failure occurs during its operation, it will cause huge losses and very serious consequences. Therefore, researching the fault diagnosis and treatment of rotating machinery to ensure the safe and reliable operation of the unit is of great significance.

[0053] In the prior art, there are some methods for processing the vibration information of rotating machinery by using holographic spectrum technology, which are mainly applied to the fault monitoring and diagnosis of large rotating machinery such as air compressors and gas turbines; the principle is to fuse key information such as vibration amplitude and phase in a cross-section in the frequency domain to fully display the overall picture of rotor vibration. For example, patent text CN202110196401.9 provides a method and system for expressing time-varying holographic features of rotating machinery, which collects two mutually perpendicular radial vibration displacement signals on a selected cross-section of the rotor and collects the rotational speed signal; uses time-varying phase demodulation and time-varying filtering to separate the characteristic orders to be extracted from the two mutually perpendicular radial vibration displacement signals into single time-varying frequency component signals; for each characteristic order, synthesizes the two mutually perpendicular radial vibration displacement signals to construct a three-dimensional helix that changes with time; arranges the three-dimensional helices corresponding to each order according to the order to construct a four-dimensional time-varying holographic spectrum of horizontal radial amplitude, vertical radial amplitude, time, and order; merges the order axis with the horizontal direction displacement axis and expresses the four-dimensional time-varying holographic spectrum as a combination of three-dimensional visual helices for observing changes in mechanical working conditions. Patent document CN202010415571.7 provides a method for diagnosing double-frequency faults of rotating machinery based on three-dimensional holographic difference spectrum. First, it determines the measuring point on the unit shafting that is most sensitive to fault response, analyzes the main exciting frequency and amplitude change trend of this measuring point using a spectral waterfall diagram or a characteristic trend diagram, and combines rules to judge whether the fault type is a double-frequency fault; then calculates the double-frequency three-dimensional holographic spectrum matrix of the shafting under normal state and multiple fault states, and constructs a double-frequency three-dimensional holographic difference spectrum matrix corresponding to the fault state; finally, draws the double-frequency three-dimensional holographic difference spectrum diagram of the shafting under multiple fault states to obtain the pure fault double-frequency ellipse of the entire shafting and the change characteristics of its features during the fault development process, and combines these characteristics to determine the location of the fault source and qualitatively diagnose the double-frequency fault. Patent document CN201610814507.X provides a model-based holographic diagnosis method for pseudo-subharmonic faults of rotating machinery. It constructs a pseudo-subharmonic vibration fault model induced by axial vibration, proposes two assumptions in the model, namely, setting a V-shaped groove defect on the surface of the vibration measuring belt and applying an axial excitation in the axial direction, then conducts a mechanism analysis of the generation of pseudo-subharmonic vibration fault signals based on the model, and finally analyzes the fault signals using a two-dimensional holographic spectrum to construct holographic spectrum characteristics at the subharmonic fault frequency to achieve the identification of pseudo-subharmonic vibration faults of rotating machinery.

[0054] Based on the above description, it can be seen that in existing research, holographic spectra and related technologies derived therefrom, although to a certain extent enrich the means of analyzing rotor vibration signals and improve the ability of fault diagnosis, their use requires on-site engineers to have a considerable degree of "ability to recognize diagrams and diagnose spectra". Specific diagnostic conclusions are obtained through the analysis of holographic diagrams, which requires a high level of professionalism for operators, has poor practicability, and the analysis process often needs to combine knowledge and prior information in multiple disciplinary fields such as mechanical structures and signal processing. It is overly dependent on professional knowledge and expert experience and cannot achieve efficient and extensive application and promotion.

[0055] Therefore, it can be seen that providing a reliable and efficient intelligent mechanical operating state analysis method based on holographic spectrum technology is an urgently needed research direction in the field.

[0056] To solve the above problems, the present invention provides a mechanical state monitoring method and system based on holographic time-series probability distribution. Early warning of the operating state of rotating machinery is realized through the analysis and calculation of the conditional probability distribution of holographic parameters in time series. The amplitude, phase, and frequency information of multiple channels in a single cross-section of rotating machinery are intelligently fused, and the tracking and processing of the dynamic evolution law of the holographic parameter hidden space in the entire life cycle of rotating machinery are realized through a parameterized probability density estimator, overcoming the defects of high professional technology requirements and insufficient practicability in the prior art, and effectively improving the accuracy and robustness of the early detection and analysis of mechanical operating states.

[0057] Next, the detailed process of the method of the embodiment of the present invention will be described in detail based on the accompanying drawings. The steps shown in the flowchart of the accompanying drawings can be executed in a computer system including a set of computer-executable instructions. Although the logical order of the steps is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0058] Embodiment 1

[0059] Figure 1 shows a schematic flowchart of a mechanical state monitoring method based on holographic time-series probability distribution provided by Embodiment 1 of the present invention. Referring to Figure 1 it can be seen that the method includes the following steps.

