Method and device for monitoring the wear state of a roll for high-frequency welded pipes

Abstract

The application provides a high-frequency pipe welding roller wear state monitoring method and device, relates to the field of artificial intelligence and data processing technology, and constructs a holographic perception system of working condition parameters and physical signals through synchronous monitoring of multiple source sensors, tracks noise spectrum shift in real time based on time-frequency expression of data in different modes, constructs an anti-misfiltering protection channel, and can effectively capture damage inherent characteristics. Furthermore, from multiple complementary dimensions such as transient impact (instantaneous frequency vector), spectral complexity (spectral entropy) and amplitude distribution skewness (skewness), multi-physical field damage characterization information is fused, more comprehensive and richer state description can be provided, precise quantitative evaluation of roller wear grade can be realized, and the ability of roller wear state real-time monitoring and accurate identification is significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and data processing, and in particular to a method and device for monitoring the wear state of a high-frequency welded pipe roller. Background Art

[0002] High-frequency welded pipe is widely used in key industrial sectors such as oil transportation, structural steel pipe, and automotive manufacturing. The forming process relies on multiple sets of rollers to continuously plastically deform the steel strip. Because the rollers are subjected to extreme conditions of high load, high temperature, and high-speed friction and impact, their surfaces are prone to scratches, flaking, cracks, and even plastic deformation, resulting in deviations in the rolling pass, which in turn affects the geometric accuracy of the pipe and weld quality. Roller wear directly determines the consistency and stability of welded pipe products and is a key element in welded pipe production that requires real-time monitoring and precise identification.

[0003] Current wear prediction technologies are generally limited by the inherent conflict between static processing logic and dynamic industrial scenarios, which manifests itself in three underlying contradictions:

[0004] 1) The incompatibility between dynamic working conditions and static assumptions: Real-time fluctuations in load, speed, and temperature at industrial sites cause the noise distribution of vibration / acoustic emission signals to continuously vary. However, existing processing logic relies on fixed rules preset by historical data (such as noise thresholds and frequency band ranges), making it impossible to establish a dynamic mapping relationship between signal characteristics and working condition parameters. As a result, microwear characteristics (such as the microsecond impact of incipient cracks) are continuously obliterated by strong background noise.

[0005] 2) Physical decoupling of the rotational mechanism from the global processing: Roller wear exhibits distinct phase-sensitive characteristics, with damage such as spalling and cracking concentrated in the rolling contact zone. Existing methods treat the rotational cycle as a homogeneous time series, which can dilute the critical damage signal with noise in non-sensitive areas.

[0006] 3) The natural conflict between transient events and statistical smoothing: Wear-related impact events (such as the impact of flaking particles) last as short as milliseconds and have weak energy. The current processing logic uses time-domain statistical averaging to force smoothing, erasing key details of the transient waveform (such as the rising edge slope and oscillation decay rate), causing event peak distortion and severely weakening the damage quantification capability. Summary of the Invention

[0007] In view of this, the object of the present invention is to provide a method and device for monitoring the wear status of high-frequency welded pipe rollers, which can significantly improve the ability to monitor and accurately identify the wear status of rollers in real time.

[0008] In a first aspect, an embodiment of the present invention provides a method for monitoring the wear state of a high-frequency welded pipe roll, the method comprising: online monitoring of the operating parameters of a target roll and operating state parameters corresponding to the operating parameters, and determining real-time multi-source sensor stream data of the target roll; the operating state parameters include a variety of dynamic physical signals during the high-frequency welded pipe rolling process; based on the time-frequency expression of the real-time multi-source sensor stream data in each mode, performing dynamic adaptive filtering processing on the real-time multi-source sensor stream data to determine a noise reduction signal of the real-time multi-source sensor stream data; and determining the instantaneous noise reduction signal according to the signal characteristics of the noise reduction signal. time-frequency feature vector, spectral entropy feature and skewness feature; and, based on the instantaneous frequency feature vector, spectral entropy feature and skewness feature, a three-dimensional feature tensor of the target roller is constructed; wherein, the signal characteristics include frequency mutation characteristics, energy migration characteristics and non-Gaussian vibration characteristics; the three-dimensional feature tensor is input into a pre-constructed high-frequency welded pipe roller wear state monitoring model, and the high-frequency welded pipe roller wear state monitoring model performs data recognition processing on the three-dimensional feature tensor, and outputs the wear state prediction probability vector corresponding to the three-dimensional feature tensor; according to the wear state prediction probability vector, the wear state of the target roller is evaluated.

[0009] In combination with the first aspect, an embodiment of the present invention provides a first implementation method of the first aspect, wherein a three-dimensional feature tensor is input into a pre-constructed high-frequency welded pipe roller wear state monitoring model, the three-dimensional feature tensor is subjected to data recognition processing by the high-frequency welded pipe roller wear state monitoring model, and a wear state prediction probability vector corresponding to the three-dimensional feature tensor is output, comprising the following steps: inputting the three-dimensional feature tensor into the high-frequency welded pipe roller wear state monitoring model, modeling the long-range wear trend of the three-dimensional feature tensor based on the periodic memory gating mechanism of the high-frequency welded pipe roller wear state monitoring model, and determining the first target feature of the three-dimensional feature tensor; calculating the residual feature of the three-dimensional feature tensor, and generating the channel dimension entropy weight coefficient and time dimension entropy weight coefficient in parallel based on the residual feature. The local amplitude mean of the inter-dimensional dimension; based on the channel dimension entropy weight coefficient and the local amplitude mean of the time dimension, the three-dimensional feature tensor is subjected to two-dimensional soft threshold filtering, and the sensitive weak features representing wear in the three-dimensional feature tensor are enhanced; the first target feature is spliced ​​with the cell state of the preset LSTM unit along the feature dimension to generate the wear state gating vector of the first target feature based on the time scale; the statistical feature vector of the wear state gating vector at each time scale is extracted, the statistical feature vector of each time scale is aggregated, and the multi-scale weighted fusion feature vector is output; the physical association between the multi-scale weighted fusion feature vector and the wear type is modeled based on the preset scale gating vector, and the wear state prediction probability vector corresponding to the three-dimensional feature tensor is determined.

[0010] In combination with the first aspect, an embodiment of the present invention provides a second implementation method of the first aspect, wherein the long-range wear trend of the three-dimensional feature tensor is modeled based on the periodic memory gating mechanism of the high-frequency welded pipe roller wear state monitoring model, and the step of determining the first target feature of the three-dimensional feature tensor includes: using an exponential decay term based on the deviation between the current phase position and the wear-sensitive phase angle, dynamically adjusting the memory strength of the forget gate of the preset LSTM unit, performing periodic perception on the three-dimensional feature tensor, and determining the forget gate output of the three-dimensional feature tensor; the forget gate output is used to characterize the first target feature of the three-dimensional feature tensor.

[0011] In combination with the first aspect, an embodiment of the present invention provides a third implementation of the first aspect, wherein the above method further includes: updating the cell state and hidden state of the preset LSTM unit based on the output of the forget gate.

[0012] In combination with the first aspect, an embodiment of the present invention provides a fourth implementation of the first aspect, wherein the component of the preset scale gating vector at each time scale is determined by learning the preference of the wear state for the corresponding time scale through an item splicing operation.

[0013] In combination with the first aspect, an embodiment of the present invention provides a fifth implementation of the first aspect, wherein the above method also includes: using a pre-calculated feature space topological constraint loss to perform model training on a high-frequency welded pipe roller wear state monitoring model; wherein the steps of calculating the feature space topological constraint loss are as follows: calculating the feature vector distance between each feature vector of the multi-scale weighted fusion feature vector, and the reference distance of the multi-scale weighted fusion feature vector based on label difference; calculating the weight coefficient of the feature vector of adjacent wear levels in the multi-scale weighted fusion feature vector, and performing adjacent wear level weight constraints on the feature vector; using a preset Hinge loss function to constrain the deviation of the feature vector distance from the reference distance, and combining the weight coefficient to calculate the feature space topological constraint loss corresponding to the high-frequency welded pipe roller wear state monitoring model.

[0014] In combination with the first aspect, an embodiment of the present invention provides a sixth implementation of the first aspect, wherein, based on the time-frequency expression of the real-time multi-source sensor stream data in each mode, the real-time multi-source sensor stream data is dynamically adaptively filtered to determine the noise reduction signal of the real-time multi-source sensor stream data, including: performing variational modal decomposition on the real-time multi-source sensor stream data to determine the time-frequency expression corresponding to the real-time multi-source sensor stream data in multiple modes; calculating the energy entropy value of the time-frequency expression, and calculating the entropy weight coefficient corresponding to each mode based on the energy entropy value; filtering the real-time multi-source sensor stream data of each mode according to the entropy weight coefficient to determine the noise reduction signal of the real-time multi-source sensor stream data.

[0015] In combination with the first aspect, an embodiment of the present invention provides a seventh implementation of the first aspect, wherein the steps of determining the instantaneous frequency characteristic vector, spectral entropy characteristic and skewness characteristic of the noise reduction signal according to the signal characteristics of the noise reduction signal include: converting the noise reduction signal into an analytical signal through a preset nonlinear transformation algorithm, capturing the frequency mutation of the noise reduction signal caused by transient impact, and determining the instantaneous frequency characteristic vector of the noise reduction signal; calculating the energy spectrum of the noise reduction signal to determine the frequency distribution entropy value of the noise reduction signal; determining the spectral entropy characteristic of the noise reduction signal based on the energy distribution disorder represented by the frequency distribution entropy value; for the signal components of the noise reduction signal in multiple sensor dimensions, respectively calculating the third-order center distance characteristics corresponding to the signal components of each sensor dimension; the third-order center distance characteristics are used to describe the data skewness of the current sensor dimension to determine the skewness characteristics of the noise reduction signal.

