Rotary machine pump vibration state trend prediction method and system based on edge computing

Abstract

The present application relates to the technical field of vibration prediction, and discloses a rotating machine pump vibration state trend prediction method and system based on edge computing, the method comprising the following steps: S101, calculating a health quantity; S102, calculating a bandwidth change rate; S103, calculating a trend slope estimate; S104, calculating a mechanical body slope; S105, calculating a future prediction endpoint; and S106, calculating a rotating machine pump vibration state trend quantization result. According to the equivalent energy bandwidth of the carrier band, the present application constructs a full-link drive, can strip the electrical pseudo-trend introduced by frequency conversion spectrum expansion, realizes the extraction of pure mechanical health trend through amplitude deviation and slope purification, avoids the trend misjudgment caused by electrical interference, and greatly improves the accuracy and timeliness of rotating machine pump health management.

Description

Technical Field

[0001] This invention relates to the field of vibration prediction technology, and more specifically, to a method and system for predicting the vibration state trend of rotating pumps based on edge computing. Background Technology

[0002] In industrial production, rotating pumps are key fluid transport equipment. To achieve energy saving and speed regulation, most equipment is equipped with variable frequency drive devices. Vibration monitoring is the core means of assessing mechanical health. By analyzing the frequency domain characteristics of vibration signals, faults such as bearing wear and impeller imbalance can be predicted.

[0003] When the variable frequency drive device is working, it uses a high-frequency carrier to achieve pulse width modulation. In order to reduce electromagnetic interference, it also jitters the carrier frequency, that is, spreads the spectrum, so that the originally concentrated carrier energy is expanded into a band and changes slowly with the load temperature rise. This electrical energy will mix in the vibration signal of the rotating pump, forming a pseudo-trend in the frequency domain that overlaps with the mechanical health frequency band.

[0004] Traditional vibration trend prediction methods have significant limitations: 1. They only focus on mechanical characteristic frequency bands, failing to identify electrical interference caused by carrier spread spectrum, easily misjudging electrical pseudo-trends as mechanical degradation trends, leading to operational and maintenance decision-making errors. 2. While complex neural network models employing multi-feature fusion can distinguish interference to some extent, they require high computational power, making real-time calculations at the edge difficult and unsuitable for deployment requirements of local monitoring of industrial equipment. Furthermore, traditional methods often use the mean or standard deviation as a baseline, making them susceptible to interference from instantaneous impacts in vibration signals, further reducing the robustness of trend prediction and severely impacting the accuracy and timeliness of health management of rotating pumps and machinery. Summary of the Invention

[0005] This invention provides a method and system for predicting the vibration state trend of a rotating pump based on edge computing, which solves the technical problems mentioned in the background.

[0006] This invention provides a method for predicting the vibration state trend of a rotating pump based on edge computing, comprising the following steps:

[0007] Step S101: Perform frequency domain transformation on the collected vibration data to obtain power spectral density, statistically analyze the power spectral density distribution shape within a preset carrier band to obtain the carrier band equivalent energy bandwidth, and statistically analyze the energy within a mechanical health band that avoids the carrier band to obtain the health quantity.

[0008] Step S102: Calculate the injection factor based on the ratio of the carrier band equivalent energy bandwidth to its historical baseline, correct the health value through the injection factor to obtain the debiased health value, and calculate the bandwidth change rate based on the carrier band equivalent energy bandwidth at adjacent times.

[0009] Step S103: The time segment composed of the debiased health quantity is used as the time input, and the vector composed of the carrier band equivalent energy bandwidth and its bandwidth change rate is used as the condition input. The vector is fed into a one-dimensional convolutional neural network. The scaling coefficient and translation coefficient generated by the condition input mapping are used to perform channel-wise linear modulation on the network convolutional features. The trend slope estimate is obtained by linear readout.

[0010] Step S104: Calculate the product of the bandwidth change rate and the preset coupling sensitivity coefficient, and subtract the product from the trend slope estimate to obtain the mechanical body slope.

[0011] Step S105: Based on the slope of the mechanical body and the current debiasing health value, perform deterministic extrapolation calculation within a preset extrapolation time to obtain the future prediction endpoint;

[0012] Step S106: Statistically calculate the median baseline and absolute median difference scale of the debiased health quantity within the historical time window, normalize the difference between the future predicted endpoint and the median baseline using the absolute median difference scale, and output the quantitative result of the vibration state trend of the rotating pump.

[0013] This invention provides a vibration state trend prediction system for rotating pumps based on edge computing, comprising:

[0014] The health quantity calculation module is used to perform frequency domain transformation on the collected vibration data to obtain the power spectral density, statistically analyze the power spectral density distribution shape within a preset carrier band to obtain the carrier band equivalent energy bandwidth, and statistically analyze the energy within a mechanical health band that avoids the carrier band to obtain the health quantity.

[0015] The bandwidth change rate calculation module is used to calculate the injection factor based on the ratio of the carrier band equivalent energy bandwidth to its historical baseline, correct the health value through the injection factor to obtain the debiased health value, and calculate the bandwidth change rate based on the carrier band equivalent energy bandwidth at adjacent times.

[0016] The trend slope estimation module takes the time segment composed of the debiased health quantity as the time input, and the vector composed of the carrier band equivalent energy bandwidth and its bandwidth change rate as the condition input, and feeds it into a one-dimensional convolutional neural network. The scaling coefficient and translation coefficient generated by the condition input mapping are used to perform channel-wise linear modulation on the network convolutional features, and the trend slope estimate is obtained by linear readout.

[0017] The mechanical body slope calculation module is used to calculate the product of the bandwidth change rate and the preset coupling sensitivity coefficient, and subtract the product from the trend slope estimate to obtain the mechanical body slope.

[0018] The future prediction endpoint calculation module is used to perform deterministic extrapolation calculations within a preset extrapolation time based on the slope of the mechanical body and the current debiased health value to obtain the future prediction endpoint.