[0060] Training data acquisition step: For the target mechanical equipment to be measured, a set normal operation period is selected as the training data period, and the vibration data of multiple channels in a single cross-section within the training data period are collected to form a training source data set;

[0061] Data processing step: Preprocess each training source data according to the set principle, and obtain the amplitude and phase information of different frequency multiples;

[0062] Holographic parameter calculation steps: construct a two-dimensional holographic spectrum based on the amplitude and phase information of different frequency multiples, and then calculate the corresponding holographic parameters. After arranging the holographic parameters of different frequency multiples in combination with the time information, perform windowing and averaging operations to obtain the time series holographic parameter matrix as the time series holographic parameter sample.

[0063] Model training steps: construct a mechanical condition monitoring operation model including an autoregressive probability distribution operation module and a time series spectrum autoencoder reconstruction module, input holographic parameter samples to calculate the objective function value, and perform cyclic optimization of the parameters of different operation modules by reverse transfer of the set indicators until the objective function value meets the set requirements;

[0064] Operation detection execution steps: collect vibration data of multiple channels of a single cross-section within the test period for the target machinery, determine the corresponding time series holographic parameters through the holographic parameter calculation step after preprocessing, input the trained mechanical state monitoring calculation model, calculate the corresponding objective function value as the monitoring indicator, and analyze the operation state monitoring results of the machine to be tested in combination with the set indicator threshold, and output warning information when the set requirements are met.

[0065] The mechanical state monitoring method based on holographic time series probability distribution provided by the embodiment of the present invention intelligently integrates the amplitude, phase and frequency information of multiple channels of a single cross-section of a rotating machinery, thereby realizing intelligent analysis and processing of holographic parameters of large-scale rotating machinery. A method for calculating the time series conditional probability distribution of holographic parameters of rotating machinery vibration data is proposed, and the tracking and processing of the dynamic evolution law of the latent space of holographic parameters of the rotating machinery throughout its life cycle is realized through a parameterized probability density estimator. The spatiotemporal characteristics of the data are comprehensively considered, combined with the structural characteristics of the rotating machinery, and the early warning indicators are calculated based on the trained and optimized mechanical state monitoring operation model, which has better warning accuracy and robustness than the solutions in the prior art.

[0066] In actual application, the training data collection step is first performed, and the normal operating cycle of the target machine to be tested is selected and set as the training data cycle, and the vibration data of multiple channels of a single cross section within the training data cycle are collected to form a training source data set.

[0067] In a preferred embodiment, in the training data collection step, vibration data of the X channel and the Y channel of a single cross section within a training data period are collected to form a training source data set.

[0068] Optionally, for the selected target machinery to be tested, the operating cycle after reliable maintenance can be selected as the normal operating cycle, and the vibration data of multiple channels of a single section within the cycle can be collected; or, the vibration data of multiple channels of a single section with a known normal operating cycle can be extracted from historical operating data as training source data.

[0069] Further, before performing subsequent operations based on the training source data, the source data is preprocessed to avoid the influence of abnormal data on the subsequent operation effect and convert the source data into the required information. Therefore, the data processing step is executed to preprocess each training source data according to the set principle to obtain the amplitude and phase information of different frequency multiples.

[0070] In a preferred embodiment, in the data processing step, the process of preprocessing each training source data includes:

[0071] Perform mean removal processing on the vibration data of different channels respectively;

[0072] Furthermore, perform fast Fourier transform on the vibration data of different channels respectively to obtain the corresponding frequency domain distribution information;

[0073] Calculate the amplitude and phase information corresponding to different frequency multiples of the vibration data in combination with the mechanical rotation speed information, and the different frequency multiples include the rotation frequency, the second harmonic frequency, and the third harmonic frequency.

[0074] Figure 2 FIG. shows the schematic diagram of the execution principle of the mechanical state monitoring method based on the holographic time series probability distribution provided by the embodiment of the present invention. In actual application, for each group of source data collected, perform mean removal processing on the X-channel data and Y-channel data of each group of samples respectively; perform fast Fourier transform (FFT) on the X-channel and Y-channel of each group of data respectively to obtain the frequency domain distribution; furthermore, calculate the amplitude and phase information corresponding to the rotation frequency, the second harmonic frequency, and the third harmonic frequency according to the unit rotation speed information, as Figure 2 shown.

[0075] Further, execute the holographic parameter operation step, construct the corresponding two-dimensional holographic spectrum based on the amplitude and phase information of different frequency multiples respectively, and then calculate its holographic parameters. After arranging the holographic parameters of different frequency multiples in combination with the time information, perform windowing processing and average operation to obtain the time series holographic parameter matrix as the time series holographic parameter sample.

[0076] In actual application, construct the rotation frequency two-dimensional holographic spectrum, the second harmonic frequency two-dimensional holographic spectrum, and the third harmonic frequency two-dimensional holographic spectrum according to the amplitude and phase information corresponding to the rotation frequency, the second harmonic frequency, and the third harmonic frequency; furthermore, calculate the rotation frequency holographic parameter, the second harmonic frequency holographic parameter, and the third harmonic frequency holographic parameter.

[0077] Preferably, in one embodiment, in the holographic parameter operation step, the corresponding two-dimensional holographic spectrum is constructed based on the amplitude and phase information of different frequency multiples according to the following formula:

[0078]

[0079] Based on this, the sine coefficient and cosine coefficient in the two-dimensional holographic spectrum algorithm are determined as holographic parameters; setting the sine coefficient and cosine coefficient in the above formula is the corresponding holographic coefficient H, as follows.