[0016] In combination with the first aspect, an embodiment of the present invention provides an eighth implementation method of the first aspect, wherein the step of constructing a three-dimensional feature tensor of the target roller based on the instantaneous frequency feature vector, spectral entropy feature and skewness feature includes: horizontally splicing the instantaneous frequency feature vector, spectral entropy feature and skewness feature along the sensor dimension of the noise reduction signal, combining them into a three-dimensional fusion feature matrix, and forming a three-dimensional feature tensor based on "time-frequency-statistics".

[0017] In a second aspect, an embodiment of the present invention provides a high-frequency welded pipe roller wear state monitoring device, which includes: a data acquisition module for online monitoring of the working parameters of the target roller and the operating state parameters corresponding to the working parameters, and determining the real-time multi-source sensor stream data of the target roller; the operating state parameters include a variety of dynamic physical signals in the high-frequency welded pipe rolling process; a preprocessing module for dynamically adaptively filtering the real-time multi-source sensor stream data based on the time-frequency expression of the real-time multi-source sensor stream data in each mode, and determining the noise reduction signal of the real-time multi-source sensor stream data; a data processing module for determining the noise reduction signal according to the signal characteristics of the noise reduction signal. The instantaneous frequency characteristic vector, spectral entropy characteristic and skewness characteristic of the noise reduction signal are obtained; and based on the instantaneous frequency characteristic vector, spectral entropy characteristic and skewness characteristic, a three-dimensional feature tensor of the target roller is constructed; wherein, the signal characteristics include frequency mutation characteristics, energy migration characteristics and non-Gaussian vibration characteristics; an execution module is used to input the three-dimensional feature tensor into a pre-constructed high-frequency welded pipe roller wear state monitoring model, perform data recognition processing on the three-dimensional feature tensor through the high-frequency welded pipe roller wear state monitoring model, and output the wear state prediction probability vector corresponding to the three-dimensional feature tensor; an output module is used to evaluate the wear state of the target roller according to the wear state prediction probability vector.

[0018] The embodiments of the present invention provide the following beneficial effects: A method and apparatus for monitoring the wear status of high-frequency welded pipe rolls is provided. This method utilizes multi-source sensors to simultaneously monitor operating parameters and physical signals, constructing a holographic perception system. Based on the time-frequency representation of data in different modes, noise spectrum offsets are tracked in real time to construct an anti-false filtering protection channel, effectively capturing inherent damage characteristics. Furthermore, by integrating multi-physical field damage characterization information from multiple complementary dimensions, including transient impact (instantaneous frequency vector), spectral complexity (spectral entropy), and amplitude distribution skew (skewness), the method provides a more comprehensive and richer state description, enabling accurate quantitative assessment of roll wear levels and significantly improving the ability to monitor and accurately identify roll wear status in real time.

[0019] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or understood by practicing the present invention. The purposes and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0020] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0022] Figure 1 A flow chart of a method for monitoring the wear status of a high-frequency welded pipe roller provided by an embodiment of the present invention;

[0023] Figure 2 A flow chart of another method for monitoring the wear status of high-frequency welded pipe rollers provided in an embodiment of the present invention;

[0024] Figure 3 A schematic diagram showing the effect of a variational mode decomposition coupled dynamic entropy weight filtering method provided by an embodiment of the present invention;

[0025] Figure 4 A schematic diagram of the modeling capability of wear state continuity based on feature space topology constraint loss provided by an embodiment of the present invention;

[0026] Figure 5 Schematic diagram of comprehensive performance of an embodiment of the present invention;

[0027] Figure 6A schematic diagram of the physical interpretability of a three-dimensional feature tensor provided by an embodiment of the present invention;

[0028] Figure 7 A schematic structural diagram of a high-frequency welded pipe roller wear state monitoring device provided by an embodiment of the present invention;

[0029] Figure 8 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0031] The present invention provides a method and device for monitoring the wear status of high-frequency welded pipe rollers, which can significantly improve the ability to monitor and accurately identify the wear status of rollers (including different damage types such as scratches, spalling, cracks, and plastic deformation) in real time.

[0032] To facilitate understanding of this embodiment, first, a method for monitoring the wear state of a high-frequency welded pipe roll disclosed in an embodiment of the present invention is described with reference to Figure 1 The flowchart of a method for monitoring the wear state of a high-frequency welded pipe roller is shown, and the method comprises the following steps:

[0033] Step S102 : Online monitoring is performed on the operating parameters of the target roll and the operating status parameters corresponding to the operating parameters to determine the real-time multi-source sensor stream data of the target roll.

[0034] To address the limited dimensionality of single signals and the unpredictable dynamics of operating conditions in traditional wear monitoring, embodiments of the present invention utilize multi-source sensors to simultaneously perform online data monitoring, establishing a holographic observation system covering vibration, acoustic emission, temperature, and rotational speed. In specific implementations, the target roll operating parameters include roll speed, pipe material (e.g., Q235 steel), and pipe wall thickness. Operating status parameters include various dynamic physical signals from the high-frequency welded pipe rolling process. For example, a vibration accelerometer (e.g., a piezoelectric sensor) is mounted on the roll bearing housing, capturing radial and axial vibration signals at a 20kHz sampling rate to monitor roll-to-pipe contact impact. An infrared temperature sensor (with a response wavelength of 8–14μm) is aligned with the roll working surface, acquiring temperature distribution at a 100Hz sampling rate to characterize temperature rise anomalies caused by frictional heating. A current sensor (Hall effect type) is integrated into the main motor power supply circuit, recording drive current fluctuations at a 1kHz sampling rate to reflect load torque changes.

[0035] In summary, a multimodal signal stream with strict time synchronization is generated to perceive the coupling relationship between operating condition fluctuations and wear signals.

[0036] Step S104 : Based on the time-frequency expression of the real-time multi-source sensor stream data in each mode, dynamic adaptive filtering is performed on the real-time multi-source sensor stream data to determine a noise reduction signal of the real-time multi-source sensor stream data.

[0037] During roller operation, sudden load changes and speed fluctuations can cause real-time shifts in the noise spectrum (e.g., rising temperature shifts background noise energy to lower frequencies). Conventional filtering techniques (such as fixed thresholds, wavelets, and small window smoothing) are unable to cope with dynamic operating conditions and strongly coupled background noise. Using static filtering with a preset frequency band (e.g., a fixed high-pass cutoff frequency) can result in: Incorrect filtering: incorrectly eliminating significant wear-related frequency bands (e.g., crack characteristic frequencies masked by thermal deformation); and insufficient filtering: residual strong background noise, obscuring microsecond-level damage signals.

[0038] To this end, an embodiment of the present invention maps the dynamic interaction of the multi-physical fields of the rolling mill system into computable time-frequency domain features, uses the time-varying spectrum of the vibration signal to dynamically track the characteristic drift caused by sudden load changes, and constructs an anti-false filtering protection channel in the time-frequency domain. It can avoid false filtering of effective features based on the inherent characteristics of material damage (such as the 28kHz resonance peak of bearing steel and the microsecond rising edge of the shock wave), and effectively resolve the conflict between the time-varying dynamic working noise and the traditional fixed filtering parameters.

[0039] Step S106, determining the instantaneous frequency feature vector, spectral entropy feature and skewness feature of the denoised signal according to the signal characteristics of the denoised signal; and constructing a three-dimensional feature tensor of the target roll based on the instantaneous frequency feature vector, spectral entropy feature and skewness feature.

[0040] The instantaneous frequency vector captures the time-frequency characteristics of transient impact events (such as timing, frequency content, and intensity variations). Spectral entropy characterizes the complexity and order of the overall spectrum (reflecting damage type and severity). Skewness describes the asymmetry of the signal amplitude distribution (reflecting impact direction, friction characteristics, and abnormal load / temperature distribution). When defects such as scratches, spalling, and cracks on the roll surface pass through the rolling contact zone, they cause instantaneous and intense impact with the pipe, generating high-frequency transient components. The instantaneous frequency captures the onset, duration, and frequency range of these impacts. The vibration spectrum of a healthy roll may be relatively stable and orderly (with low spectral entropy). However, the presence of various types of damage, such as wear, pitting, and cracks, generates new and complex frequency components, making the spectral energy distribution more dispersed and increasing the spectral entropy. Impacts caused by spalling may produce a broadband response, significantly increasing the spectral entropy. Impacts in the roll contact zone (especially those caused by surface defects) typically manifest as large-amplitude pulses in the vibration signal. Skewness indicates the primary direction (positive / negative) and degree of asymmetry of these impacts. Severe spalling can produce extremely strong positive or negative impacts, resulting in a significant increase in the absolute value of the skewness. Based on this, the embodiments of the present invention characterize the operating status and potential damage of the roll from three different and complementary perspectives: "transient details" (time-frequency), "global structure" (frequency domain), and "statistical distribution" (amplitude). This provides a more comprehensive and richer description of the status, overcoming the limitations of a single feature. This significantly improves the ability to monitor and accurately identify the real-time wear status of the roll (including different damage categories such as scratches, spalling, cracks, and plastic deformation).

[0041] Step S108: input the three-dimensional feature tensor into a pre-built high-frequency welded pipe roller wear state monitoring model, perform data recognition processing on the three-dimensional feature tensor through the high-frequency welded pipe roller wear state monitoring model, and output a wear state prediction probability vector corresponding to the three-dimensional feature tensor.

[0042] Step S110 : evaluating the wear state of the target roller according to the wear state prediction probability vector.