[0019] The vibration state quantification module of the rotating pump is used to statistically analyze the median baseline and absolute median difference scale of the debiased health quantity within the historical time window. It normalizes the difference between the future prediction endpoint and the median baseline using the absolute median difference scale and outputs the quantification result of the vibration state trend of the rotating pump.

[0020] The beneficial effects of this invention are as follows: This invention constructs a full-link drive based on the carrier band's equivalent energy bandwidth, which can strip away the electrical pseudo-trends introduced by frequency conversion spectrum spreading. Through amplitude de-biasing and slope purification, it achieves the extraction of pure mechanical health trends, avoiding trend misjudgments caused by electrical interference. Simultaneously, it adopts a lightweight one-dimensional convolution and linear operation architecture, without complex models or redundant calculations, fully adapting to low-computing-power edge computing scenarios and ensuring the response efficiency of real-time on-site monitoring. Furthermore, it completes trend quantification using the median baseline and absolute median difference scale, possessing strong resistance to outliers. The single output quantification result provides clear trend criteria, offering precise basis for operation and maintenance decisions, significantly improving the accuracy and timeliness of rotating pump health management, and balancing technical precision with practicality in industrial settings. Attached Figure Description

[0021] Figure 1 This is a flowchart of the edge computing-based method for predicting the vibration state trend of a rotating pump according to the present invention.

[0022] Figure 2 This is a schematic diagram of the vibration state trend prediction system for rotating pumps based on edge computing according to the present invention.

[0023] In the diagram: Health quantity calculation module 201, bandwidth change rate calculation module 202, trend slope estimation module 203, mechanical body slope calculation module 204, future prediction endpoint calculation module 205, and rotating pump vibration state quantification module 206. Detailed Implementation

[0024] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0025] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0026] like Figures 1-2 As shown, the edge computing-based method for predicting the vibration state trend of a rotating pump includes the following steps:

[0027] Step S101: Perform frequency domain transformation on the collected vibration data to obtain power spectral density, statistically analyze the power spectral density distribution shape within a preset carrier band to obtain the carrier band equivalent energy bandwidth, and statistically analyze the energy within a mechanical health band that avoids the carrier band to obtain the health quantity.

[0028] Step S102: Calculate the injection factor based on the ratio of the carrier band equivalent energy bandwidth to its historical baseline, correct the health value through the injection factor to obtain the debiased health value, and calculate the bandwidth change rate based on the carrier band equivalent energy bandwidth at adjacent times.

[0029] Step S103: The time segment composed of the debiased health quantity is used as the time input, and the vector composed of the carrier band equivalent energy bandwidth and its bandwidth change rate is used as the condition input. The vector is fed into a one-dimensional convolutional neural network. The scaling coefficient and translation coefficient generated by the condition input mapping are used to perform channel-wise linear modulation on the network convolutional features. The trend slope estimate is obtained by linear readout.

[0030] Step S104: Calculate the product of the bandwidth change rate and the preset coupling sensitivity coefficient, and subtract the product from the trend slope estimate to obtain the mechanical body slope.

[0031] Step S105: Based on the slope of the mechanical body and the current debiasing health value, perform deterministic extrapolation calculation within a preset extrapolation time to obtain the future prediction endpoint;

[0032] Step S106: Statistically calculate the median baseline and absolute median difference scale of the debiased health quantity within the historical time window, normalize the difference between the future predicted endpoint and the median baseline using the absolute median difference scale, and output the quantitative result of the vibration state trend of the rotating pump.

[0033] In one embodiment of the present invention, the current analysis window and frequency resolution are determined according to the sampling frequency parameter, the analysis window length parameter and the sliding step size parameter. A window function is applied to the vibration data in the current analysis window and a fast Fourier transform is performed. The square of the modulus of the transform result is normalized according to the analysis window length parameter and the sampling frequency parameter to obtain the power spectral density.

[0034] It should be noted that the sampling frequency parameter represents the number of times the vibration signal of the rotating pump is collected per unit time, measured in Hertz; the analysis window length parameter represents the total number of vibration data sample points selected in each frequency domain analysis; the sliding step size parameter represents the number of vibration data sample points between two adjacent consecutive analysis windows; and the current analysis window represents the data segment extracted from the continuous vibration data for the current frequency domain analysis based on the analysis window length parameter and the sliding step size parameter. Frequency resolution represents the frequency interval between two adjacent frequency points in the frequency domain analysis, reflecting the ability to distinguish signals of different frequencies; frequency resolution is equal to the sampling frequency parameter divided by the analysis window length parameter, that is, the frequency difference between adjacent frequency points is obtained by the ratio of the number of samples per unit time to the number of sample points in a single analysis.

[0035] It should be noted that the window function represents the weighting function applied to the data in the current analysis window before the Fast Fourier Transform (FFT). It is used to reduce frequency domain leakage caused by aperiodic truncation of the data. Common types include Hanning windows and Hamming windows; this invention preferentially chooses the Hanning window, but if the high-frequency components are weak, the Hamming window can be selected. Vibration data represents the time-domain vibration signal of the rotating pump collected by sensors. The Fast Fourier Transform (FFT) is used to convert the time-domain vibration data of the current analysis window into a frequency-domain complex signal, obtaining the complex result corresponding to each frequency point. The square of the modulus of the transform result represents the value obtained by first calculating the modulus (the square root of the square of the real part plus the square of the imaginary part) for each frequency point obtained by the FFT, and then squaring the modulus, reflecting the signal energy intensity at the corresponding frequency point. The power spectral density represents the power within a unit frequency range, used to describe the power distribution of the signal at different frequencies; the power spectral density is equal to the square of the modulus of the transform result multiplied by 2, and then divided by the product of the analysis window length parameter and the sampling frequency parameter, i.e., normalizing the frequency domain energy intensity to obtain the power value per unit frequency.