[0080] H t =[sx1, sy1, cx1, cy1, sx2, sy2, cx2, cy2, …, sx i , sy i , cx i , cy i

[0081] In the formula, H t represents the holographic coefficient set, i represents the multiple frequency type in the vibration signal, i = 1 represents the rotation frequency, i = 2 represents the second harmonic frequency, i = 3 represents the third harmonic frequency, sx i , sy i respectively represent the sine holographic parameter corresponding to the two-dimensional holographic spectrum of the i-th multiple frequency of the X-channel data of the current source data and the sine holographic parameter corresponding to the two-dimensional holographic spectrum of the i-th multiple frequency of the Y-channel data, cx i , cy i respectively represent the cosine holographic parameter corresponding to the two-dimensional holographic spectrum of the i-th multiple frequency of the X-channel data of the current source data and the cosine holographic parameter corresponding to the two-dimensional holographic spectrum of the i-th multiple frequency of the Y-channel data, ω represents the frequency, and t represents the time.

[0082] In an optional embodiment, in the holographic parameter operation step, the process of performing windowing processing and averaging operation includes:

[0083] After arranging the holographic parameters of the rotation frequency, second harmonic frequency, and third harmonic frequency at each moment to form a holographic parameter vector, windowing processing is performed on the holographic parameters at multiple moments according to the set windowing parameters, and the holographic parameters within a single time window are averaged along the time axis.

[0084] Specifically, in an optional embodiment, in the holographic parameter operation step, the holographic parameters are averaged along the time axis according to the following formula:

[0085]

[0086] In the formula, H T represents the result of windowing and averaging processing of the holographic parameters, T ω represents the window length, t represents the time, represents the n-th multiple frequency data t - T ω ​Holographic parameters at each moment; in practical applications, the rotational frequency, second harmonic frequency, and third harmonic frequency holographic parameters at each moment are arranged to form a holographic parameter vector. The holographic parameters at multiple moments are windowed, and the holographic parameters within a single time window are averaged along the time axis to obtain a time-series holographic parameter matrix. Preferably, the window length and the overlapping parameter of adjacent windows involved in the process are set according to the sample requirements. Generally, the values of the window length and the overlapping parameter of adjacent windows are related to the data length and the performance of the processing hardware. Generally, the longer the data, the longer the window length required, and the larger the overlapping parameter of adjacent windows, the more samples are finally formed, and correspondingly, the more challenging it is to test the performance of the processing hardware.

[0087] Next, perform the model training step. Based on a mechanical state monitoring operation model that includes an autoregressive probability distribution operation module and a time-series spectrum autoencoding reconstruction module, design an objective function by combining the loss terms of different operation modules. Input each group of obtained time-series holographic parameter samples into the autoregressive probability distribution operation module and the time-series spectrum autoencoding reconstruction module respectively for forward operations, calculate the objective function value, and then perform cyclic optimization of the parameters of different operation modules by backpropagation in combination with the set indicators until the objective function value meets the set requirements.

[0088] In a preferred embodiment, in the model training step, the autoregressive probability distribution operation module of the mechanical state monitoring operation model constructs a masked fully connected layer based on the classical fully connected layer, and makes each output of the parameterized probability density estimator only depend on the input of its previous holographic parameters through a sequential masking strategy.

[0089] On the other hand, in the model training step, set the time-series spectrum autoencoding module of the mechanical state monitoring operation model to be responsible for the frequency-domain autoencoding reconstruction of vibration data. Perform a fast Fourier transform on the vibration signals of the X channel and the Y channel, and perform time-series windowing on the frequency-domain distribution to calculate the reconstructed time-series spectrum. Among them, both the encoder and the decoder adopt a long short-term memory network structure.

[0090] In the mechanical state monitoring operation model, input the obtained time-series holographic parameter samples into a parameterized probability density estimator for conditional probability operations to make the time-series holographic parameter samples have better interpretability and recognition accuracy. To achieve this goal, in an alternative embodiment, construct a masked fully connected layer based on the classical fully connected layer:

[0091]

[0092] In the formula, T represents the operation time, d represents the number of samples input at the current time, represents the output of the fully connected layer of the neural network, c and l represent the order of the neural network connection layer, and e and j represent the order of the neurons in the fully connected layer.

[0093] With this sequential masking strategy, each output of the parametric probability density estimator can depend only on the input of its previous holographic parameters. The loss term l of the sequential holographic parameter autoregressive probability distribution module holo is set to:

[0094]

[0095] In the formula, H represents the holographic parameter, θ g represents the parameters of the parametric probability density estimator, g(·) represents the parametric probability density estimator, K represents the total number of layers of the parametric probability density estimator, and k represents the current layer number of the parametric probability density estimator.

[0096] Furthermore, the obtained sequential holographic parameter samples are input into the sequential spectrum autoencoder module;

[0097] The sequential spectrum autoencoder module is mainly responsible for the frequency-domain autoencoding reconstruction of vibration data, performs fast Fourier transform on the vibration signals of the X channel and Y channel, and performs sequential windowing processing on the frequency-domain distribution. The window length T ω is the same as the window length in the aforementioned holographic sequential windowing process.