[0043] In one embodiment, the wear state prediction probability vector includes the following categories: 1) Mild wear (profile deviation ≤ 0.05mm): surface micro-scratches; 2) Moderate wear (0.05mm < deviation ≤ 0.15mm): local material peeling; 3) Severe wear (deviation > 0.15mm): macro cracks / plastic deformation. The above categories can be determined by data annotation of the training sample set, and the corresponding labels for the above categories can be 0, 1, and 2. The data annotation can be based on offline precision measurement and expert diagnosis. Furthermore, after a period of downtime (such as every 8 hours), a laser profiler can be used to measure the roller hole profile deviation to divide the above wear state categories. The multivariate data set with time series labels of multi-source sensor data can be expressed as: ;in, is the time series data vector of the kth sensor, lab represents the roller wear level determined by offline precision measurement, encoded using discrete integer values, k represents the sensor dimension index, k=1,2,⋯,K, where K is the total number of sensors.

[0044] In summary, the present invention provides a high-frequency welded pipe roll wear monitoring method. This method utilizes multi-source sensor synchronization to build a holographic perception system for operating parameters and physical signals. It also uses the time-frequency representation of data in different modes to track noise spectrum shifts in real time, constructing a protection channel against false negatives and effectively capturing inherent damage characteristics. Furthermore, it integrates multi-physical field damage characterization information from multiple complementary dimensions, including transient impact (instantaneous frequency vector), spectral complexity (spectral entropy), and amplitude distribution skew (skewness), to achieve a quantitative assessment of roll wear levels, significantly enhancing the ability to monitor and accurately identify roll wear in real time.

[0045] Furthermore, based on the above embodiment, the embodiment of the present invention also provides another method for monitoring the wear state of high-frequency welded pipe rollers. Figure 2 A flowchart of another high-frequency welded pipe roll wear state monitoring method provided by an embodiment of the present invention is shown. Figure 2 , the method comprises the following steps:

[0046] Step S202 : Online monitoring of the operating parameters of the target roll and the operating status parameters corresponding to the operating parameters is performed to determine the real-time multi-source sensor stream data of the target roll.

[0047] Step S204 : performing variational modal decomposition processing on the real-time multi-source sensor stream data to determine the time-frequency expressions corresponding to the real-time multi-source sensor stream data in multiple modalities.

[0048] Multi-source sensor data, such as roll vibration, temperature, and current, suffer from strongly coupled noise and transient fluctuations in operating conditions. Conventional filtering methods struggle to effectively separate noise from impact signatures. Existing technologies use fixed-threshold filtering, which is inadequate for dynamic conditions, while sliding average filtering smooths out sudden anomalies. Wavelet denoising requires pre-defined basis functions and is sensitive to shifts in noise distribution. To address these issues, the present invention employs variational mode decomposition coupled with dynamic entropy weight filtering.

[0049] In practice, embodiments of the present invention process raw multi-source sensor data through variational modal decomposition. In one implementation, the number of modes and initial center frequencies are preset, and then a variational constrained optimization process is used to determine the time-frequency representation of each mode. This process automatically adapts to the frequency band characteristics of the sensor signals, decomposing strongly coupled mixed signals such as vibration and temperature into a set of independent modal components.

[0050] Step S206 , calculating the energy entropy value of the time-frequency expression, and calculating the entropy weight coefficient corresponding to each mode based on the energy entropy value.

[0051] The energy entropy value of each modal component can reflect the significance of the modal impact feature. The embodiment of the present invention converts the energy entropy value into an entropy weight coefficient, dynamically weights each modal component and reconstructs the signal to filter out baseline drift and high-frequency noise. At the same time, the key information of the transient impact is retained, and the signal vector filtered by entropy weight is output to eliminate baseline drift and high-frequency noise, while retaining the transient impact feature of the roller. Among them, the noise component corresponding to the high entropy value mode is suppressed, and the effective impact feature contained in the low entropy value mode is enhanced. Specifically, the steps for calculating the entropy weight coefficient are as follows:

[0052]

[0053] in, For the The entropy weight coefficient of each mode, For the The energy entropy of the mode, For the The energy entropy of the mode. Taking the energy entropy of a mode as an example, the calculation method is expressed as . is the instantaneous energy ratio, which is calculated as ; is the total length of the time series; For logarithmic functions, the default base is a natural constant.

[0054] It should be noted that energy entropy Quantitative decomposition of modal components The transient shock concentration, The smaller it is, the more significant the impact characteristics are. It should also be noted that the entropy weight coefficient The term increases the weight of the low-entropy mode containing effective shocks and attenuates the weight of the high-entropy mode dominated by noise. While retaining the transient shock of the roller, it adaptively suppresses the low-frequency baseline drift and high-frequency white noise. It should also be noted that the energy entropy Based on instantaneous energy ratio It is calculated without the need for preset thresholds and can adapt to changes in working conditions.

[0055] Step S208 : filtering the real-time multi-source sensor stream data of each modality according to the entropy weight coefficient to determine a noise reduction signal of the real-time multi-source sensor stream data.

[0056] Furthermore, the embodiment of the present invention also uses KL divergence-constrained variational modal decomposition to adaptively separate signal frequency bands and maintain distribution consistency, and then dynamically reconstructs based on modal energy entropy to suppress high-entropy noise modes and enhance low-entropy impact characteristics, thereby realizing dynamic adaptive filtering of multi-source sensor data.

[0057] Among them, the KL divergence constraint is used to ensure the consistency of data distribution before and after decomposition to avoid feature distortion. The final output is the set of filtered modal components that retain the roller impact characteristics and the adaptively optimized center frequency, which is expressed as:

[0058]

[0059] Where, For the The time series data vector of each sensor, such as vibration acceleration signal; Represents the sensor dimension index, , is the total number of sensors; is the first The set of filtered modal components corresponding to the sensors can be expressed as ; For the Sensor No. The decomposed modal components of the variational modes.

[0060] is the total number of modes of variational mode decomposition, which is adaptively selected by the center frequency convergence criterion. When the center frequency of the newly added mode is less than the preset threshold from the existing mode, the decomposition is stopped to avoid overfitting. For example, the preset threshold is ; is the mode index of variational mode decomposition; For the The center frequency of the variational mode decomposition corresponding to each sensor is dynamically adjusted during the optimization process to match the dominant frequency band of the sensor signal. Indicates about variables and Minimization operation of ; It is the L2 norm, which is calculated in the same way as the Euclidean distance; is the time variable index; is the time derivative operator; is the Dirac function, which is used to construct the derivative of the analytical signal; is a natural constant; is the modal correlation weight, which regulates the degree of noise separation, such as, .

[0061] express The probability distribution of can be obtained through kernel density estimation or empirical distribution function, which is used to describe the statistical characteristics of the data, such as skewness and kurtosis; express The probability distribution of can be obtained by fitting after variational mode decomposition to ensure that the data distribution before and after decomposition is consistent and avoid feature distortion; is the KL divergence calculation function; It is a convolution operator, which is a time domain convolution operation used to generate analytical signals and eliminate negative frequency components; Is an imaginary unit.

[0062] It should be noted that The term serves as a KL divergence constraint term, forcing the data distribution to be consistent before and after decomposition, avoiding false features due to modal decomposition, and being able to maintain the time series data vector The statistical characteristics of the signal, such as skewness and kurtosis, are preserved while suppressing the distortion of impact characteristics caused by frequency band aliasing, especially protecting high-frequency transient components.

[0063] It should also be noted that in In the term, the time derivative operator Used to capture the instantaneous rate of change of the signal and enhance the local sensitivity of transient impact characteristics. The term characterizes the convolution to construct the analytical signal, where the Dirac function The term preserves the signal amplitude. The term is used as the imaginary part to convert the real signal into an analytical signal, which is convenient for frequency domain processing. The signal is modulated to baseband so that the center frequency It can be optimized to match the dominant frequency band of the sensor signal. The various parts cooperate to achieve adaptive frequency band separation. The time derivative strengthens the transient characteristics. The convolution constructs the analytical signal to ensure a unilateral spectrum. The complex exponential modulation dynamically adjusts the center frequency. It can effectively retain the high-frequency impact characteristics and suppress the frequency band aliasing caused by operating condition fluctuations.

[0064] Furthermore, the real-time multi-source sensor stream data of each modality is filtered based on the entropy weight coefficient to determine the noise reduction signal of the real-time multi-source sensor stream data. In one embodiment, it is represented by the following formula:

[0065]

[0066] Where, For the The signal vector of each sensor is filtered by entropy weight; is the total number of modes of variational mode decomposition; For the Sensor No. The decomposed modal components of the variational modes.

[0067] In one embodiment, the embodiment of the present invention also verifies the advantages of the variational mode decomposition coupled dynamic entropy weight filtering method proposed in the present invention in retaining key impact characteristics. Figure 3 A schematic diagram showing the effect of a variational mode decomposition coupled dynamic entropy weight filtering method is shown. Figure 3 The embodiment of the present invention is specifically aimed at the vibration acceleration signal of the roller bearing seat. The signal contains four typical transient impact events (representing the characteristic vibration caused by wear) and complex environmental noise (including Gaussian noise and periodic working condition interference). The gray curve in the figure shows the original acquisition signal. It can be seen that the impact characteristics are completely submerged by the noise. The orange curve shows the fixed threshold filtering result. Although this method suppresses some noise, it incorrectly filters out the real impact signal below the threshold (especially the second narrow pulse) and produces signal truncation distortion. The green curve is the sliding average filtering effect. This method is over-smoothed, resulting in the disappearance of the steep edges of all impact pulses (especially the first and second narrow pulses). The rising edges of the three shocks become gentle slopes), with severe loss of transient features. The purple curve shows the wavelet denoising result, with obvious periodic noise still remaining and false oscillations occurring after the fourth shock. In contrast, the method of the present invention (red curve) perfectly preserves the complete form of the four shock pulses (especially the steep rising edge of the second narrow pulse and the fourth shock), while completely suppressing background noise and periodic interference. The signal baseline is flat and drift-free, proving that the present invention achieves precise separation of shock features and noise through KL divergence-constrained variational mode decomposition and dynamic weighting based on energy entropy, overcoming the fundamental contradiction between transient feature preservation and noise suppression in conventional methods.