[0036] In one embodiment of the present invention, a carrier band is defined based on the carrier center frequency parameter and the carrier neighborhood half bandwidth parameter. The square of the sum of the power spectral density values ​​of all frequency points in the carrier band is calculated, the sum of the squares of the power spectral density values ​​of all frequency points in the carrier band is calculated, the square of the sum of the values ​​is divided by the sum of the squares of the values ​​and multiplied by the frequency resolution to obtain the equivalent energy bandwidth of the carrier band.

[0037] Define a mechanical health band that does not overlap with the carrier band, calculate the sum of the products of the power spectral density values ​​and the frequency resolution of all frequency points within the mechanical health band, and perform a square root operation on the sum to obtain the health value.

[0038] It should be noted that the carrier center frequency parameter represents the center frequency of the high-frequency carrier signal of the variable frequency drive device for the rotating pump. The carrier center frequency parameter is equal to the rated carrier frequency of the variable frequency drive device, which can be found in the device's product manual. If not specified, the frequency corresponding to the peak power spectral density of the vibration data of the rotating pump under no-load operation is the carrier center frequency parameter value. The carrier neighborhood half-bandwidth parameter represents the frequency width extending in both high and low frequency directions based on the carrier center frequency, used to determine the boundary range of the carrier band. For low-voltage variable frequency drives (380V), the carrier center frequency parameter is taken as 100 Hz; for high-voltage variable frequency drives (10kV and above), the carrier center frequency parameter is taken as 150 Hz. In general scenarios, a range of 50 to 200 Hz can be selected. The carrier band represents a specific frequency range around the center frequency of the carrier. It is used to locate electrical interference signals introduced by the frequency converter and is the interval for calculating the equivalent energy bandwidth of the carrier band. The lower limit frequency of the carrier band is equal to the carrier center frequency parameter minus the carrier neighborhood half bandwidth parameter, and the upper limit frequency is equal to the carrier center frequency parameter plus the carrier neighborhood half bandwidth parameter. That is, the carrier band is a continuous frequency segment from the lower limit frequency to the upper limit frequency.

[0039] It should be noted that the sum of the power spectral density values ​​at all frequency points within the carrier band reflects the total power within the carrier band; the sum of the squares of the power spectral density values ​​at all frequency points within the carrier band reflects the concentration of power distribution within the carrier band; the carrier band equivalent energy bandwidth represents an indicator of the concentration or dispersion of power spectral density distribution within the carrier band, measured in Hertz, and is used to represent the spread spectrum intensity of the carrier signal of the frequency converter. The mechanical health band represents the frequency domain range that does not overlap with the carrier band, used to locate the signal frequency components reflecting the health status of the rotating pump's mechanical components; the lower limit frequency of the mechanical health band is equal to the upper limit frequency of the carrier band plus 100 Hertz; the upper limit frequency for centrifugal pumps is set at 2000 Hertz, and the upper limit frequency for centrifugal fans is set at 1500 Hertz to ensure coverage of the characteristic frequencies of components such as bearings and impellers. The health quantity represents a quantitative indicator of the mechanical health status of the rotating pump.

[0040] Specifically, the equivalent energy bandwidth of the waveband at time t. The calculation formula is as follows:

[0041]

[0042] in Indicates frequency resolution. This represents the k-th frequency point. Indicates carrier band, Let represent the power spectral density at the k-th frequency point at time t.

[0043] Specifically, the health level at time t. The calculation formula is as follows:

[0044]

[0045] in Indicates frequency resolution. This represents the k-th frequency point. Indicates mechanical health belt, Let represent the power spectral density at the k-th frequency point at time t.

[0046] In one embodiment of the present invention, within a time range specified by the reference time window parameter, the carrier band equivalent energy bandwidth sequence is extracted and its median is calculated, and the median is used as the historical baseline.

[0047] Divide the current carrier band equivalent energy bandwidth value by the historical baseline value to obtain the injection factor; calculate the difference between the injection factor and one, multiply the difference by the preset fixed debiasing coefficient and add it to one to obtain the correction denominator, and divide the health value by the correction denominator to obtain the debiased health value.

[0048] Calculate the natural logarithm of the carrier band equivalent energy bandwidth at the current moment, calculate the natural logarithm of the carrier band equivalent energy bandwidth at the previous moment, subtract the natural logarithm of the previous moment from the natural logarithm of the current moment to obtain the logarithmic difference, divide the logarithmic difference by the time interval between adjacent outputs to obtain the bandwidth change rate.

[0049] It should be noted that the reference time window parameter is used to define the time range for calculating the historical baseline. The reference time window parameter is the sum of 30 to 60 adjacent output time intervals. In typical rotating pump scenarios, the sum of 50 adjacent output time intervals is preferred to ensure coverage of at least one complete equipment operating condition fluctuation cycle. The carrier band equivalent energy bandwidth sequence represents the set of carrier band equivalent energy bandwidth values ​​arranged chronologically within the time range specified by the reference time window parameter. The historical baseline represents the average level of the carrier band equivalent energy bandwidth within the reference time window, serving as a benchmark for measuring the relative change of the current carrier band equivalent energy bandwidth; the historical baseline is equal to the median of all values ​​in the carrier band equivalent energy bandwidth sequence extracted within the reference time window. The injection factor is used to quantify the degree of change of the current carrier band equivalent energy bandwidth relative to the historical baseline, reflecting the strength of the electrical spread spectrum; the injection factor is equal to the current carrier band equivalent energy bandwidth value divided by the historical baseline value, i.e., the ratio of the current bandwidth value to the baseline value.

[0050] It should be noted that the fixed debiasing coefficient represents a preset constant parameter used to adjust the debiasing intensity of the health quantity, with a value ranging from 0 to 1, determining the deduction ratio of the electrical spectrum spread effect. The fixed debiasing coefficient is determined according to the type of rotating pump: 0.6 to 0.8 for centrifugal pumps, 0.4 to 0.6 for centrifugal fans, and 0.5 to 0.7 for general-purpose variable frequency drive rotating pumps. If a precise value is required, the equipment can be run under no-load for three times, adjusting the coefficient and calculating the fluctuation amplitude of the health quantity after debiasing in each run, and taking the coefficient value with the smallest fluctuation amplitude as the final fixed debiasing coefficient. The correction denominator is used to correct the amplitude of the health quantity, achieving adaptive debiasing by changing the injection factor and the fixed debiasing coefficient; the correction denominator equals one plus (the difference between the injection factor and one multiplied by the fixed debiasing coefficient), that is, first calculate the result of the injection factor minus one, then multiply by the fixed debiasing coefficient, and finally add one. The debiased health quantity represents a quantitative indicator of the mechanical health status after eliminating the electrical spectrum spread effect.