[0098] The sequential spectrum signal is input into the autoencoder for reconstruction, and the reconstructed sequential spectrum is calculated Both the encoder and decoder adopt the long short-term memory network structure.

[0099] The loss term of the shaft center orbit discrimination module is

[0100] In the model training step, the objective function is designed as follows:

[0101] l = α holo l holo + α rec l rec

[0102] where

[0103] In the formula, l represents the target operation output, α represents the weight coefficient of different penalty terms, F T represents the windowed frequency-domain distribution, θ g represents the parameters of the parametric probability density estimator, and g(·) represents the parametric probability density estimator.

[0104] In the actual training process, the obtained sequential holographic parameters corresponding to the normal state vibration data are input into the constructed mechanical state monitoring operation model for forward propagation operation. The model output error is calculated according to the objective function (comprehensive loss function) and the model parameters are optimized by backpropagation, that is, further according to the objective function l = αholo l holo +α rec l rec The calculation results are used to realize back propagation, update the parameters of the autoregressive probability distribution module and the time series spectrum autoencoder module. In actual operation, it is preferred to set α holo =0.01,α rec =1.

[0105] The above forward transfer operation and back propagation optimization process are repeated until the objective function value obtained by the operation meets the set optimization index and the number of cycles reaches the predetermined requirement, indicating that the current mechanical condition monitoring operation model training is completed.

[0106] Further execute the operation detection execution step, collect vibration data of multiple channels of a single cross-section within the test period for the target machinery, determine the corresponding time series holographic parameters through the holographic parameter calculation step after preprocessing, input the trained mechanical state monitoring calculation model, calculate the corresponding objective function value as the monitoring index, and analyze the operation state monitoring results of the machine to be tested in combination with the set index threshold, and output warning information when the set requirements are met.

[0107] In actual application, the test cycle of the current target machinery can be selected according to needs, and the vibration data of multiple channels of a single section can be collected to obtain the corresponding time series holographic parameters. The trained mechanical state monitoring operation model can be input to calculate the corresponding objective function value as the monitoring indicator, which is compared with the setting to determine whether an early warning is needed.

[0108] By adopting the technical means in the embodiments of the present invention, early warning of the operating status of large units in industries such as petrochemicals can be achieved, and key components such as rotors can be judged to be in a normal or abnormal state, providing guidance for predictive maintenance of mechanical equipment, ensuring long-term stable operation of the device and safe industrial production, which has important engineering application value and a relatively broad application prospect. In addition, it should be noted that when training the mechanical state monitoring calculation model in the scheme of the present invention, only the vibration data in the normal state needs to be used as the basis for the sample, which gets rid of the dependence and limitation on the fault data samples and labels.

[0109] The present invention is further described below in conjunction with examples of implementation. The scope of the present invention is not limited by the examples, but is set forth in the claims.

[0110] The solution of this patent will be illustrated by a syngas unit combined with the known full life cycle data of a certain chemical enterprise in China. First, the training data acquisition step is performed. For this syngas unit, the set normal operation cycle is selected as the training data cycle, and the vibration data of multiple channels in a single cross-section within the training data cycle is collected to form a training source data set. In actual operation, since the full life cycle data of the syngas unit is known, the vibration data of multiple channels in a single cross-section within the full life cycle can be directly collected. Specifically, the current rotational speed of the syngas unit is about 10,500 r / min. The vibration data of channels X and Y at the non-coupled end of the high-pressure cylinder of the compressor is selected for analysis, and the sampling frequency is 1024 Hz, obtaining a total of 7,588 sets of data. The time domain waveform of the full life cycle data is as Figure 3 shown. The vibration amplitude of the compressor continues to increase. After shutdown for maintenance, fouling is found at the rotor impeller.

[0111] Furthermore, the collected data is preprocessed according to the set principle to obtain the amplitude and phase information of different multiple frequencies. Specifically, the following operations are performed on all 7,588 sets of data for preprocessing and time series data construction:

[0112] (1) The mean value of the data of channel X and channel Y in each group of samples is removed respectively;

[0113] (2) The fast Fourier transform is performed on channel X and channel Y of each group of data respectively to obtain the corresponding frequency domain distribution;

[0114] (3) Furthermore, the amplitude and phase information corresponding to the rotation frequency, double frequency, and triple frequency are calculated according to the rotational speed information of the unit.

[0115] Furthermore, the holographic parameter calculation step is performed. Two-dimensional holographic spectra are constructed respectively based on the amplitude and phase information of different multiple frequencies, and then the corresponding holographic parameters are calculated. After arranging the holographic parameters of different multiple frequencies in combination with the time information, windowing processing and averaging operation are carried out to obtain a time series holographic parameter matrix as the time series holographic parameter sample.