[0068] Step S210 , converting the noise reduction signal into an analytical signal through a preset nonlinear transformation algorithm, capturing the frequency mutation of the noise reduction signal caused by the transient impact, and determining the instantaneous frequency eigenvector of the noise reduction signal.

[0069] Single time-domain or frequency-domain features cannot fully characterize the frequency band energy migration, transient response, and non-Gaussian characteristics caused by wear. In existing technologies, the Fast Fourier Transform (FFT) ignores temporal correlation, and the short-time FFT suffers from inconsistent time-frequency resolution. Regarding step S106, the present invention constructs a three-dimensional joint feature space, extracts the transient frequency of the nonlinear transformation algorithm to capture the transient impact of the data, calculates the FFT spectral entropy to quantify the change in frequency band energy concentration, and combines the absolute value of the third-order central moment to enhance the sensitivity to non-Gaussian vibration. Finally, horizontal splicing is performed to form a "time-frequency-statistics" tensor.

[0070] In one embodiment, the preset nonlinear transformation algorithm is the Hilbert transform algorithm. Accordingly, this embodiment of the present invention applies the Hilbert transform to the sensor data after dynamic adaptive filtering, constructing an analytical signal to obtain the signal's time-varying frequency components. This embodiment of the present invention accurately captures frequency fluctuations caused by transient impacts on the roller, ultimately outputting a transient frequency eigenvector that characterizes the local time-frequency characteristics of the vibration signal, providing a transient response basis for wear state analysis.

[0071] In specific implementation, the instantaneous frequency eigenvector is determined by the following formula:

[0072]

[0073] Where, is the time differential operator, which is realized by discrete difference; is the Hilbert transform; is the instantaneous frequency feature vector. It should be noted that when extracting the instantaneous frequency feature vector When the wear signal is detected, the Hilbert transform is used to convert the real signal into an analytical signal, accurately capturing the frequency mutation caused by transient impact, such as the high-frequency oscillation caused by microcracks in the roller, and then providing time-varying frequency components to enhance the characterization capability of wear transient response, which is better than the fixed window spectrum method.

[0074] Step S212 , calculating the energy spectrum of the noise reduction signal and determining the frequency distribution entropy value of the noise reduction signal; and determining the spectral entropy feature of the noise reduction signal based on the energy distribution disorder represented by the frequency distribution entropy value.

[0075] Specifically, the present invention uses a fast Fourier transform to calculate the energy spectrum of filtered data and uses this energy spectrum to calculate the frequency distribution entropy. Furthermore, by quantifying changes in frequency band energy concentration, it detects energy migration caused by wear and outputs a spectral entropy feature that reflects the degree of energy dislocation in the frequency domain. High entropy values ​​indicate a uniform distribution, while low entropy values ​​correspond to concentrated energy. Specifically, the steps for calculating the spectral entropy feature are as follows:

[0076]

[0077] Where, It is the spectrum entropy feature, which is a scalar that characterizes the energy distribution disorder of the entire spectrum. In the early stage of wear, the frequency band energy is concentrated, and the spectrum entropy feature When the wear is intensified, the energy will diffuse. Big. Among them, is the energy proportion probability of the fth frequency, and the calculation method is expressed as ; is the fast Fourier transform spectrum coefficient, and the output signal vector Perform fast Fourier transform to calculate it.

[0078] In step S214 , for the signal components of the noise reduction signal in multiple sensor dimensions, the third-order center distance feature corresponding to the signal component of each sensor dimension is calculated respectively; the third-order center distance feature is used to describe the data skewness of the current sensor dimension.

[0079] The embodiment of the present invention calculates the third-order central moment of the vibration signal through absolute value cube operation to enhance the sensitivity to outliers, capture the non-Gaussian vibration characteristics caused by wear, and output the skewness feature that characterizes the asymmetry of the signal distribution. It can be expressed as:

[0080]

[0081] Where, The cubic operation is used to amplify abnormal values, such as transient shock caused by wear, and enhance the sensitivity to non-Gaussian vibration, such as impact cracks, which is better than the second-order moment using square operation. The t-th moment is a discrete sampling point, and the unit is milliseconds. The filtered data of the k-th dimension sensor at time t; is the mean of the filtered data, and the calculation method is expressed as ; It is the third-order central moment characteristic, which represents the skewness of the data.

[0082] In step S216, the instantaneous frequency feature vector, spectral entropy feature, and skewness feature are horizontally spliced ​​along the sensor dimension of the noise reduction signal to form a three-dimensional fusion feature matrix, thereby forming a three-dimensional feature tensor based on "time-frequency-statistics".

[0083] Among them, the extracted instantaneous frequency feature vectors, spectral entropy features, and skewness features can be horizontally spliced ​​along the sensor dimension to form a three-dimensional fusion feature matrix of the noise reduction signal. This matrix form can characterize the transient response, frequency band migration, and nonlinear characteristics of roll wear, and can maintain the independence of the physical meaning of each feature. Specifically, the three-dimensional feature tensor based on "time-frequency-statistics" is expressed as follows:

[0084]

[0085] Where, is the three-dimensional fusion feature matrix; Indicates that 、 、 The three eigenvectors are combined into a matrix by horizontal splicing to form a three-dimensional feature tensor of "time-frequency-statistics".

[0086] Furthermore, the three-dimensional feature tensor is subjected to data recognition processing through a pre-built model, and a wear state prediction probability vector corresponding to the three-dimensional feature tensor is output. In an embodiment of the present invention, a new processing mechanism is proposed in the embodiment of the present invention, which embeds a rotation cycle perception mechanism in the LSTM forget gate and introduces a phase modulation factor, so that the model can dynamically enhance the memory capacity under the wear-sensitive phase, and significantly improve the wear trend modeling and early detection sensitivity under periodic load scenarios. Furthermore, a two-dimensional soft threshold residual block is constructed, and dynamic compression of redundant information and enhancement of weak wear signals are achieved through channel-dimensional entropy weight learning + time-dimensional local amplitude perception, effectively solving the problem of weak features being submerged in deep networks. Furthermore, the topological constraints of the feature space are introduced into the wear classification task, and the Hinge loss driven by cross entropy and label difference is combined to establish a consistent mapping between the wear level and the geometric structure of the feature space, solving the problems of fuzzy levels and unclear boundaries in traditional classification methods.

[0087] In step S218, the three-dimensional feature tensor is input into a pre-built high-frequency welded pipe roller wear state monitoring model, and the high-frequency welded pipe roller wear state monitoring model performs data recognition processing on the three-dimensional feature tensor, and outputs a wear state prediction probability vector corresponding to the three-dimensional feature tensor.

[0088] For specific implementation, refer to steps 1 to 5 below.

[0089] 1) The three-dimensional feature tensor is input into the high-frequency welded pipe roller wear state monitoring model. The long-range wear trend of the three-dimensional feature tensor is modeled based on the periodic memory gating mechanism of the high-frequency welded pipe roller wear state monitoring model, and the first target feature of the three-dimensional feature tensor is determined.

[0090] Roller wear is associated with periodic load variations, but standard LSTMs struggle to model the relationship between rotation period and wear phase. Furthermore, the LSTM forget gate lacks phase awareness, resulting in delayed response under periodic loads and the loss of long-range correlation features. This embodiment of the present invention dynamically adjusts the memory strength of the forget gate of a preset LSTM unit using an exponential decay term based on the deviation between the current phase position and the wear-sensitive phase angle. This performs periodic sensing on the three-dimensional feature tensor and determines the forget gate output for the three-dimensional feature tensor. The forget gate output is used to represent the first target feature of the three-dimensional feature tensor.

[0091] In a specific implementation, the present invention embeds the roller rotation cycle parameter in the forget gate of the LSTM unit to determine the cycle attenuation factor, and dynamically adjusts the memory strength through an exponential term. Specifically, the wear-sensitive phase strengthens feature retention, while the insensitive phase accelerates noise forgetting. Specifically, the memory strength of the forget gate is dynamically adjusted by adopting an exponential decay term based on the deviation between the current phase position and the wear-sensitive phase angle. When the current phase position is aligned with the wear-sensitive phase angle, feature retention is strengthened; when in the insensitive phase, noise forgetting is accelerated. The final output is a modulated forget gate signal, realizing cycle-aware memory control, expressed as:

[0092]

[0093] Where, is the output of the forget gate at time t (that is, the first target feature mentioned above), which controls the degree of retention of historical memory. is the Sigmoid activation function, which compresses the output to interval; is the weight matrix of the forget gate, which is a trainable parameter; H t-1 is the hidden state at time t−1, representing the network state vector at time t-1, encoding historical temporal information, such as long-term wear trends, for gating decisions; is the fusion feature at the tth moment, which is the three-dimensional fusion feature matrix No. The slice vector at the moment is used as the input feature of the LSTM at the current moment; is the bias vector of the forget gate, which is a trainable parameter; is the Hadamard product; is the periodic attenuation strength coefficient, which regulates the influence of the periodic alignment degree, such as, =0.05; It is a modulo operation used to align the roller rotation period; is the roller rotation period; is the wear-sensitive phase angle.

