A method and system for multi-fault diagnosis of a lyocell fiber spinning metering pump

Abstract

This invention relates to the field of intelligent diagnostic technology for textile equipment, specifically to a multi-fault diagnosis method and system for a lyocell fiber spinning metering pump. The method includes the following steps: based on the structure and fluid parameters of the metering pump, a computational fluid dynamics digital twin model is constructed to generate simulated physical field signals under various fault conditions; the physical field signals generated by the metering pump during operation are acquired, and the physical field signals and the simulated physical field signals are fused in the time and frequency domains respectively to construct their respective two-dimensional feature matrices; a feature extraction network with an attention mechanism is constructed, and the feature extraction network is trained using a two-stage training paradigm; the fault diagnosis result is obtained using the two-dimensional feature matrix to be diagnosed through the trained feature extraction network, and the physical component where the fault occurs is located based on the attention mechanism.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology for textile equipment, specifically to a multi-fault diagnostic method and system for a lyocell fiber spinning metering pump. Background Technology

[0002] In the production process of lyocell fiber, the metering pump is the core precision conveying equipment. Its operating status directly affects the flow stability and pressure fluctuation of the spinning solution, which in turn has an important impact on the fiber forming quality. At present, the fault diagnosis of metering pump mainly relies on the following methods: (1) manual experience judgment and regular maintenance; (2) spectrum analysis method based on a single signal (such as vibration or pressure); (3) simple classification model based on traditional machine learning.

[0003] However, existing technologies have the following shortcomings: First, fault data is difficult to obtain, as it is difficult to artificially create faults in industrial settings, resulting in extremely limited data samples that can be used for training.

[0004] Secondly, the signal feature representation capability is insufficient. Traditional methods mostly use single time domain or frequency domain analysis, which cannot fully reflect the complex flow and structural coupling effects inside the metering pump.

[0005] Furthermore, lacking physical mechanism constraints, purely data-driven models are prone to misjudgment under complex working conditions or noise interference, and may even output results that violate the basic laws of fluid mechanics.

[0006] Finally, it has weak anti-interference ability. Due to the existence of pressure fluctuations and system disturbances, the pressure signal is often mixed with a lot of noise, which causes the fault characteristics to be submerged.

[0007] Therefore, there is an urgent need for a multi-fault diagnosis method for metering pumps that can achieve data scalability, sufficient feature representation, physical consistency, and strong anti-interference ability. Summary of the Invention

[0008] To address the shortcomings of existing methods and the needs of practical applications, this invention provides a multi-fault diagnosis method for lyocell fiber spinning metering pumps, comprising the following steps: Based on the structure and fluid parameters of the metering pump, a computational fluid dynamics digital twin model is constructed to generate simulated physical field signals under various fault conditions. The physical field signals generated by the metering pump during operation are acquired, and the physical field signals and the simulated physical field signals are fused in the time and frequency domains respectively to construct their respective two-dimensional feature matrices. A feature extraction network with an attention mechanism is constructed, and a two-stage training paradigm is used to train the feature extraction network. In the first stage, a supervised contrastive loss function is used to perform feature clustering learning on the training set composed of the two-dimensional feature matrices of the simulated physical field signals and / or the two-dimensional feature matrices of the physical field signals. The fault diagnosis result is obtained by using the two-dimensional feature matrix to be diagnosed through the trained feature extraction network, and the physical component where the fault occurs is located based on the attention mechanism.

[0009] Optionally, the computational fluid dynamics digital twin model is constructed to generate simulated physical field signals under various fault conditions, including: Construct the computational fluid dynamics digital twin model, and by adjusting the gear geometric boundary parameters in the digital twin model, inject and generate simulated physical field signals that include at least one of the faults: wear, tooth breakage, and pitting. The process involves adjusting the gear geometric boundary parameters in the digital twin model to inject and generate simulated physical field signals that include at least one of the following faults: wear, tooth breakage, and pitting. Wear failure was simulated by increasing radial clearance by reducing tooth thickness. A broken tooth failure is simulated by cutting off the tooth tip to reduce the tooth height. Pitting corrosion is simulated by creating localized pits in the tooth surface contact area.

[0010] Optionally, the physical field signal is fused in the time and frequency domains to construct the two-dimensional feature matrix, including: Extracting the frequency domain feature sequence of the pressure signal and the time domain feature sequence of the flow signal from the physical field signal; After standardizing the frequency domain feature sequence and the time domain feature sequence, they are folded and recombined into a two-dimensional feature matrix.