[0116] In actual application, a rotation frequency two-dimensional holographic spectrum, a double frequency two-dimensional holographic spectrum, and a triple frequency two-dimensional holographic spectrum are constructed for all 7,588 sets of data; the rotation frequency holographic parameter, the double frequency holographic parameter, and the triple frequency holographic parameter are calculated;

[0117] In an optional embodiment, according to the structure and operation characteristics of large rotating machinery, substituting the amplitude and phase of the vibration data of channels X and Y perpendicular to each other in a single cross-section at time t into the following formula can obtain the two-dimensional holographic spectrum,

[0118]

[0119] where i = 1, 2, 3... represents the rotation frequency component and multiple frequency components in the vibration signal;

[0120] The sine coefficients and cosine coefficients in the above formula are the holographic coefficients H described in this patent.

[0121] H t = [sx1, sy1, cx1, cy1, sx2, sy2, cx2, cy2, …, sx i , sy i , cx i , cy i ;

[0122] Furthermore, the rotating frequency, double frequency, and triple frequency holographic parameters at each moment are arranged to form a holographic parameter vector. Windowing processing is performed on the holographic parameters at multiple moments, and the holographic parameters within a single time window are averaged along the time axis to obtain a time-sequence holographic parameter matrix. Preferably, the window length is selected as 40, the adjacent windows overlap by 39, and a total of 7549 groups of time-sequence holographic parameter samples are obtained.

[0123] Among them, the time-sequence averaging process is performed on the extracted holographic coefficient H according to the following formula:

[0124]

[0125] Furthermore, the spectral data of the X channel and Y channel at each moment are spliced to form a dual-channel spectral vector. Windowing processing is performed on the spectra at multiple moments. Preferably, the window length is selected as 40, the adjacent windows overlap by 39, and a total of 7549 groups of time-sequence spectral samples are obtained.

[0126] Next, a set of time-sequence spectral samples of a set size is selected for the model training step. A mechanical state monitoring operation model including an autoregressive probability distribution operation module and a time-sequence spectral autoencoder reconstruction module is constructed. The objective function is designed by combining the loss terms of different operation modules. The selected groups of time-sequence holographic parameter samples are respectively input into the autoregressive probability distribution operation module and the time-sequence spectral autoencoder reconstruction module for forward operation, the objective function value is calculated, and the parameters of different operation modules are cyclically optimized by backpropagation in combination with the set indicators until the objective function value meets the set requirements.

[0127] In an optional embodiment, the training process of the mechanical state monitoring operation model specifically includes:

[0128] (1) Training set construction. The first 1500 groups of samples are regarded as normal state samples, and their corresponding time-sequence holographic parameters are input into the autoregressive probability distribution module;

[0129] The time-sequence holographic parameters are input into a parametric probability density estimator for conditional probability operation. To achieve this goal, a masked fully connected layer is constructed on the basis of the classical fully connected layer.

[0130]

[0131] Through this sequential masking strategy, each output of the parametric probability density estimator only depends on the input of its previous holographic parameters. The loss term of the temporal holographic parameter autoregressive probability distribution module is set as

[0132]

[0133] where H represents the holographic parameter, θ g represents the parameter of the parametric probability density estimator, g(·) represents the parametric probability density estimator, K represents the total number of layers of the parametric probability density estimator, and k represents the current layer number of the parametric probability density estimator.

[0134] (2) Input the temporal frequency spectrum corresponding to the first 1500 groups of samples into the temporal frequency spectrum autoencoder module;

[0135] The temporal frequency spectrum autoencoder module is mainly responsible for the frequency domain autoencoding reconstruction of vibration data, performs fast Fourier transform on the vibration signals of the X channel and Y channel, and performs temporal windowing processing on the frequency domain distribution. The window length T ω is the same as the window length in the aforementioned holographic temporal windowing processing.

[0136] Input the temporal frequency spectrum signal into the autoencoder for reconstruction, and calculate the reconstructed temporal frequency spectrum Both the encoder and decoder adopt the long short-term memory network structure.

[0137] The loss term of the shaft orbit discrimination module is

[0138] (3) Forward operation to obtain the parametric probability distribution g(H; θ g ) and the reconstructed temporal frequency spectrum

[0139] In the training stage, input the vibration data in the X direction and Y direction in the normal state into the constructed mechanical state monitoring operation model for forward propagation operation, and calculate the model output error according to the loss function.

[0140] (4) Backpropagation. According to the optimization objective function l = α holo l holo + α rec l rec Update the parameters of the autoregressive probability distribution module and the temporal frequency spectrum autoencoder module according to the values of and the set index requirements. Preferably, α holo = 0.01, α rec = 1;

[0141] (5) Repeat the above steps (3) to (4) until the objective function value meets the set index requirements and the number of executions reaches the predetermined number of loops.

[0142] Furthermore, the full life cycle data of the syngas unit can be input into the trained mechanical operation state monitoring operation model for forward transfer operation, and the mechanical operation state monitoring operation model can be tested in combination with the actual operation state data to verify the effectiveness of the proposed model and the accuracy of the operation results.

[0143] In practical applications, the testing of the large rotating machinery operation state monitoring operation model specifically includes:

[0144] (1) Take all 7549 groups of samples as the test data set, input the corresponding time series holographic parameters into the trained autoregressive probability distribution module, and input the corresponding time series spectrum into the trained time series spectrum autoencoder module;

[0145] (2) Calculate forward to obtain the parametric probability distribution g(H; θ g ) and the reconstructed time series spectrum

[0146] (3) Perform index operations, using the optimized objective function l = α holo l holo + α rec l rec as the early warning index, using 5 times the average value of the index in the normal state as the warning threshold line, and giving an alarm when the index exceeds the threshold line, it can be found that the distribution of the operation results is consistent with the actual operation state of the synthesis unit.