[0094] It should be noted that The term represents the current moment in the roller rotation cycle The phase position in Term characterization phase and wear sensitive phase angle The deviation in The nearby forget gate attenuates weakly, strengthening the wear feature memory, and staying away from It is also necessary to note that the conventional LSTM cannot distinguish between wear-sensitive areas and non-sensitive areas under periodic loads because the forget gate has no phase perception ability. The project combines prior knowledge of rotation phase to solve the response lag problem and improve the sensitivity of early wear detection.

[0095] The embodiment of the present invention also updates the cell state and hidden state of the preset LSTM unit based on the output of the forget gate, so as to transmit the long-term wear trend in combination with the gated state update. In the specific implementation, the modulated forget gate, input gate and output gate are jointly updated to update the cell state and hidden state of the LSTM. Among them, the input gate controls the writing degree of the candidate cell state, and the candidate cell state generates a new feature vector based on the fusion feature of the current moment and the hidden state of the previous moment. The updated cell state fuses the historical memory with the current input, and the output gate controls the degree of cell state output to the hidden state, filters irrelevant noise, and the final output hidden state vector carries the filtered time series features and is passed to the downstream network, which is expressed as:

[0096]

[0097]

[0098]

[0099]

[0100]

[0101] Where, is the input gate output at time t, controlling the state of the candidate cell The degree of writing determines the retention strength of the current feature; is the candidate cell state at time t, which is a new feature vector generated based on the fusion feature at time t and the hidden state vector at time t-1, representing the potential update information; is the cell state updated at time t, encoding the long-range wear trend; is the cell state updated at time t-1; is the output gate at time t, controlling the cell state Output to The degree of hidden state at the moment, filtering irrelevant noise.

[0102] is the hidden state vector at time t, carrying the filtered temporal features and passing them to the downstream network; is the weight matrix of the input gate, which is a trainable parameter; is the weight matrix of the candidate state, which is a trainable parameter; is the weight matrix of the output gate, which is a trainable parameter. is the bias vector of the input gate, which is a trainable parameter; is the bias vector of the candidate state, which is a trainable parameter. is the bias vector of the output gate, which is a trainable parameter; is the hyperbolic tangent activation function.

[0103] It should be noted that the conventional LSTM forget gate has no phase perception capability and cannot distinguish between wear-sensitive areas and non-sensitive areas under periodic loads such as roller rotation, resulting in response lag and feature loss. Item encoding roller rotation period and wear-sensitive phase angle ,when Hours, representing sensitive phase, weak attenuation, The term is close to 1, strengthening the feature memory, and when When it is large, it indicates an insensitive phase with strong attenuation. The term is close to 0, which accelerates noise forgetting. Based on this, it can effectively solve the response lag problem, improve the sensitivity of early wear detection, and reduce computational redundancy.

[0104] 2) Calculate the residual features of the three-dimensional feature tensor and, based on the residual features, generate the channel-dimensional entropy weight coefficient and the time-dimensional local amplitude mean in parallel. Based on the channel-dimensional entropy weight coefficient and the time-dimensional local amplitude mean, perform two-dimensional soft threshold filtering on the three-dimensional feature tensor to enhance the sensitive and weak features representing wear in the three-dimensional feature tensor.

[0105] In deep networks, subtle wear features are easily overwhelmed by redundant information. Conventional residual connections lack an adaptive feature selection mechanism, and conventional channel attention ignores temporal correlations, resulting in the suppression of key temporal features. This paper constructs a channel-time dual-dimensional soft-threshold residual block. This block first calculates residual features and concurrently generates channel-dimensional entropy weights and temporal local amplitude means. It then shrinks irrelevant features using a dual-dimensional threshold and uses a sign function to preserve polarity, enhancing wear-sensitive responses.

[0106] The specific implementation includes the following steps:

[0107] a- Perform two layers of convolution on the input features, add the convolution results to the original input features, and finally output the residual feature tensor, which is expressed as:

[0108]

[0109] Where, is the residual input feature tensor at the tth moment, which is equal to the output of the front-layer LSTM network. The hidden state vector at time ; It is a residual transformation function, which includes two layers of convolution operations. The first layer is a 1×1 convolution, which mainly plays a role in dimensionality reduction. The second layer is a 3×1 temporal convolution, which mainly plays a role in extracting local features. The activation function is the ReLU activation function. is the set of convolutional layer weight parameters, including the convolution weight matrices of the first and second layers of the two convolutional layers corresponding to the residual transformation function; is the residual feature tensor at time t, which retains the original input and convolution transformation features. It should be noted that the residual feature tensor It contains the low-order information of the original input and the high-order abstract features extracted by the convolution transformation, thereby realizing the residual connection of the features.

[0110] b- Parallel calculation of the threshold coefficients for the channel and time dimensions. In the specific implementation, the channel-dimensional soft threshold coefficient vector is obtained by processing the feature amplitude after global average pooling through a multi-layer perceptron, reflecting the importance of each channel; the time-dimensional feature amplitude mean is obtained by calculating the L1 norm mean of the feature amplitude within a local time window, representing the intensity of the temporal activity within the window. The final output is the channel-dimensional soft threshold coefficient vector and the time-dimensional feature amplitude mean, expressed as:

[0111]

[0112]

[0113] Where, is the channel-dimensional soft threshold coefficient vector, generated by multi-layer perceptron and global average pooling; Represents the absolute value of the residual feature tensor at time t, which is an element-level calculation operation used to quantize the feature amplitude; is an exponential function; It is a channel-dimensional multilayer perceptron, that is, a multi-layer fully connected neural network, and the number of layers can be set to 3; is the global average pooling operation.

[0114] is the mean of the time dimension feature amplitude at the tth moment, in the window Internal calculation; is the time window radius, which controls the local time range, such as =5; is the L1 norm, used to calculate the feature amplitude; For the The residual feature tensor at time t; is the time index within the time window, which is a discrete sampling point with a range of ; is the total number of moments in the window, and the calculation method is expressed as .

[0115] c- Shrink the residual features based on the calculated channel-dimensional soft threshold coefficient vector and the mean time-dimensional feature amplitude. In its implementation, the difference between the feature amplitude and the two-dimensional threshold is first calculated. The noise components below the threshold are then eliminated by taking the maximum value of this difference and zero. The sign function is used to preserve the original polarity of the features. Finally, the residual output features are output, enhancing the response of wear-sensitive areas and suppressing interference from irrelevant information. This is expressed as:

[0116]

[0117] Where, is a sign function, which is an element-wise operation used to preserve characteristic polarity; is the outer product broadcast operation, Scalar Expand to Same dimension. Indicates taking the larger value between the input value and 0; is the residual output feature at the tth moment, which enhances the wear-sensitive feature response. It should be noted that the channel-dimensional soft threshold coefficient vector Realize channel-dimensional adaptive feature selection and use the time-dimensional feature amplitude mean Capturing the local feature strength in the time dimension, the dual-dimensional threshold synergistically compresses noise and redundant information, solving the problem of weak wear features being drowned in deep networks. It should also be noted that the channel-dimensional soft threshold coefficient vector Learning channel importance and time dimension feature amplitude mean through multi-layer perceptron The local window amplitude mean is calculated to capture the transient intensity. The combination of the two effectively solves the problem that conventional methods such as SENet (Squeeze-and-Excitation-Networks) only consider the channel dimension and ignore the temporal locality. The temporal dimension feature amplitude mean Characterize the time window to dynamically perceive transient events, such as wear impact, to avoid over-smoothing caused by fixed thresholds, and to coordinate adaptively compress noise such as working condition fluctuations in two dimensions to enhance the significance of weak wear features. It should also be noted that Through the outer product broadcast, the channel-time dual-dimensional threshold is realized, which is different from the single-dimensional soft threshold in conventional technology. The dual-dimensional joint dynamic suppression redundancy is combined with The item retains significant features and can protect weak wear features such as early cracks in the deep network to avoid the weak wear features being submerged.

[0118] 3) Concatenate the first target feature with the cell state of the preset LSTM unit along the feature dimension to generate a wear state gating vector of the first target feature based on the time scale.

[0119] 4) Extract the statistical feature vector of the wear state gating vector at each time scale, aggregate the statistical feature vectors at each time scale, and output a multi-scale weighted fusion feature vector.

[0120] Wear requires the fusion of microscopic transient features and macroscopic trends, but conventional multi-scale fusion and other weighted splicing result in weak features being submerged. Conventional attention mechanisms do not establish the association between wear type and physical scale, making it difficult to adaptively distinguish scale sensitivity. The present invention adopts a scale-aware gating strategy to splice residual features with LSTM cell states to generate a scale gating vector. The residual features are then pooled at short / medium / long scales, weightedly fused according to the gating weights, and the wear probability is output, so that the network automatically focuses on short-scale transient features at the initial stage of wear. In specific implementation, the residual output features and LSTM cell states are first spliced ​​along the feature dimension. The spliced ​​features are processed by linear transformation and sigmoid activation function to generate a scale gating vector. The components of the scale gating vector characterize the preference strength of different wear states for different time scale features, which is expressed as:

[0121]

[0122] Where, The input feature of the prediction layer at time t is equal to the updated cell state at time t output by the previous LSTM network. ; It is a splicing operation along the feature dimension; is the gate weight matrix, which is a trainable parameter; is the gate bias vector, which is a trainable parameter; is the scale gate vector at the tth moment. It should be noted that the feature Encoding long-range wear trends, characteristics of macro-wear The transient characteristics of micro cracks are included, and the connection between the two establishes an explicit mapping between the wear state and the physical scale, making the scale gate vector Dynamically perceive the sensitivity of different wear types to scale. Different from the conventional attention mechanism that only weights the importance of features, the scale gate vector By fusing macroscopic and microscopic features, the spatiotemporal correlation of wear evolution is embedded in the gating signal. When including early wear trends, the scale gate vector Automatically enhance the weight of short-scale features to improve the detection rate of micro cracks.