[0051] Specifically, at time t, the partial health value is... The calculation formula is as follows:

[0052]

[0053] in Let represent the health status at time t. This indicates a fixed debiasing coefficient. Let represent the injection factor at time t.

[0054] It should be noted that the natural logarithm of the carrier band equivalent energy bandwidth at the current moment is used to reduce the scale difference of bandwidth values ​​and stabilize the rate of change calculation. The logarithmic difference represents the magnitude of the logarithmic change in the carrier band equivalent energy bandwidth between two adjacent moments, reflecting the relative magnitude of the bandwidth change. The time interval between adjacent outputs is equal to the analysis window length parameter divided by the sampling frequency parameter, and then multiplied by the sliding step size parameter. That is, first calculate the time of a single analysis window (analysis window length ÷ sampling frequency), and then multiply it by the interval multiple corresponding to the sliding step size; for example, when the sampling frequency parameter is 1000 Hz, the analysis window length parameter is 1024 points, and the sliding step size parameter is 512 points, the time of a single analysis window is 1024 ÷ 1000 = 1.024 seconds, and the time interval between adjacent outputs is 1.024 × (512 ÷ 1024) = 0.512 seconds. That is, when the sliding step size is half the analysis window length, the time interval is half the time of a single analysis window. The rate of change of bandwidth represents how quickly the carrier band equivalent energy bandwidth changes with time, reflecting the time-varying speed of the electrical spread spectrum.

[0055] Specifically, the rate of change of bandwidth at time t. The calculation formula is as follows:

[0056]

[0057] in Let represent the natural logarithm of the carrier band's equivalent energy bandwidth at time t. This indicates the time interval between adjacent outputs.

[0058] It should be noted that the reference time window parameter needs to be determined in conjunction with the operating fluctuation cycle of the centrifugal pump. For a centrifugal pump with a load fluctuation cycle of 10 minutes, the reference time window parameter should be 500 seconds (approximately 8.3 minutes); for a centrifugal fan with a load fluctuation cycle of 5 minutes, the reference time window parameter should be 250 seconds (approximately 4.2 minutes). The minimum reference time window should be the sum of 30 adjacent output time intervals, and the maximum should not exceed 60 adjacent output time intervals to avoid a baseline that is too short (easily affected by disturbances) or too long (unable to reflect the latest operating conditions). The value of the fixed debiasing coefficient needs to be determined in conjunction with the variable frequency drive power of the equipment. For small centrifugal pumps with a power of 11 kW and below, the coefficient should be 0.4 to 0.5; for medium-sized centrifugal pumps with a power of 15 kW to 55 kW, the coefficient should be 0.5 to 0.6; and for large centrifugal pumps with a power of 75 kW and above, the coefficient should be 0.7 to 0.8. If the operating environment of the equipment is subject to strong electromagnetic interference (such as multiple frequency converters operating together), the coefficient can be increased by 0.1 to enhance the debiasing strength.

[0059] In one embodiment of the present invention, the most recent debiased health quantity sequence of a preset length is used to form a time-series input vector;

[0060] Calculate the natural logarithm of the equivalent energy bandwidth of the carrier band at the current moment, and combine the natural logarithm with the bandwidth change rate to form a conditional input vector;

[0061] One-dimensional convolution operations are performed on the temporal input vector using a preset one-dimensional convolution kernel and convolution bias, and a non-linear activation function is applied to the operation result to obtain basic convolution features.

[0062] It should be noted that the most recent preset length is a positive integer parameter used to extract the debiased health quantity sequence to construct the time-series input vector. It determines the number of historical health data points included in the time-series input, affecting the network's ability to capture trends. The most recent preset length is equal to the equipment fault warning period divided by the time interval between adjacent outputs, typically an integer between 30 and 50. For centrifugal pumps, a preferred value is 40 to 50, and for centrifugal fans, a preferred value is 30 to 40, ensuring complete time-series characteristics covering the typical early stages of equipment degradation. The debiased health quantity sequence represents the set of debiased health quantity values ​​for the most recent preset length of consecutive moments arranged chronologically. The conditional input vector is a column vector composed of the natural logarithm of the carrier band's equivalent energy bandwidth at the current moment and the bandwidth change rate.

[0063] It should be noted that the preset one-dimensional convolutional kernel represents the weight matrix used to extract temporal local features in a one-dimensional convolutional neural network. It is a three-dimensional structure (single input channel) with kernel length × 1 × number of kernels. The kernel length is an odd number from 3 to 7 (5 is preferred), and the number of kernels is from 8 to 16 (12 is preferred). The kernel weights are trained using historical device data (including healthy and slightly degraded states). The training objective is to minimize the trend slope estimation error. The initial weights are initialized using a Xavier normal distribution, meaning the weight values ​​follow a normal distribution with a mean of 0 and a standard deviation of 2 ÷ the square root of (input dimension + output dimension). The convolution bias represents the set of scalar biases used in conjunction with the one-dimensional convolutional kernel. The number of biases is the same as the number of kernels. It is used to adjust the baseline of the convolution operation output to avoid feature shift. The initial value of each convolution bias is between 0.01 and 0.05. Through training optimization, the final value range is controlled between -0.1 and 0.1. The optimization objective is consistent with the kernel weights, ensuring that the baseline of each output feature channel is close to zero. One-dimensional convolution operation involves sliding a one-dimensional convolution kernel along the time dimension of the temporal input vector, calculating the inner product of the kernel and the corresponding local temporal segment at each position, and then adding the convolution bias. This is used to extract local trend features from temporal data. The ReLU function is preferred as the non-linear activation function. The basic convolutional feature represents the feature matrix obtained after processing the one-dimensional convolution result with a non-linear activation function, with dimensions of: temporal input length × number of convolution kernels.