[0147] On the other hand, to verify the advantages of the operation results of the present invention, the model output results are compared and analyzed with the peak-to-peak value of the original vibration signal of the same unit. As Figure 4 shown are the peak-to-peak values of the vibration in the X channel and Y channel of the non-coupled end of the high-pressure cylinder of the syngas unit during its full life cycle. It can be seen from the figure that the peak-to-peak value fluctuates continuously before 4000 points and only shows an obvious upward trend after 4000 points. It is difficult to directly and accurately judge the operation state of the unit from the peak-to-peak value. The distribution of the index results calculated by the method proposed in the present invention is as Figure 5 shown. At about 2100 samples, the constructed index undergoes an obvious jump, exceeding the set threshold, and an alarm message is sent. Before that, the constructed index is relatively stable and has better robustness than the peak-to-peak value. At about 4000 samples, the constructed index shows an obvious upward trend again, indicating that the unit enters the next degradation stage. The index constructed in the present invention shows an obvious three-stage jump during the full life cycle and has better early warning accuracy.

[0148] For each of the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should be aware that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0149] It should be noted that in other embodiments of the present invention, the method can also be combined with one or more of the above embodiments to obtain a new mechanical state monitoring method based on holographic time series probability distribution, so as to realize early abnormal monitoring and optimization of the operating state of large rotating machinery.

[0150] Embodiment 2:

[0151] It should be noted that based on the method in any one or more of the above embodiments of the present invention, the present invention also provides a storage medium, on which program code capable of implementing the method described in any one or more of the above embodiments is stored. When the code is executed by an operating system, it can implement the mechanical state monitoring method based on holographic time series probability distribution as described above.

[0152] Embodiment 3:

[0153] In the above embodiments disclosed by the present invention, the method is described in detail. The method of the present invention can be implemented by various forms of devices or systems. Therefore, based on other aspects of the method in any one or more of the above embodiments, the present invention also provides a mechanical state monitoring system based on holographic time series probability distribution, which is used to execute the mechanical state monitoring method based on holographic time series probability distribution described in any one or more of the above embodiments. Specific embodiments are given below for detailed description.

[0154] Specifically, Figure 6 shows a schematic structural diagram of the mechanical state monitoring system based on holographic time series probability distribution provided in the embodiments of the present invention, as Figure 6 shown, the system includes:

[0155] A training data acquisition module, configured to select a set normal operation cycle as the training data cycle for the mechanical to be measured, and collect vibration data of multiple channels of a single cross-section within the training data cycle to form a training source data set;

[0156] A data processing module, configured to preprocess each training source data according to a set principle, and obtain amplitude and phase information of different multiples of frequency;

[0157] A holographic parameter calculation module is configured to construct a two-dimensional holographic spectrum based on the amplitude and phase information of different frequency multiples, and then calculate the corresponding holographic parameters. After arranging the holographic parameters of different frequency multiples in combination with the time information, windowing processing and averaging operation are performed to obtain a time series holographic parameter matrix as a time series holographic parameter sample;

[0158] A model training module is configured to construct a mechanical condition monitoring operation model including an autoregressive probability distribution operation module and a time series spectrum autoencoder reconstruction module, input holographic parameter samples to calculate the target function value, and perform cyclic optimization on the parameters of different operation modules in combination with the set index reverse transfer until the target function value meets the set requirements;

[0159] The operation detection execution module is configured to collect vibration data of multiple channels of a single cross-section within the test period for the target machinery, determine the corresponding time series holographic parameters by enabling the holographic parameter operation module after preprocessing, input the trained mechanical state monitoring operation model, calculate the corresponding objective function value as the monitoring indicator, analyze the operation state monitoring results of the machine to be tested in combination with the set indicator threshold, and output warning information when the set requirements are met.

[0160] Optionally, in one embodiment, the training data acquisition module is configured to: select a normal operating cycle from historical operating data, or set a normal operating cycle after maintenance as a training data cycle, and collect vibration data of a single-section X channel and Y channel within the training data cycle to form a training source data set.

[0161] In a selected embodiment, the data processing module pre-processes each training source data according to the following operations:

[0162] De-average the vibration data of different channels respectively;

[0163] Then, fast Fourier transform is performed on the vibration data of different channels to obtain the corresponding frequency domain distribution information;

[0164] The amplitude and phase information corresponding to different multiple frequencies of the vibration data are calculated in combination with the mechanical rotation speed information, and the different multiple frequencies include rotation frequency, double frequency and triple frequency.