[0123] Furthermore, a temporal global average pooling operation is performed on each time scale of the residual output feature to extract the statistical feature vectors at each scale. In one embodiment, the residual output features can be processed according to three scales: short, medium, and long. The component sub-vectors corresponding to each scale in the scale gating vector can be used to weight the pooled feature vectors of the corresponding scale. Finally, the weighted pooled feature vectors of each scale are aggregated to finally output a multi-scale weighted fusion feature vector, retaining the scale information most relevant to the current wear state. It can be expressed as:

[0124]

[0125] Where, is the temporal pooling feature vector of the s-th scale, and the feature is output from the residual by global average pooling with a step size of s , where s∈{4,8,16} corresponds to the short, medium and long time scales respectively; is the scale gate vector Corresponding to The component sub-vector of the scale; s is the pooling scale parameter, which controls the time window size, for example, s=4 represents a short scale, s=8 represents a medium scale, and s=16 represents a long scale; It is a multi-scale weighted fusion feature vector that retains the scale information most relevant to the current wear state.

[0126] 5) Based on the preset scale gating vector model, the physical association between the multi-scale weighted fusion feature vector and the wear type is modeled to determine the wear state prediction probability vector corresponding to the three-dimensional feature tensor.

[0127] In the specific implementation, the multi-scale weighted fusion feature vector is input into the fully connected layer for transformation, and then processed by the normalized exponential function to output the wear state prediction probability vector. The result reflects the probability of the roller being in different wear levels and is expressed as:

[0128]

[0129] Where, is the classification weight matrix, which is a trainable parameter; is the classification weight bias vector, which is a trainable parameter; To normalize the exponential function, the output is mapped into a probability distribution; is the wear state prediction probability vector. Among them, conventional multi-scale fusion splicing of features with equal weights can easily cause weak micro features to be submerged. The above-mentioned preset scale gating vector of the embodiment of the present invention learns the preference of the wear state for the corresponding time scale through the item splicing operation at each time scale. For example, the scale gating vector used is Modeling the physical association between wear type and characteristic scale, scale gating vector The amount of The stitching operation learns the wear state of the first For example, for micro cracks caused by short-term impact, As the weight increases, the short-scale features are strengthened. , for the macro wear of the long-term trend, As the weight increases, the long-scale features are strengthened. , which enables the network to automatically focus on short-scale features, such as transient frequency mutations caused by cracks, in the early stages of wear, and on long-scale trends in the later stages of wear.

[0130] Step S220: evaluating the wear state of the target roller according to the wear state prediction probability vector.

[0131] This step is similar to the above-mentioned step S110 and will not be described in detail here. Furthermore, the wear levels of mild, moderate, and severe are continuous, but the imbalance of samples leads to blurred decision boundaries. Conventional classification loss ignores the correlation between levels, and ordinal regression lacks feature space topological constraints. Samples of similar levels may be over-separated or confused. The embodiment of the present invention also uses the pre-calculated feature space topological constraint loss to train the high-frequency welded pipe roller wear state monitoring model. In the specific implementation, the classification cross entropy loss and the feature space topological constraint loss are combined to force samples of similar levels to cluster in the feature space to maintain topological continuity.

[0132] Specifically, the embodiment of the present invention performs weighted fusion of classification cross entropy loss and feature space topology constraint loss according to set weights to form the final multi-task joint loss objective function, which is used to simultaneously optimize classification accuracy and feature distribution rationality, expressed as:

[0133]

[0134] Where, is the feature regularization loss weight, which regulates the strength of topological constraints. ; It is a multi-task joint loss. Among them, based on the one-hot encoding of the wear state prediction probability vector and the wear state true label, the cross entropy loss between the predicted probability and the true label can be calculated, and the classification cross entropy loss is defined as . Furthermore, an embodiment of the present invention calculates the feature space topology constraint loss through the following steps: 1) First, calculate the feature vector distance between each feature vector of the multi-scale weighted fusion feature vector, as well as the benchmark distance of the multi-scale weighted fusion feature vector based on the label difference. 2) Calculate the weight coefficients of the feature vectors of adjacent wear levels in the multi-scale weighted fusion feature vector, and perform adjacent wear level weight constraints on the feature vectors. 3) Use the preset Hinge loss function to constrain the deviation of the feature vector distance and the benchmark distance, and combine the weight coefficient to calculate the feature space topology constraint loss corresponding to the high-frequency welded pipe roller wear state monitoring model.

[0135] In its implementation, the present invention adaptively constrains feature space distances based on sample label differences. First, the distance between the sample's aggregated feature vectors and a baseline distance based on label differences are calculated. Then, a sample pair weight coefficient is used to assign greater weight to pairs of adjacent wear level samples. Finally, a Hinge loss function is used to constrain the deviation between the feature distance and the baseline distance. This loss term aims to maintain the physical interpretability of the feature space and enhance the continuity of wear levels, keeping samples with similar wear levels close together in the feature space and samples with significantly different wear levels further apart.

[0136] Specifically, the calculation method of the feature space topology constraint loss is expressed as:

[0137]

[0138] Where, is the feature space topology constraint loss. is the total number of valid sample pairs in the batch, and the valid sample pairs are defined as the label difference sample pairs to avoid interference from irrelevant samples. is the boundary margin, controlling the distance tolerance interval, such as, ; is the Hinge loss function, which is equivalent to , used to constrain distance deviation. is the label difference benchmark distance. is the aggregated feature vector of the i-th sample, which is the multi-scale weighted fusion feature vector of the i-th sample at all times Aggregated, the aggregation method uses the global average pooling operation, expressed as . is the global average pooling operation; is the multi-scale weighted fusion feature vector; is the multi-scale weighted fusion feature vector at the tth moment. is the aggregated feature vector of the jth sample; i is the sample index; j is the sample index different from i. Represents the set of multi-scale weighted fusion feature vectors of the i-th sample at all times from t=1 to t=T.

[0139] Label difference benchmark distance The calculation method is expressed as:

[0140]

[0141] in, is the scaling factor, such as, ; Indicates the original label value, such as mild = 0, moderate = 1, severe = 2, is the original label value of the i-th sample, is the original label value of the jth sample. The calculation method of the weight coefficient is expressed as:

[0142]

[0143] in, is the sample pair weight coefficient, which makes the adjacent wear level sample pairs have a larger weight. is the weight attenuation coefficient, such as, It should be noted that conventional ordinal regression has no feature space constraints, and the present invention uses sample pair weight coefficients and label difference benchmark distance Jointly strengthen the clustering of adjacent level samples, such as, When the distance is small, the fuzzy decision boundary caused by sample imbalance is solved, so that the feature space maintains topological continuity. Furthermore, the number of deep network parameters is large, and non-stationary training data can easily cause the optimization to fall into local optimality. Conventional regularization compresses parameters equally, which damages the expression ability of key channels. In addition, the Adam optimizer is sensitive to sudden changes in working conditions, and weight decay ignores the importance differences of wear-sensitive channels. The embodiment of the present invention also adopts gradient normalization and selective spectral norm constraints, normalizes the gradient of the sliding window to eliminate non-stationary fluctuations, and then dynamically detects the weight spectral norm. Sign-preserving regularization is applied only to super-threshold redundant channels to protect the wear-sensitive feature extraction capability. The specific steps are as follows:

[0144] a- Normalize the original gradient in the time dimension. Specifically, use the historical gradient amplitude statistics in the sliding time window to normalize the current original gradient and output the normalized gradient to improve training stability and avoid parameter update oscillation caused by non-stationary working conditions. It is expressed as:

[0145]

[0146] Where, To standardize the gradient and eliminate the interference of non-stationary working conditions; Multi-task joint loss Parameters The original gradient of is a trainable parameter, in the form of a parameter set, representing the set of all trainable parameters; It is the iteration step index within the gradient smoothing window, which is a discrete value; is the length of the gradient smoothing time window, for example, taking the most recent 50 iterations, indicating that the gradient statistics are calculated in the 50 iterations before the current iteration; For parameters In the historical iteration The original gradient of the moment; is a numerical stability constant that prevents the denominator from being zero, such as, .

[0147] It should be noted that in the process of normalized gradient calculation, The term represents the sliding window statistical term, which is used to capture historical fluctuations. Compared with the conventional Adam optimizer which only normalizes the current gradient, it can effectively improve the convergence stability under non-stationary data such as sudden changes in working conditions.

[0148] b- Calculate the spectral norm of the weight matrix in real time and determine whether the spectral norm exceeds the preset activation threshold. If it exceeds the threshold, the indicator function is used to mark the weight matrix as requiring regularization constraints, which is expressed as:

[0149]

[0150] Where, is the weight matrix The spectral norm of is the set of convolutional layer weight parameters The weight matrix in ; is the spectral norm activation threshold, such as, ; is the indicator function, In the item, when Time output , triggering regularization.