[0064] In one embodiment of the present invention, a scaling factor is obtained by performing a linear weighted calculation on the conditional input vector using a preset first set of fixed weight parameters, and a translation factor is obtained by performing a linear weighted calculation on the conditional input vector using a preset second set of fixed weight parameters.

[0065] The scaling factor is obtained by adding the scaling factor to the numerical value. The scaling factor is then multiplied element-wise by the basic convolutional features to obtain the product. The product is then added element-wise by the translation coefficient to obtain the modulation feature. The modulation feature is then multiplied by the preset readout weight vector and accumulated with the preset readout bias to obtain the trend slope estimate.

[0066] It should be noted that the first set of fixed weight parameters represents a set of scalar weights used to linearly map the conditional input vector to scaling factors, containing three parameters (corresponding to two elements of the conditional input vector and a basic term); in the first set of fixed weight parameters, the weight corresponding to the natural logarithm value takes a range of -0.2 to 0.3, the weight corresponding to the bandwidth change rate takes a range of -0.1 to 0.2, and the basic term takes a range of 0.05 to 0.15; the training and optimization are performed using the gradient descent algorithm, and the training data must contain at least three sets of complete time-series data from device health to slight degradation, with the optimization objective being to minimize the difference between the modulation characteristics and the true mechanical trend characteristics. The scaling factor represents a scalar obtained by linearly weighting the conditional input vector using the first set of fixed weight parameters, used to adjust the magnitude scale of the base convolutional features. The second set of fixed weight parameters represents a set of scalar weights used to linearly map the conditional input vector to translation coefficients, containing three parameters (two elements corresponding to the conditional input vector and the base term). In the second set of fixed weight parameters, the weights corresponding to the natural logarithm value range from -0.1 to 0.2, the weights corresponding to the bandwidth change rate range from -0.05 to 0.15, and the base term ranges from -0.05 to 0.05. The training data and optimization objective are the same as the first set of fixed weight parameters, ensuring that the translation coefficients can compensate for the baseline deviation of the scaled features.

[0067] Specifically, the scaling factor obtained from conditional input mapping The calculation formula is as follows:

[0068]

[0069] in Represents the conditional input vector. Let represent the natural logarithm of the carrier band's equivalent energy bandwidth at time t. This represents the rate of change of bandwidth at time t. , and These represent the first, second, and third weight parameters in the first set of fixed weight parameters, respectively.

[0070] Specifically, the translation coefficients obtained from the conditional input mapping The calculation formula is as follows:

[0071]

[0072] in Represents the conditional input vector. , and These represent the first, second, and third weight parameters in the second set of fixed weight parameters, respectively.

[0073] It should be noted that the translation coefficient represents a scalar obtained by linearly weighting the conditional input vector using the second set of fixed weight parameters. This is used to adjust the baseline of the modulated features and eliminate feature shift caused by electrical spread. The multiplication factor represents a scalar used to amplify the basic convolutional features, obtained by adding 1 to the scaling factor, ensuring that some feature information is retained even when the scaling factor is negative. The modulation feature represents the feature matrix obtained by adding the product result and the translation coefficient element-wise. Its dimension is the same as the basic convolutional features, and it is the final convolutional feature after adaptive adjustment by electrical conditions. The readout weight vector represents a column vector used to map the modulation features to the trend slope estimate. Its dimension is the same as the number of channels (number of convolutional kernels) of the modulation features. Each element of the readout weight vector has a value between -0.5 and 0.5. Through training and optimization, the optimization objective is to minimize the mean square error between the estimated trend slope and the actual mechanical trend slope. The initial weights are initialized using a Xavier normal distribution, and the absolute value of the sum of the vector elements is controlled between 0.1 and 0.3 to avoid overfitting. The readout bias represents a scalar used to adjust the baseline of the trend slope estimation; the readout bias ranges from -0.02 to 0.02 and is optimized through training, with the optimization objective being to minimize the deviation of the mean of the trend slope estimation for the healthy data segment from zero; the initial value is 0.01. The trend slope estimation represents a scalar of the rate of change of the mechanical health status of the rotating pump.

[0074] Specifically, modulation characteristics The calculation formula is as follows:

[0075]

[0076] in This represents the scaling factor obtained from the conditional input mapping. This represents the translation coefficient obtained from the conditional input mapping. Represents the basic convolutional features. This indicates element-wise multiplication.

[0077] It should be noted that the training data must include full lifecycle data (from factory health status to the first minor degradation alarm) for at least three rotating pumps of the same model. The data duration for each device should be no less than 30 days, the sampling frequency no less than 1000 Hz, and it must cover more than three typical load conditions (e.g., 50%, 75%, and 100% rated load). During data preprocessing, data segments experiencing instantaneous shocks (e.g., startup and shutdown) should be removed, retaining only stable operating data. When there is a lot of high-frequency noise in the rotating pump vibration data (e.g., initial bearing wear), the ReLU function should be selected first to suppress the negative invalid features caused by noise. When there are obvious negative trend features in the data (e.g., slight impeller corrosion leading to a decrease in vibration amplitude), the LeakyReLU function (negative slope 0.01) should be selected to retain the effective negative trend. The proportion of negative elements in the convolution output of the healthy data segment can be calculated; if the proportion exceeds 5%, LeakyReLU should be used; otherwise, ReLU should be used.

[0078] In one embodiment of the present invention, the bandwidth change rate is multiplied by a preset coupling sensitivity coefficient to obtain a pseudo-increment of the slope; the slope pseudo-increment is subtracted from the trend slope estimate to obtain the mechanical body slope.