[0011] Optionally, after constructing their respective two-dimensional feature matrices and before inputting them into the feature extraction network, the method further includes: A generative adversarial network guided by physical information is used to augment the data of the two-dimensional feature matrix; The generator loss function of the generative adversarial network incorporates a physical consistency penalty term to constrain the generated data to conform to the laws of fluid dynamics. The physical consistency penalty includes at least one of the following: Flow continuity constraints are used to penalize the error between the flow rate in the generated data and the theoretical displacement. Meshing frequency energy constraint is used to force the energy of the pressure frequency domain in the generated data to be concentrated at the meshing frequency and higher harmonics. Phase consistency constraints are used to penalize the phase difference between pressure and flow signals in the generated data that exceeds a preset physical tolerance.

[0012] Optionally, the feature extraction network with an attention mechanism includes: The first layer uses an anisotropic convolution kernel to differentially extract the features of the two-dimensional feature matrix in the time and frequency dimensions; Following the first layer, a dual mechanism of channel attention and spatial attention is integrated to adaptively weight the features of different channels and spatial regions; The two-stage training paradigm includes: In the first stage, based on the training set, the structure other than the projection head connected after the feature extraction network is frozen, and the feature extraction network and the projection head are trained using a supervised contrastive loss function to achieve clustering of similar features in a high-dimensional space. In the second stage, the projection head is removed and the feature extraction network is frozen. A linear classifier is then connected, and the linear classifier is trained using the cross-entropy loss function.

[0013] Optionally, the supervised contrastive loss function used in the first stage satisfies:

[0014] in, This represents the supervised contrastive loss function. This is the set of indices for all samples within the current training micro-batch. For batch size, For anchor point A set of positive samples with the same working condition labels. The two-dimensional feature matrices represent the two-dimensional feature matrices of the anchor point samples respectively. Positive samples and arbitrary comparison sample two-dimensional feature matrix Embedding features of a two-dimensional feature matrix in the projection space, and satisfying of Normalization conditions It is a temperature coefficient used to amplify small similarity differences between samples and control the sensitivity to penalties for difficult negative samples.

[0015] Optionally, before performing the time-domain and frequency-domain fusion processing, the following steps are also included: Based on the rotation cycle of the metering pump, the window length and sliding step size of the synchronous sliding window are determined. The physical field signal is sliced ​​using the synchronous sliding window to obtain multiple local transient signal slices; The time-domain and frequency-domain fusion processing of the physical field signal is performed on each of the local transient signal slices.

[0016] Optionally, after acquiring the physical field signal generated by the metering pump during operation, the method further includes: Noise is injected into the physical field signal using a phase modulation function based on the gear meshing frequency to generate a physical seed signal with real physical variance. The time-domain and frequency-domain fusion processing of the physical field signal is performed on the physical seed signal.

[0017] Optionally, the attention mechanism includes: Channel attention processing is performed on the input feature map to generate channel weights, which are then multiplied with the input feature map to obtain channel-weighted features; Spatial attention processing is applied to the channel weighted features to generate spatial weights, which are then multiplied with the channel weighted features to obtain the final weighted features. The input feature map is compressed using global average pooling and max pooling, and then nonlinearly mapped using a multilayer perceptron with shared weights to obtain the channel attention weights and the channel weight coefficient matrix. ,satisfy:

[0018] in, This represents the Sigmoid activation function. and For the shared weights of the multilayer perceptron, and These represent the channel feature descriptors after pooling.

[0019] Secondly, to efficiently execute the multi-fault diagnosis method for a lyocell fiber spinning metering pump provided by this invention, this invention also provides a multi-fault diagnosis system for a lyocell fiber spinning metering pump, comprising: an input device, an output device, a processor, and a memory, wherein the input device, output device, processor, and memory are interconnected, and the memory stores program instructions used for the multi-fault diagnosis method for the lyocell fiber spinning metering pump. The multi-fault diagnosis system for a lyocell fiber spinning metering pump of this invention has a compact structure and stable performance, and can stably execute the multi-fault diagnosis method for a lyocell fiber spinning metering pump provided by this invention, further enhancing the overall applicability and practical application capability of this invention.