[0151] c- Update the overall network parameters. First, a basic update is performed using normalized gradients. At the same time, for weight matrices marked as requiring regularization by spectral norm dynamic detection, a selective spectral regularization term based on their sign and L2 norm is additionally applied. This regularization term aims to compress channel parameters identified as redundant. The final updated network parameters suppress overfitting while protecting the expressive power of key features, expressed as:

[0152]

[0153] Where, Update the assignment operation for the parameters; is the basic learning rate, such as, ; is the spectral regularization strength coefficient, such as, ; For parameters The sign function is an element-level calculation operation that preserves the gradient update direction. It should be noted that, unlike the conventional weight decay that compresses all parameters equally, the present invention only uses Time application The incremental value of the term, combined with the spectral norm activation threshold Dynamically identify redundant channels, such as noise-related weights, protect the crack feature extraction layer as a wear-sensitive channel, and improve the expression ability of key features. It should also be noted that L2 regularization may offset the feature distribution. This invention adopts The term is used to maintain the consistency of the physical meaning of the parameters, avoid regularization from distorting the feature space, and improve the robustness of the model. The model iteratively trains by cyclically executing the forward propagation process to calculate the loss and the backpropagation process to update the parameters. Training continues until any of the following stopping conditions is met: 1) the change in the multi-task joint loss value is continuously lower than the preset threshold, for example, the change in the multi-task joint loss value for five consecutive iterations is lower than 0.01; 2) the preset maximum number of iterations is reached, for example, the maximum number of iterations is 5000.

[0154] In summary, the embodiments of the present invention have the following features: 1. A joint filtering method combining KL divergence-constrained variational mode decomposition and energy entropy-weighted reconstruction is proposed, overcoming the limitations of traditional fixed threshold and wavelet methods in handling high-frequency shocks and non-stationary noise, and achieving precise separation of transient shock features from strongly coupled noise. 2. A rotational cycle perception mechanism is embedded in the LSTM forget gate, and a phase modulation factor is introduced, enabling the model to dynamically enhance its memory capacity during wear-sensitive phases, significantly improving wear trend modeling and early detection sensitivity in cyclic load scenarios. 3. A two-dimensional soft-threshold residual block is constructed. Through channel-dimensional entropy weight learning and time-dimensional local amplitude perception, redundant information is dynamically compressed and weak wear signals are enhanced, effectively addressing the problem of weak features being submerged in deep networks. 4. Feature space topology constraints are introduced into the wear classification task, combining cross-entropy with a label-difference-driven Hinge loss to establish a consistent mapping between wear level and feature space geometry, addressing the issues of fuzzy levels and unclear boundaries in traditional classification methods.

[0155] In one embodiment, the present invention also provides a schematic diagram of the modeling capability of the wear state continuity of the feature space topology constraint loss, referring to Figure 4The scatter plot on the left shows the feature space distribution of the model without constraints. It can be seen that the three types of wear state samples (green points for light wear, orange points for moderate wear, and pink points for heavy wear) are in a disordered and mixed state. In particular, the light and moderate samples overlap in a large area in the central area, and some heavy samples are abnormally embedded in the light sample cluster. This chaotic distribution leads to blurred classification boundaries and is the root cause of misjudgment of adjacent states. The scatter plot on the right shows the feature space after topological constraints are applied, showing a clear hierarchical progressive structure. The light wear samples gather in the lower left to form a dense cluster, and the moderate wear samples are distributed in the upper middle area as a transition zone, which is different from the light samples. There is reasonable overlap in the clusters, and the heavily worn samples are concentrated in the upper right quadrant. More importantly, the sample distribution distance strictly corresponds to the degree of wear. The center distance between the light and moderate sample clusters is smaller than the center distance between the moderate and heavy sample clusters, indicating that the dual mechanism of topological constraint loss is at work. The sample pair weight coefficient makes adjacent level samples (such as light-moderate) have stronger mutual attraction, and the label difference baseline distance forces the separation of samples with significant differences (such as light-heavy). Experiments have shown that this constraint successfully transforms the physical continuity of the wear state into the geometric continuity of the feature space, solving the decision boundary distortion problem caused by conventional classification loss.

[0156] In one embodiment, the present invention further provides a schematic diagram of the comprehensive performance of the present invention, referring to Figure 5The embodiment of the present invention experimentally compares 5 technical solutions to evaluate the comprehensive performance of the monitoring model of the present invention in the wear status classification task. The specific analysis is as follows: support vector machine based on statistical features, random forest fused with manual features, standard long short-term memory network, long short-term memory network with attention mechanism and the complete model of the present invention. The 4 groups of colored columns in the figure respectively represent the performance of different methods in light wear (green), moderate wear (orange), severe wear (pink) and average accuracy (blue). It can be seen that the traditional machine learning method (the first two columns) has obvious shortcomings in the recognition of moderate and severe wear, and the performance drops sharply due to the inability to model temporal dependencies. Although the two long short-term memory network improvement schemes (the third and fourth columns) are better than the traditional method, they still have problems in the heavy wear scenario. There is a gap of about 10% in accuracy. The method of the present invention (the rightmost column group) is significantly ahead in all three types of wear states, especially in the identification of severe wear. It is worth noting that the accuracy distribution of the model of the present invention under different wear states is more balanced (the three types of columns are highly close), proving that it has stable cross-state recognition capabilities. Experimental results show that the improved periodic memory gating unit strengthens the feature memory of the wear-sensitive area through rotational phase perception, the deep residual contraction structure protects the weak crack features through the channel-time two-dimensional threshold, and the multi-scale weighted fusion adaptively coordinates the feature contributions of transient impact and macro trends. Experiments have proved that the model of the present invention still maintains high robustness under complex working conditions, solving the performance degradation problem of conventional methods in the identification of severe wear states.

[0157] In one embodiment, the present invention further provides a physical interpretability diagram of a three-dimensional feature tensor, referring to Figure 6The embodiment of the present invention verifies the physical interpretability of the three-dimensional characteristic tensor system of "time-frequency-statistics" through the characteristic evolution law within a continuous working cycle. The three sub-graphs respectively show the changing trends of instantaneous frequency (blue), spectral entropy (red), and third-order central moment (green) within 8 hours. The background color indicates the actual wear state (light wear is light green, moderate wear is light orange, and severe wear is light red). In the instantaneous frequency graph, the curve of the light wear stage is stable and maintained at a low level; after turning to moderate wear, periodic peaks appear (reflecting transient oscillations caused by local spalling), and when entering severe wear, continuous high-frequency fluctuations are shown (corresponding to high-frequency resonance caused by macroscopic cracks). The spectral entropy graph shows an inverse trend. The entropy value is low during light wear (frequency band energy is concentrated), and the curve rises step by step as the wear intensifies (30 in the moderate stage). %, and then increased by 40% in the severe stage), indicating that the energy is diffused into a wide frequency band. The third-order central moment diagram at the bottom shows that the value is stable and close to zero in the mild stage (the vibration signal is Gaussian distributed), a positive deviation begins to appear in the moderate stage (the impact event causes the distribution to be right-skewed), and the severe stage continues to fluctuate at a high level (abnormal impact dominates the signal distribution). The three characteristics all undergo significant jumps at the state transition boundaries (100 hours and 200 hours) and maintain a stable trend within the same state. The strong correlation feature association proves that the instantaneous frequency is sensitive to transient impact (a characterization of crack propagation), the spectral entropy reflects the frequency band energy migration (a manifestation of material fatigue), and the third-order central moment quantifies non-Gaussian vibration (the statistical feature of impact accumulation). The complementarity of three-dimensional features provides a multi-perspective observation basis for wear state monitoring.

[0158] Furthermore, based on the above embodiment, the embodiment of the present invention also provides a high-frequency welded pipe roller wear state monitoring device, Figure 7 The schematic diagram of the structure corresponding to the embodiment of the present invention is shown. Figure 7The device includes: a data acquisition module 100, which is used to perform online monitoring of the operating parameters of the target roll and the operating status parameters corresponding to the operating parameters, and determine the real-time multi-source sensor flow data of the target roll; the operating status parameters include a variety of dynamic physical signals in the high-frequency welded pipe rolling process; a pre-processing module 200, which is used to perform dynamic adaptive filtering processing on the real-time multi-source sensor flow data based on the time-frequency expression of the real-time multi-source sensor flow data in each mode, and determine the noise reduction signal of the real-time multi-source sensor flow data; a data processing module 300, which is used to determine the instantaneous frequency feature vector of the noise reduction signal based on the signal characteristics of the noise reduction signal , spectral entropy features and skewness features; and, based on the instantaneous frequency feature vector, spectral entropy features and skewness features, a three-dimensional feature tensor of the target roller is constructed; wherein, the signal characteristics include frequency mutation characteristics, energy migration characteristics and non-Gaussian vibration characteristics; an execution module 400 is used to input the three-dimensional feature tensor into a pre-constructed high-frequency welded pipe roller wear state monitoring model, perform data recognition processing on the three-dimensional feature tensor through the high-frequency welded pipe roller wear state monitoring model, and output a wear state prediction probability vector corresponding to the three-dimensional feature tensor; an output module 500 is used to evaluate the wear state of the target roller based on the wear state prediction probability vector. The embodiment of the present invention provides a high-frequency welded pipe roller wear state monitoring device, the implementation principle and technical effects of which are the same as those of the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the device embodiment, reference can be made to the corresponding content in the aforementioned method embodiment.