[0079] It should be noted that the preset coupling sensitivity coefficient represents a dimensionless constant that characterizes the correlation strength between the electrical spread spectrum change (bandwidth change rate) and the pseudo-component of the trend slope, determining the calculation amplitude of the slope pseudo-increment, and needs to be determined through offline equipment calibration. The preset coupling sensitivity coefficient is equal to the median of the ratio of the measured value of the slope pseudo-increment to the measured value of the corresponding bandwidth change rate under rated load and different carrier spread spectrum intensities. The specific calibration steps are as follows: 1. Run the equipment under no-load and record the estimated trend slope at this time (considered as a pure mechanical slope, denoted as A); 2. Adjust the spread spectrum intensity of the frequency converter so that the bandwidth change rates are 0.01, 0.02, and 0.03 respectively, and record the estimated trend slope for each state under rated load (denoted as B1, B2, and B3); 3. Calculate the slope pseudo-increment for each state (B1-A, B2-A, B3-A); 4. Calculate the ratio of each slope pseudo-increment to the corresponding bandwidth change rate, and take the median of these ratios as the preset coupling sensitivity coefficient. The slope pseudo-increment represents the spurious component introduced by electrical spectrum expansion and superimposed on the true mechanical trend slope; the slope pseudo-increment equals the bandwidth change rate multiplied by a preset coupling sensitivity coefficient. The mechanical body slope represents an index that reflects only the rate of change of the health status of the rotating pump after deducting the electrical pseudo-component; the mechanical body slope equals the trend slope estimate minus the slope pseudo-increment.

[0080] In one embodiment of the present invention, the slope of the mechanical body is multiplied by a preset extrapolation duration parameter to obtain the extrapolation increment; the extrapolation increment is added to the current debiased health value to obtain the future prediction endpoint.

[0081] It should be noted that the preset extrapolation duration parameter represents the pre-set time span used to predict future mechanical health status, determined by equipment maintenance needs and fault development cycle. The preset extrapolation duration parameter is equal to the equipment's average fault development cycle multiplied by a coefficient of 0.6 to 0.8 (to allow for maintenance preparation time). The equipment's average fault development cycle is obtained through historical fault data statistics, i.e., collecting the time intervals from slight deterioration to fault alarm for the same model of equipment, and taking the median of these intervals as the average fault development cycle. If no historical data is available, the preset extrapolation duration parameter for general rotating pumps (such as centrifugal pumps and fans) is preferably set to 72 to 168 hours (3 to 7 days). The extrapolation increment represents the incremental value of the mechanical health quantity as a function of the slope of the mechanical body within the preset extrapolation duration. The extrapolation increment is equal to the slope of the mechanical body multiplied by the preset extrapolation duration parameter, i.e., the product of rate and time, reflecting the expected change in the mechanical health quantity within the preset duration. The future forecast endpoint represents the predicted mechanical health value at the end of the preset extrapolation period. The future forecast endpoint is equal to the debiased health value at the current moment plus the extrapolation increment, which is the sum of the current baseline value and the expected change.

[0082] In one embodiment of the present invention, within a preset robust baseline construction time window, the debiased healthy quantity sequence is extracted and its median is calculated, and the median is used as the median baseline.

[0083] The absolute value of the difference between the biased health value and the median baseline value at each moment within the time window is calculated. The median of the sequence of all absolute values ​​is calculated to obtain the absolute median difference. The absolute median difference is multiplied by a preset fixed scaling coefficient to obtain a robust scale.

[0084] The normalized numerator is obtained by subtracting the median baseline from the future predicted endpoint. The normalized denominator is obtained by adding the robust scale to the preset minimum constant. The normalized numerator is then divided by the normalized denominator to obtain the quantitative result of the vibration state trend of the rotating pump.

[0085] It should be noted that the preset robust baseline construction time window is equal to 3 to 5 times the preset extrapolation duration parameter; if the preset extrapolation duration is 72 hours (3 days), then the robust baseline construction time window is 216 to 360 hours (9 to 15 days); in general scenarios, the minimum is no less than 72 hours (to ensure sufficient historical data coverage) and the maximum is no more than 336 hours (14 days, to avoid the baseline being too old to reflect the current operating conditions). The median baseline represents a robust indicator of the average level of debiased health within the preset robust baseline construction time window, serving as a benchmark for measuring the deviation of future prediction endpoints. The absolute median difference represents a robust indicator of the fluctuation of debiased health relative to the median baseline within the preset robust baseline construction time window; the absolute median difference is equal to the median of the sequence formed by the absolute values ​​of the differences between the debiased health value and the median baseline value at each moment within the time window, i.e., first calculate the absolute deviation of each value from the baseline, and then take the median of these deviations. The robust scale represents a metric for the fluctuation of the scaled debiased health; the robust scale is equal to the absolute median difference multiplied by a preset fixed scaling coefficient. The quantitative results of the vibration state trend of the rotating pump indicate the degree and direction of deviation of the future mechanical health state of the rotating pump from the historical baseline.

[0086] Specifically, the quantitative results of the vibration state trend of the rotating pump at time t. The calculation formula is as follows:

[0087]

[0088] in Indicates the first The future prediction endpoint at each moment This represents the baseline of the median at time t. This represents the robust metric at time t. This represents a very small constant.

[0089] It should be noted that the preset robust baseline construction time window needs to be adjusted according to the stability of equipment operating conditions: 1. For stable operating conditions (load fluctuation less than or equal to ±5%), take 5 times the preset extrapolation time. For example, if the extrapolation time is 72 hours, the time window is 360 hours (15 days) to ensure that the baseline reflects the long-term stable level; 2. For fluctuating operating conditions (load fluctuation greater than ±5%), take 3 times the preset extrapolation time. For example, if the extrapolation time is 72 hours, the time window is 216 hours (9 days) to avoid old operating condition data interfering with the current baseline; 3. For frequent equipment start-up and shutdown scenarios, shorten the original time window by 20%. For example, shorten the original 360 hours to 288 hours to reduce the impact of instantaneous start-up and shutdown data on the baseline.