[0165] Furthermore, in one embodiment, the holographic parameter calculation module constructs the corresponding two-dimensional holographic spectrum according to the following formula based on the amplitude and phase information of different frequency doublings:

[0166]

[0167] Based on this, the sine coefficient and cosine coefficient in the two-dimensional holographic spectrum algorithm are determined as holographic parameters;

[0168] Wherein, i represents the multiple frequency type in the vibration signal, i = 1 represents the rotation frequency, i = 2 represents the double frequency, i = 3 represents the triple frequency, A is the vibration amplitude in the x direction, α- is the vibration phase in the x direction, B is the vibration amplitude in the y direction, β is the vibration phase in the y direction; ω represents the frequency, t represents the time; sx i , sy i respectively represent the sine holographic parameter corresponding to the two-dimensional holographic spectrum of the i-th multiple frequency of the X-channel data of the current source data and the sine holographic parameter corresponding to the two-dimensional holographic spectrum of the i-th multiple frequency of the Y-channel data, cx i , cy i respectively represent the cosine holographic parameter corresponding to the two-dimensional holographic spectrum of the i-th multiple frequency of the X-channel data of the current source data and the cosine holographic parameter corresponding to the two-dimensional holographic spectrum of the i-th multiple frequency of the Y-channel data.

[0169] Furthermore, in one embodiment, the holographic parameter operation module performs windowing processing and averaging operation according to the following operations:

[0170] After arranging the holographic parameters of the rotation frequency, double frequency, and triple frequency at each moment to form a holographic parameter vector, windowing processing is performed on the holographic parameters at multiple moments according to the set windowing parameters, and the holographic parameters within a single time window are averaged along the time axis.

[0171] In a preferred embodiment, the holographic parameter operation module performs time series averaging processing on the holographic parameters according to the following formula:

[0172]

[0173] Wherein, H T represents the result of the holographic parameter averaging processing, T ω represents the window length, t represents the time, represents the holographic parameter at the time t - T of the n-th multiple frequency data ω moment.

[0174] Optionally, in one embodiment, the model training module is configured to design an objective function based on the mechanical state monitoring operation model including an autoregressive probability distribution operation module and a time series spectrum autoencoder reconstruction module, and combine the loss terms of different operation modules. The obtained groups of time series holographic parameter samples are respectively input into the autoregressive probability distribution operation module and the time series spectrum autoencoder reconstruction module for forward operation, and the objective function value is calculated.

[0175] Furthermore, in one embodiment, the model training module sets the autoregressive probability distribution operation module in the mechanical state monitoring operation model to construct a masked fully connected layer on the basis of the classical fully connected layer, and makes each output of the parameterized probability density estimator only depend on the input of its previous holographic parameter through the sequential masking strategy.

[0176] In one embodiment, the time-series spectrum auto-encoding module in the mechanical state monitoring operation model set by the model training module is responsible for the frequency-domain auto-encoding reconstruction of vibration data, performing fast Fourier transform on the vibration signals of different channels of the target machine, and performing time-series windowing processing on the frequency-domain distribution to calculate the reconstructed time-series spectrum. Among them, both the encoder and the decoder adopt long short-term memory network structures.

[0177] In an alternative embodiment, the model training module designs the objective function as follows:

[0178] l = α holo l holo + α rec l rec

[0179] Wherein,

[0180] In the formula, l represents the target operation output, α holo and α rec respectively represent the weight coefficients of different penalty terms, F T represents the windowed frequency-domain distribution, H represents the holographic parameter, θ g represents the parameters of the parametric probability density estimator, and g(·) represents the parametric probability density estimator.

[0181] In the mechanical state monitoring system based on holographic time-series probability distribution provided by the embodiments of the present invention, each module or unit structure can operate independently or in combination according to actual data processing requirements and model operation requirements to achieve corresponding technical effects.

[0182] It should be understood that the embodiments disclosed in the present invention are not limited to the specific structures, processing steps or materials disclosed herein, but should extend to equivalent alternatives of these features understood by those of ordinary skill in the relevant art. It should also be understood that the terms used herein are only for the purpose of describing specific embodiments and do not imply limitation.

[0183] The phrase "one embodiment" mentioned in the specification means that the specific features, structures or features described in connection with the embodiment are included in at least one embodiment of the present invention. Therefore, the phrase "one embodiment" appearing throughout the specification does not necessarily refer to the same embodiment.

[0184] Although the disclosed embodiments of the present invention are as above, the content described is only an embodiment adopted for the convenience of understanding the present invention and is not used to limit the present invention. Any person skilled in the technical field to which the present invention pertains can make any modifications and changes in the form and details of the implementation without departing from the spirit and scope disclosed by the present invention. However, the scope of patent protection of the present invention shall still be subject to the scope defined by the appended claims.