[0020] This invention first constructs a computational fluid dynamics digital twin model of a metering pump, generating physical seed data through virtual fault injection. Then, it employs modulation noise enhancement based on meshing phase, combined with a synchronous slicing method based on gear rotation cycles, to construct a time-frequency fusion feature tensor. Next, it proposes a PI-WGAN-GP model embedding three physical priors: flow continuity, meshing frequency energy, and phase synchronization, achieving large-scale intelligent data expansion consistent with physical mechanisms. Then, it constructs a two-stage diagnostic model based on anisotropic attention and supervised contrastive learning, forcing similar fault features to cluster on a hypersphere. Finally, it accurately locates faulty gear teeth through spatial attention inverse mapping, achieving interpretable diagnosis and active closed-loop compensation. This invention solves the diagnostic challenges under conditions of few samples and high noise, significantly improving recognition accuracy and model robustness, and enabling root cause tracing and process self-optimization. Attached Figure Description

[0021] Figure 1 A flowchart of a multi-fault diagnosis method for a Lyocell fiber spinning metering pump provided in an embodiment of the present invention; Figure 2 This is a framework diagram of a multi-fault diagnosis system for a Lyocell fiber spinning metering pump provided in an embodiment of the present invention. Detailed Implementation

[0022] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0023] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0024] Please see Figure 1 This invention provides a method for diagnosing multiple faults in a lyocell fiber spinning metering pump, comprising the following steps: S1. Based on the structure and fluid parameters of the metering pump, construct a computational fluid dynamics digital twin model to generate simulated physical field signals under various fault conditions.

[0025] In this embodiment, a computational fluid dynamics digital twin model is established based on the actual structural parameters of the metering pump (such as module, number of teeth, tooth width, etc.). The gear meshing area and fluid channel are divided into fluid domains, and a non-Newtonian fluid model is used to describe the rheological properties of the Lyocell spinning solution.

[0026] Because the flow rate of the Lyocell spinning solution is low, the fluid state is laminar, the flow is transient, and it is an incompressible non-Newtonian fluid.

[0027] Based on this, the flow field governing equations are constructed, including: Construct the fluid continuity equation based on the law of conservation of mass:

[0028] in, For fluid velocity vector, For Hamiltonian operators; Construct the momentum conservation equation based on fluid velocity, density, and pressure:

[0029] Where t is time, The fluid density typically ranges from [value missing]. , For the transient pressure of the flow field, For viscous deviatoric stress tensor, It is the acceleration due to gravity; Because the Lyocell spinning solution flows incompressibly within the pump, its viscous deviatoric stress tensor The mathematical expression for it is defined as:

[0030]

[0031] in, Apparent viscosity; Describe the heat transfer and temperature change within the pump, using the energy conservation equation:

[0032] in, The value of the specific heat capacity at constant pressure is approximately , For temperature, Thermal conductivity, This is a viscous dissipation term. For the temperature diffusion term (Laplace operator); To accurately describe the high viscosity and shear-thinning characteristics of lyocell spinning solutions, a power-law model is used to characterize their apparent viscosity, based on the constitutive equations for non-Newtonian fluids.

[0033] in, This is the consistency coefficient, and its value typically ranges from [value range missing]. between, Shear rate, The rheological index (for shear-thinned fluids, ).

[0034] To ensure the tensor invariance of the constitutive equations in numerical calculations of multidimensional complex flow fields (such as gear meshing regions), the shear rate... The mathematical differential formula is derived from the second invariant of the flow field deformation rate tensor and is defined as follows:

[0035] in, and These are the fluid velocity vectors. Velocity components in different spatial dimensions of the Cartesian coordinate system and For the corresponding spatial coordinate components, This represents the double dot product operation of tensors. This calculation method rigorously integrates the velocity gradient changes of the fluid in all directions into a scalar shear rate, ensuring physical consistency and solution accuracy when dealing with strong shear flow in the radial clearance of a metering pump gear.

[0036] Furthermore, to ensure the accuracy of multiphysics transient numerical simulations and the realism of physical processes, a rigorous mathematical definition of the fluid boundary conditions of the computational domain was established for the geometric topology of the Lyocell spinning metering pump, and the overall boundary of the computational domain was defined. Specifically, it is divided into: entrance boundary Export border Gear motion boundary and the static wall boundary of the pump casing Specific flow constraints are applied to each boundary: Regarding the inlet boundary conditions, since the lyocell spinning solution is typically supplied by a booster device and the fluid is a viscous, incompressible fluid, the left suction end of the pump is set as a constant pressure inlet at the inlet boundary. At that point, the static pressure and velocity gradient of the fluid satisfy the following equation:

[0037] in, This represents the hydrostatic pressure at time t of a point with absolute spatial coordinates (x, y) in the computational domain. For a given inlet supply pressure (e.g., 20 bar), the equation constrains the flow field pressure at the inlet to be constant, and the fluid flows into the computational domain perpendicular to the inlet cross section.

[0038] For the outlet boundary conditions, the spinning solution enters the downstream spinning assembly after being squeezed by gears, resulting in a high system back pressure. The right-side discharge end of the pump is set as the pressure outlet boundary. At this location, the following constraints are satisfied:

[0039] in, For a given outlet back pressure (e.g., 35 bar), this pressure differential condition ( ( ) is the main power source for driving internal leaks in pumps (such as tooth tip clearance leaks and end face leaks).