[0090] It should be noted that the interpretation criteria for trend quantification results are as follows: 1. A value between -0.5 and 0.5 indicates a stable trend, with no significant changes in the equipment's health status, requiring no intervention; 2. A value greater than 0.5 and less than or equal to 1.5 indicates a positive trend (e.g., a decrease in vibration amplitude), requiring continuous monitoring to confirm whether it is caused by maintenance or operational optimization; 3. A value greater than 1.5 indicates an abnormally positive trend, requiring investigation into whether data acquisition or calculation is abnormal (e.g., a loose sensor causing a low vibration value); 4. A value less than -0.5 and greater than or equal to -1.5 indicates a slight deterioration in the trend, requiring the development of a short-term monitoring plan (e.g., monitoring once a day); 5. A value less than -1.5 indicates a severely deteriorated trend, requiring immediate equipment inspection and maintenance to prevent the fault from escalating.

[0091] In one embodiment of the present invention, a vibration state trend prediction system for a rotating pump based on edge computing includes:

[0092] The health quantity calculation module 201 is used to perform frequency domain transformation on the collected vibration data to obtain the power spectral density, statistically analyze the power spectral density distribution shape within a preset carrier band to obtain the carrier band equivalent energy bandwidth, and statistically analyze the energy within a mechanical health band that avoids the carrier band to obtain the health quantity.

[0093] The bandwidth change rate calculation module 202 is used to calculate the injection factor based on the ratio of the carrier band equivalent energy bandwidth to its historical baseline, correct the health quantity through the injection factor to obtain the debiased health quantity, and calculate the bandwidth change rate based on the carrier band equivalent energy bandwidth at adjacent times.

[0094] The trend slope estimation module 203 is used to take the time segment composed of the debiased health quantity as the time input, and the vector composed of the carrier band equivalent energy bandwidth and its bandwidth change rate as the condition input, and input it into a one-dimensional convolutional neural network. The scaling coefficient and translation coefficient generated by the condition input mapping are used to perform channel-wise linear modulation on the network convolutional features, and the trend slope is estimated by linear readout.

[0095] The mechanical body slope calculation module 204 is used to calculate the product of the bandwidth change rate and the preset coupling sensitivity coefficient, and subtract the product from the trend slope estimate to obtain the mechanical body slope.

[0096] The future prediction endpoint calculation module 205 is used to perform deterministic extrapolation calculation within a preset extrapolation time based on the slope of the mechanical body and the current debiased health value to obtain the future prediction endpoint.

[0097] The rotating pump vibration state quantification module 206 is used to statistically analyze the median baseline and absolute median difference scale of the debiased health quantity within the historical time window. It normalizes the difference between the future prediction endpoint and the median baseline using the absolute median difference scale and outputs the quantification result of the rotating pump vibration state trend.

[0098] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0099] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A method for predicting the trend of the vibration state of a rotary machine pump based on edge computing, characterized by, The method comprises the following steps: Step S101, performing frequency domain transformation on the collected vibration data to obtain a power spectrum density, and statistically obtaining a power spectrum density distribution shape in a preset carrier band to obtain a carrier band equivalent energy bandwidth, and statistically obtaining an energy in a mechanical health band avoiding the carrier band to obtain a health quantity; Step S102, calculating an injection factor based on a ratio of the carrier band equivalent energy bandwidth relative to its historical baseline, correcting the health quantity by the injection factor to obtain a de-biased health quantity, and calculating a bandwidth change rate according to the carrier band equivalent energy bandwidth of adjacent time instants; Step S103, taking a time sequence segment composed of the de-biased health quantity as a time sequence input, taking a vector composed of the carrier band equivalent energy bandwidth and the bandwidth change rate as a conditional input, and inputting into a one-dimensional convolutional neural network, performing channel linear modulation on network convolutional features by scaling coefficients and translation coefficients generated by conditional input mapping, and obtaining a trend slope estimate through linear reading; Step S104, calculating a product of the bandwidth change rate and a preset coupling sensitivity coefficient, and subtracting the product from the trend slope estimate to obtain a mechanical body slope; Step S105, based on the mechanical body slope and the de-biased health quantity at the current time instant, performing a deterministic extrapolation calculation in a preset extrapolation time length to obtain a future prediction endpoint; Step S106, statistically obtaining a median baseline and an absolute median difference scale of the de-biased health quantity in a historical time window, normalizing the difference between the future prediction endpoint and the median baseline by the absolute median difference scale, and outputting a rotating machine pump vibration state trend quantization result; The step of statistically obtaining the power spectrum density distribution shape in the preset carrier band to obtain the carrier band equivalent energy bandwidth, and statistically obtaining the energy in the mechanical health band avoiding the carrier band to obtain the health quantity comprises: The carrier band is determined according to the carrier center frequency parameter and the carrier neighborhood half-bandwidth parameter, the square of the sum of the power spectrum density values of all frequency points in the carrier band is calculated, the sum of the squares of the power spectrum density values of all frequency points in the carrier band is calculated, the square of the sum of the values is divided by the sum of the squares of the values and multiplied by the frequency resolution to obtain the carrier band equivalent energy bandwidth; The mechanical health band is determined to be mutually non-overlapping with the carrier band, the cumulative sum of the product of the power spectrum density values of all frequency points in the mechanical health band and the frequency resolution is calculated, and the cumulative sum is subjected to square root operation to obtain the health quantity.

2. The edge computing-based rotating machine pump vibration state trend prediction method according to claim 1, characterized by, The power spectrum density is obtained by determining the current analysis window and the frequency resolution according to the sampling frequency parameter, the analysis window length parameter and the sliding step parameter, applying a window function to the vibration data in the current analysis window and performing fast Fourier transform, and normalizing the square of the modulus of the transformation result according to the analysis window length parameter and the sampling frequency parameter.