[0159] An embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned Figures 1 to 2 The embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to execute the above Figures 1 to 2 The embodiment of the present invention also provides a structural diagram of an electronic device, such as Figure 8 FIG. 8 is a schematic diagram of the structure of the electronic device, wherein the electronic device includes a processor 81 and a memory 80, the memory 80 stores computer executable instructions that can be executed by the processor 81, and the processor 81 executes the computer executable instructions to implement the above Figures 1 to 2 Either of the methods shown. Figure 8In the illustrated embodiment, the electronic device further includes a bus 82 and a communication interface 83. The processor 81, communication interface 83, and memory 80 are connected via bus 82. Memory 80 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk drive. Communication between the system network element and at least one other network element is achieved via at least one communication interface 83 (which may be wired or wireless). This communication interface may utilize the Internet, a wide area network, a local area network, a metropolitan area network, or the like. Bus 82 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus. It can also be an AMBA (Advanced Microcontroller Bus Architecture) bus. AMBA defines three types of buses, including APB (Advanced Peripheral Bus), AHB (Advanced High-performance Bus), and AXI (Advanced Xtensible Interface). Bus 82 can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, Figure 8 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0160] The processor 81 may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by hardware integrated logic circuits in the processor 81 or by software instructions. The above processor 81 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of the present application can be directly implemented and executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium mature in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor 81 reads the information in the memory and combines its hardware to complete the above Figures 1 to 2 Any of the methods shown.

[0161] The computer program product of a high-frequency welded pipe roller wear condition monitoring method and apparatus provided in an embodiment of the present invention includes a computer-readable storage medium storing program code. The program code includes instructions that can be used to execute the methods described in the aforementioned method embodiments. For detailed implementation, please refer to the method embodiments and will not be described in detail here. Those skilled in the art will clearly understand that, for ease of description and brevity, the specific operating process of the system described above can refer to the corresponding process in the aforementioned method embodiments and will not be described in detail here. If the functions described are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes instructions for causing a computer device (such as a personal computer, server, or network device) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk. In the description of the present invention, it should be noted that the terms "first", "second" and "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. Finally, it should be noted that the above embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that any person skilled in the art who is familiar with this technical field can still modify the technical solutions described in the aforementioned embodiments within the technical scope disclosed by the present invention, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for monitoring the wear state of a high-frequency welded pipe roller, characterized in that: The method comprises: Online monitoring of the target roll's operating parameters and operating status parameters corresponding to the operating parameters is performed to determine real-time multi-source sensor stream data of the target roll; the operating status parameters include multiple dynamic physical signals during high-frequency welded pipe rolling; Based on the time-frequency expression of the real-time multi-source sensor stream data in each modality, performing dynamic adaptive filtering processing on the real-time multi-source sensor stream data to determine a noise reduction signal of the real-time multi-source sensor stream data; Determining, based on the signal characteristics of the noise reduction signal, an instantaneous frequency feature vector, a spectral entropy feature, and a skewness feature of the noise reduction signal; and constructing, based on the instantaneous frequency feature vector, the spectral entropy feature, and the skewness feature, a three-dimensional feature tensor of the target roll; wherein the signal characteristics include frequency mutation characteristics, energy migration characteristics, and non-Gaussian vibration characteristics; Inputting the three-dimensional feature tensor into a pre-built high-frequency welded pipe roller wear state monitoring model, performing data recognition processing on the three-dimensional feature tensor through the high-frequency welded pipe roller wear state monitoring model, and outputting a wear state prediction probability vector corresponding to the three-dimensional feature tensor; The wear state of the target roll is evaluated according to the wear state prediction probability vector.

2. The method according to claim 1, characterized in that The steps of inputting the three-dimensional feature tensor into a pre-built high-frequency welded pipe roller wear state monitoring model, performing data recognition processing on the three-dimensional feature tensor through the high-frequency welded pipe roller wear state monitoring model, and outputting a wear state prediction probability vector corresponding to the three-dimensional feature tensor include: Inputting the three-dimensional feature tensor into the high-frequency welded pipe roller wear state monitoring model, modeling the long-range wear trend of the three-dimensional feature tensor based on the periodic memory gating mechanism of the high-frequency welded pipe roller wear state monitoring model, and determining the first target feature of the three-dimensional feature tensor; Calculating residual features of the three-dimensional feature tensor, and generating, based on the residual features, a channel-dimensional entropy weight coefficient and a time-dimensional local amplitude mean in parallel; performing two-dimensional soft threshold filtering on the three-dimensional feature tensor based on the channel-dimensional entropy weight coefficient and the time-dimensional local amplitude mean, and enhancing sensitive weak features in the three-dimensional feature tensor that represent wear; Splicing the first target feature with the cell state of the preset LSTM unit along the feature dimension to generate a wear state gating vector of the first target feature based on the time scale; Extracting statistical feature vectors of the wear state gating vector at each time scale, aggregating the statistical feature vectors at each time scale, and outputting a multi-scale weighted fusion feature vector; The physical association between the multi-scale weighted fusion feature vector and the wear type is modeled based on a preset scale gating vector, and a wear state prediction probability vector corresponding to the three-dimensional feature tensor is determined.

3. The method according to claim 2, characterized in that The steps of modeling the long-range wear trend of the three-dimensional feature tensor based on the periodic memory gating mechanism of the high-frequency welded pipe roller wear state monitoring model and determining the first target feature of the three-dimensional feature tensor include: An exponential decay term based on the deviation between the current phase position and the wear-sensitive phase angle is used to dynamically adjust the memory strength of the forget gate of the preset LSTM unit, perform periodic perception on the three-dimensional feature tensor, and determine the forget gate output of the three-dimensional feature tensor; the forget gate output is used to characterize the first target feature of the three-dimensional feature tensor.

4. The method according to claim 3, characterized in that The method further comprises: Based on the forget gate output, the cell state and hidden state of the preset LSTM unit are updated.

5. The method according to claim 2, characterized in that The components of the preset scale gating vector at each time scale are determined by learning the preference of the wear state for the corresponding time scale through the term splicing operation.

6. The method according to claim 2, characterized in that The method further comprises: Using pre-calculated feature space topology constraint loss to train the high-frequency welded pipe roller wear state monitoring model; The calculation steps of the feature space topology constraint loss are as follows: Calculating a feature vector distance between each feature vector of the multi-scale weighted fusion feature vector, and a reference distance of the multi-scale weighted fusion feature vector based on label differences; Calculating weight coefficients of feature vectors of adjacent wear levels in the multi-scale weighted fusion feature vector, and performing adjacent wear level weight constraints on the feature vectors; The preset Hinge loss function is used to constrain the deviation between the characteristic vector distance and the reference distance, and the characteristic space topology constraint loss corresponding to the high-frequency welded pipe roller wear state monitoring model is calculated in combination with the weight coefficient.

7. The method according to claim 1, characterized in that The step of performing dynamic adaptive filtering on the real-time multi-source sensor stream data based on the time-frequency expression of the real-time multi-source sensor stream data in each modality to determine a noise reduction signal of the real-time multi-source sensor stream data includes: Performing variational modal decomposition processing on the real-time multi-source sensor stream data to determine time-frequency expressions corresponding to the real-time multi-source sensor stream data in multiple modes; Calculating the energy entropy value of the time-frequency expression, and calculating the entropy weight coefficient corresponding to each of the modes based on the energy entropy value; The real-time multi-source sensor stream data of each modality is filtered according to the entropy weight coefficient to determine a noise reduction signal of the real-time multi-source sensor stream data.

8. The method according to claim 1, characterized in that The step of determining the instantaneous frequency feature vector, spectral entropy feature, and skewness feature of the noise reduction signal according to the signal characteristics of the noise reduction signal comprises: The noise reduction signal is converted into an analytical signal by a preset nonlinear transformation algorithm, the frequency mutation of the noise reduction signal caused by the transient impact is captured, and the instantaneous frequency eigenvector of the noise reduction signal is determined; Calculating the energy spectrum of the noise reduction signal to determine the frequency distribution entropy value of the noise reduction signal; and determining the spectral entropy feature of the noise reduction signal based on the energy distribution disorder represented by the frequency distribution entropy value; For the signal components of the denoised signal in multiple sensor dimensions, the third-order center distance feature corresponding to the signal component of each sensor dimension is calculated respectively; the third-order center distance feature is used to describe the data skewness of the current sensor dimension to determine the skewness feature of the denoised signal.

9. The method according to claim 8, characterized in that The step of constructing a three-dimensional feature tensor of the target roll based on the instantaneous frequency feature vector, the spectral entropy feature, and the skewness feature comprises: The instantaneous frequency feature vector, the spectral entropy feature, and the skewness feature are horizontally spliced ​​along the sensor dimension of the noise reduction signal and combined into a three-dimensional fusion feature matrix to form a three-dimensional feature tensor based on "time-frequency-statistics".

10. A high-frequency welded pipe roller wear state monitoring device, characterized in that: The device comprises: A data acquisition module is used to perform online monitoring of the operating parameters of the target roll and the operating status parameters corresponding to the operating parameters, and determine the real-time multi-source sensor stream data of the target roll; the operating status parameters include various dynamic physical signals during the high-frequency welded pipe rolling process; a preprocessing module, configured to perform dynamic adaptive filtering on the real-time multi-source sensor stream data based on the time-frequency expression of the real-time multi-source sensor stream data in each modality, and determine a noise reduction signal for the real-time multi-source sensor stream data; a data processing module, configured to determine, based on the signal characteristics of the noise reduction signal, an instantaneous frequency feature vector, a spectral entropy feature, and a skewness feature of the noise reduction signal; and construct a three-dimensional feature tensor of the target roll based on the instantaneous frequency feature vector, the spectral entropy feature, and the skewness feature; wherein the signal characteristics include frequency mutation characteristics, energy migration characteristics, and non-Gaussian vibration characteristics; an execution module, configured to input the three-dimensional feature tensor into a pre-built high-frequency welded pipe roller wear state monitoring model, perform data recognition processing on the three-dimensional feature tensor through the high-frequency welded pipe roller wear state monitoring model, and output a wear state prediction probability vector corresponding to the three-dimensional feature tensor; The output module is used to evaluate the wear state of the target roller according to the wear state prediction probability vector.

Images