Claims

1. A mechanical state monitoring method based on holographic time series probability distribution, characterized in that, The method comprises: Training data collection steps: for the target machine to be tested, a normal operating cycle is selected and set as the training data cycle, and vibration data of multiple channels of a single cross section within the training data cycle are collected to form a training source data set; Data processing steps: pre-process each training source data according to the set principles, and obtain the amplitude and phase information of different frequency multiples; Holographic parameter calculation steps: construct a two-dimensional holographic spectrum based on the amplitude and phase information of different frequency multiples, and then calculate the corresponding holographic parameters. After arranging the holographic parameters of different frequency multiples in combination with the time information, perform windowing and averaging operations to obtain a time series holographic parameter matrix as a time series holographic parameter sample; Model training steps: construct a mechanical condition monitoring operation model including an autoregressive probability distribution operation module and a time series spectrum autoencoder reconstruction module, input holographic parameter samples to calculate the objective function value, and perform cyclic optimization of the parameters of different operation modules by reverse transfer of the set indicators until the objective function value meets the set requirements; Operation detection execution steps: collect vibration data of multiple channels of a single cross-section within the test period for the target machinery, determine the corresponding time series holographic parameters through the holographic parameter calculation step after preprocessing, input the trained mechanical state monitoring calculation model, calculate the corresponding objective function value as the monitoring indicator, and analyze the operation state monitoring results of the machine to be tested in combination with the set indicator threshold, and output warning information when the set requirements are met.

2. The method according to claim 1, characterized in that, In the training data collection step, a normal operation cycle is selected from historical operation data, or a normal operation cycle is set after maintenance as a training data cycle, and vibration data of a single cross-section X channel and Y channel within the training data cycle are collected to form a training source data set.

3. The method according to claim 1, wherein In the data processing step, the process of preprocessing each training source data includes: De-average the vibration data of different channels respectively; Then, fast Fourier transform is performed on the vibration data of different channels to obtain the corresponding frequency domain distribution information; The amplitude and phase information corresponding to different multiple frequencies of the vibration data are calculated in combination with the mechanical rotation speed information, and the different multiple frequencies include rotation frequency, double frequency and triple frequency.

4. The method according to claim 1, characterized in that, In the holographic parameter calculation step, the corresponding two-dimensional holographic spectrum is constructed according to the following formula based on the amplitude and phase information of different frequency doublings: Determine the sine coefficient and cosine coefficient in the two-dimensional holographic spectrum algorithm as holographic parameters; Wherein, i represents the multiple frequency type in the vibration signal, i = 1 represents the rotation frequency, i = 2 represents the double frequency, i = 3 represents the triple frequency, A is the vibration amplitude in the x direction, α- is the vibration phase in the x direction, B is the vibration amplitude in the y direction, β is the vibration phase in the y direction; ω represents the frequency, t represents the time; sx i , sy i respectively represent the sine holographic parameters corresponding to the two-dimensional holographic spectrum of the i-th multiple frequency of the X-channel data of the current source data and the sine holographic parameters corresponding to the two-dimensional holographic spectrum of the i-th multiple frequency of the Y-channel data, cx i , cy i respectively represent the cosine holographic parameters corresponding to the two-dimensional holographic spectrum of the i-th multiple frequency of the X-channel data of the current source data and the cosine holographic parameters corresponding to the two-dimensional holographic spectrum of the i-th multiple frequency of the Y-channel data.

5. The method according to claim 1, characterized in that, In the holographic parameter calculation step, the process of windowing and averaging includes: After arranging the frequency conversion, double frequency and triple frequency holographic parameters at each moment to form a holographic parameter vector, the multi-moment holographic parameters are windowed according to the set windowing parameters, and the holographic parameters in a single time window are time-series averaged along the time axis.

6. The method according to claim 1, characterized in that, In the holographic parameter calculation step, the holographic parameters are processed by time-series averaging according to the following formula: In the formula, H T represents the result of holographic parameter averaging, T ω represents the window length, t represents the time, represents the holographic parameter at the time of t - T of the n - fold frequency data ω .

7. The method according to claim 1, characterized in that, In the model training step, based on the mechanical condition monitoring operation model including the autoregressive probability distribution operation module and the time series spectrum self-encoding reconstruction module, the objective function is designed in combination with the loss terms of different operation modules, and each group of time series holographic parameter samples obtained are respectively input into the autoregressive probability distribution operation module and the time series spectrum self-encoding reconstruction module for forward operation to calculate the objective function value.

8. The method according to claim 1, wherein In the model training step, in the mechanical state monitoring operation model, the autoregressive probability distribution operation module constructs a masked fully connected layer on the basis of the classical fully connected layer, and makes each output of the parameterized probability density estimator depend only on the input of its previous holographic parameter through the sequential masking strategy.

9. The method according to claim 1, wherein In the model training step, in the mechanical state monitoring operation model, the time series frequency spectrum autoencoder module is responsible for the frequency domain autoencoding reconstruction of vibration data, performs fast Fourier transform on the vibration signals of different channels of the target machine, and performs time series windowing processing on the frequency domain distribution to calculate the reconstructed time series frequency spectrum. Among them, both the encoder and the decoder adopt the long short-term memory network structure.

10. The method according to claim 7, characterized in that, In the model training step, the objective function is designed as follows: l＝α holo l holo +α rec l rec Among them, where, in the formula, l represents the target operation output, α represents the weight coefficient of different penalty terms, F T represents the frequency domain distribution after windowing, H represents the holographic parameter, θ g represents the parameter of the parametric probability density estimator, and g(·) represents the parametric probability density estimator.

11. A storage medium, characterized in that, Program code for implementing the method according to any one of claims 1 to 10 is stored on the storage medium.

12. A mechanical state monitoring system based on holographic time series probability distribution, characterized in that, The system executes the method according to any one of claims 1 to 10.

Images