3. The edge computing-based rotating machine pump vibration state trend prediction method according to claim 1, characterized by, In the time range specified by the reference time window parameter, the carrier band equivalent energy bandwidth sequence is extracted and the median is calculated, and the median is taken as the historical baseline; The carrier band equivalent energy bandwidth value at the current time instant is divided by the historical baseline value to obtain the injection factor; The difference between the injection factor and one is calculated, the difference is multiplied by a preset fixed de-biasing coefficient and then added to one to obtain a modified denominator, and the health quantity is divided by the modified denominator to obtain the de-biased health quantity; The natural logarithm value of the equivalent energy bandwidth of the carrier band at the current time is calculated, the natural logarithm value of the equivalent energy bandwidth of the carrier band at the previous time is calculated, the natural logarithm value at the current time is subtracted from the natural logarithm value at the previous time to obtain a logarithm difference value, and the logarithm difference value is divided by the time interval of adjacent outputs to obtain a bandwidth change rate. 4.The edge computing based rotating machine pump vibration state trend prediction method according to claim 1, characterized in that, A time sequence input vector is formed by intercepting a sequence of the latest preset length of the health quantity after deviation removal; The natural logarithm value of the equivalent energy bandwidth of the carrier band at the current time is calculated, and the natural logarithm value and the bandwidth change rate are combined to form a condition input vector; A one-dimensional convolution operation is performed on the time sequence input vector by a preset one-dimensional convolution kernel and a convolution bias, and a nonlinear activation function is applied to the operation result to obtain a basic convolution feature.

5. The edge computing-based rotating machine pump vibration state trend prediction method according to claim 4, characterized by, A scaling coefficient is obtained by performing linear weighting calculation on the condition input vector by a preset first group of fixed weight parameters, and a translation coefficient is obtained by performing linear weighting calculation on the condition input vector by a preset second group of fixed weight parameters; A multiplication factor is obtained by adding the scaling coefficient and the number one, and the multiplication factor and the basic convolution feature are multiplied element by element to obtain a product result, the product result and the translation coefficient are added element by element to obtain a modulation feature, and the modulation feature and a preset readout weight vector are subjected to an inner product operation and a preset readout bias is accumulated to obtain a trend slope estimate.

6. The edge computing-based rotating machine pump vibration state trend prediction method according to claim 1, characterized by, A slope pseudo-increment is obtained by performing a multiplication operation on the bandwidth change rate and a preset coupling sensitivity coefficient; and a mechanical body slope is obtained by subtracting the slope pseudo-increment from the trend slope estimate.

7. The edge computing based rotating machine pump vibration condition trend prediction method of claim 1, wherein, A extrapolation increment is obtained by performing a multiplication operation on the mechanical body slope and a preset extrapolation time length parameter; and a future prediction endpoint is obtained by performing an addition operation on the extrapolation increment and the health quantity after deviation removal at the current time.

8. The edge computing based rotating machine pump vibration condition trend prediction method of claim 1, wherein, In a preset robust baseline construction time window, the health quantity after deviation removal sequence is extracted and the median is calculated, and the median is taken as the median baseline; The absolute value of the difference between the health quantity after deviation removal at each time in the time window and the median baseline value is calculated, the median of the sequence of all absolute values is obtained to obtain an absolute median difference, and the absolute median difference is multiplied by a preset fixed scale factor to obtain a robust scale. A normalized numerator is obtained by subtracting the median baseline from the future prediction endpoint, a normalized denominator is obtained by performing an addition operation on the robust scale and a preset minimum constant, and a rotating machine pump vibration state trend quantization result is obtained by dividing the normalized numerator by the normalized denominator.

9. A rotating machine pump vibration condition trend prediction system based on edge computing, characterized by, The edge computing-based rotating machine pump vibration state trend prediction method according to any one of claims 1 to 8 is executed, comprising: A health quantity calculation module is configured to perform frequency domain transformation on the collected vibration data to obtain a power spectral density, calculate a carrier band equivalent energy bandwidth by counting the power spectral density distribution shape in a preset carrier band, and calculate a health quantity by counting the energy in a mechanical health band avoiding the carrier band; A bandwidth change rate calculation module is configured to calculate an injection factor based on the ratio of the carrier band equivalent energy bandwidth to its historical baseline, correct the health quantity to obtain a health quantity after deviation removal by the injection factor, and calculate a bandwidth change rate according to the carrier band equivalent energy bandwidth of adjacent time. A health quantity calculation module is configured to perform frequency domain transformation on the collected vibration data to obtain a power spectral density, calculate a carrier band equivalent energy bandwidth by counting the power spectral density distribution shape in a preset carrier band, and calculate a health quantity by counting the energy in a mechanical health band avoiding the carrier band; A bandwidth change rate calculation module is configured to calculate an injection factor based on the ratio of the carrier band equivalent energy bandwidth to its historical baseline, correct the health quantity to obtain a health quantity after deviation removal by the injection factor, and calculate a bandwidth change rate according to the carrier band equivalent energy bandwidth of adjacent time. a trend slope estimation module configured to input a time series segment of the de-biased health quantity as a time series input, input a vector of the carrier band equivalent energy bandwidth and the bandwidth change rate thereof as a conditional input, input the vector into a one-dimensional convolutional neural network, perform channel-wise linear modulation on the network convolutional features by using the scaling coefficient and the translation coefficient generated by the conditional input mapping, and obtain the trend slope estimation through linear reading; a mechanical body slope calculation module configured to calculate a product of the bandwidth change rate and a preset coupling sensitivity coefficient, subtract the product from the trend slope estimation, and obtain the mechanical body slope; a future prediction endpoint calculation module configured to perform a deterministic extrapolation calculation within a preset extrapolation time length based on the mechanical body slope and the de-biased health quantity at the current time, and obtain the future prediction endpoint; a rotating machine pump vibration state quantification module configured to calculate a median baseline and an absolute median difference scale of the de-biased health quantity within a historical time window, normalize the difference between the future prediction endpoint and the median baseline by using the absolute median difference scale, and output a rotating machine pump vibration state trend quantification result.

Images