[0040] For gear dynamic boundary conditions, the core of volumetric delivery in a metering pump lies in the periodic counter-rotation of the driving and driven gears. Due to the use of dynamic mesh technology, the tooth surfaces of the driving and driven gears are defined as non-slipping rigid body boundaries. Above, the local velocity of the fluid The linear velocity is exactly the same as that of the gear surface:

[0041] in, It is the rotational angular velocity vector of the gear (the driving gear and the driven gear are in opposite directions). This represents the spatial position vector of the monitoring point on the gear surface. The coordinates are the rotation center coordinates of the corresponding gear. This boundary condition drives the fluid carrying, volumetric compression, and jetting into the discharge chamber in real time.

[0042] For the static wall boundary condition of the pump casing, the non-moving contact surfaces such as the inner and outer casings of the metering pump are defined as static rigid walls. Due to the adhesive properties of high-viscosity non-Newtonian fluids, the no-slip condition is satisfied at the wall surface:

[0043] That is, the relative velocity of the fluid when it contacts the pump casing wall is zero, and the normal pressure gradient is zero, ensuring that the fluid can only generate strong shear leakage flow within the narrow radial gap formed by the tooth tip and the pump casing.

[0044] In the embodiments, macroscopic mechanical dynamics and multi-physics field cooperative control equations are also constructed, including transient volumetric displacement and internal leakage flow equations, gear meshing excitation frequency equations and pressure-flow pulsation phase synchronization equations.

[0045] For incompressible, high-viscosity, non-Newtonian fluids, the macroscopic theoretical flow rate of a gear pump driven by pressure difference is... It is formed by the coupling and superposition of ideal geometric displacement and leakage within physical clearances:

[0046] in, The preferred nominal displacement is the inherent geometric displacement per revolution of the metering pump. , For steady-state operating speed (preferably rated speed) ), The pressure difference between the inlet and outlet is 20 bar, and the back pressure at the outlet is 30 bar. The combined leakage coefficient, related to radial and axial clearances, is used when gear wear failure occurs. The increase in physical clearance is nonlinear, leading to an irreversible drop in flow rate, which depends on the physical assembly clearance between the gears in the radial and axial directions (often within...). (between), belonging to positive real empirical constants determined by geometric constraints, typically on the order of magnitude of... , For effective viscosity, the order of magnitude at the tooth tip clearance is typically in the range of... This equation defines the conservation boundary of physical flow.

[0047] The periodic high-frequency pulsations in the flow field at the metering pump outlet originate from the alternating meshing and squeezing of the driving and driven gears, causing the fundamental frequency (i.e., the theoretical meshing frequency) where the main energy of the flow field is concentrated. (Strictly defined by mechanical kinematics)

[0048] Regardless of the health or damage state of the flow field, the core energy of its pressure and flow signals must be anchored at the fundamental frequency and its higher harmonics. This equation constitutes the absolute physical benchmark for the frequency domain energy distribution of the flow field.

[0049] In a positive displacement gear pump, both the instantaneous pressure build-up and the transient flow discharge are driven uniformly by the rotor. Let the gear meshing excitation phase angle be... The phase of the main harmonic of the outlet pressure is The main harmonic phase of the flow is Due to the limitations of fluid microcompressibility and fixed channel geometry, there is a fixed intrinsic physical time difference (i.e., phase delay) between the pressure pulse spike and the flow drop. ):

[0050] in, The characteristic length of the flow channel from the engagement point to the outlet monitoring surface is typically determined by the pump body structure dimensions. , For the fluid sound velocity, at standard process temperature (e.g. Under these conditions, the velocity of sound in the Lyocell spinning solution is typically taken as a value of... , Substituting the above parameters, since... extremely short and Extremely large, with a theoretical time difference on the order of microseconds, which translates to an extremely small phase angle delay. In the WGAN-GP penalty constraint, its physical tolerance band upper and lower limits are very small. Usually set to This ensures that the generated "pseudo-pressure" and "pseudo-flow" are in a stable synchronous relationship within a finite phase tolerance range on the time axis. The equation shows that pressure and flow are not two isolated variables, and their transient changes must meet strict phase-locking tolerance zone limits on the time axis.

[0051] Furthermore, to ensure the numerical stability and physical accuracy of the digital twin model in simulating the transient discharge process of the metering pump, this invention constructs an adaptive dynamic mesh update and high-frequency solution mechanism for the high viscosity laminar flow characteristics of Lyocell spinning solution.

[0052] Specifically, to address the drastic compression and expansion of fluid volume in the gear meshing zone of the metering pump, a mesh update strategy coupled with spring smoothing and local mesh reconstruction is adopted. When the mesh between the tooth tip and the pump casing, and between tooth surfaces, becomes severely distorted due to strong shearing and compression, and the mesh orthogonality quality falls below a set safety threshold, a local mesh rewriting mechanism is automatically triggered. This mechanism effectively avoids solution divergence caused by negative volume, ensuring the stability of continuous operation simulation under multiple operating conditions.

[0053] Given the incompressibility and high viscosity of lyocell spinning solution, the flow field control equations are discretized using a fully implicit transient scheme, and the iterative solution of flow field momentum and mass is obtained using a pressure-based coupled algorithm to capture the minute high-frequency pressure pulsations generated at the outlet end by the periodic meshing of gears with high fidelity, thus preserving a complete physical impact fingerprint for subsequent time-frequency feature extraction.

[0054] To accurately capture the characteristic frequencies corresponding to local mechanical defects (such as pitting and broken teeth), the time step of the transient calculation is strictly constrained based on the rated speed of the gear and the minimum mesh size, ensuring that the displacement of the mesh in each time step does not exceed one-third of the local mesh feature size.

[0055] Meanwhile, the residual convergence criteria for both the continuity equation and the momentum equation are set to be less than [value missing]. This ensures that the transient flow state at each time step strictly adheres to the conservation of mass and momentum.

[0056] To address the difficulty in obtaining real-world fault data, and referencing physical failure evolution criteria, a geometric boundary parameterization reconstruction method is used in the digital twin model to target and inject four typical health / failure states, including: Healthy operating conditions: Standard involute tooth profile, no damage.

[0057] Wear failure: By reducing the tooth thickness, the radial clearance of the gear is increased, simulating the wear of the tooth tip and tooth surface caused by long-term operation.

[0058] Tooth breakage failure: The tooth height is reduced by cutting off the tooth tip to simulate the partial fracture of the tooth body caused by fatigue cracks.

[0059] Pitting fault: Local pit structures are set in the contact area of ​​the tooth surface to simulate local fatigue damage.

[0060] Using dynamic mesh technology in conjunction with boundary rotation speed data, transient pressure-velocity coupled calculations are performed with adaptive step size. The transient physical quantity, the outlet pressure signal, is extracted after the mesh has been verified to be independent. With export flow signal .

[0061] S2. Obtain the physical field signal generated by the metering pump during operation, and perform time-domain and frequency-domain fusion processing on the physical field signal and the simulated physical field signal respectively to construct their respective two-dimensional feature matrices.

[0062] To address the issue that strong background noise can easily mask subtle temporal fluctuations, and the need for high-quality data distribution in deep learning models, this invention proposes a feature space coupling strategy with dual-domain driving and synchronous truncation, and combines it with a physically guided generative adversarial network (PI-WGAN-GP) to complete the construction of the sample library.

[0063] First, modulation noise injection based on the meshing phase. To enable WGAN-GP to learn a data distribution with true operating condition variance, a phase modulation noise injection mechanism combining the gear meshing frequency is adopted to inject noise into the raw physical field signal, i.e., the pure pressure signal, generated by the metering pump during operation. With flow signal Introducing noise to construct an enhanced physical seed signal:

[0064]

[0065] in, The phase modulation envelope function, used to simulate the background flow field fluctuations caused by the periodic operation of the gear pump, is defined as follows: The noise model described above follows a normal distribution with zero mean:

[0066] In the formula, The total duration of the captured steady-state valid data. This represents the number of teeth on the gear (take 14). For rotational speed (take) ), To obey Uniformly distributed random initial phase, Based on Gaussian white noise, its variance Controlled by signal-to-noise ratio at different gradients:

[0067]

[0068] The effective power of the original signal; to provide WGAN-GP with an appropriate water level variance, the noise gradient level is set to... The corresponding current signal-to-noise ratio calculation formula is: .

[0069] To accurately capture the local features of the periodic meshing of gears, while avoiding unreliable features learned by the network due to window overlap, this invention proposes a synchronous sliding slicing method based on physical period.

[0070] Define window length With sliding step size :

[0071]

[0072] in, To capture the number of gear rotations covered by the window (take...) ), ensuring that it contains complete low-frequency rotational features; The number of single-tooth meshing cycles spanned by the sliding step (take) ).

[0073] Then, switch to the discrete sampling point domain, the number of data points in a single window. Step size , Sampling frequency (take The set of local transient signals after windowing is represented as follows:

[0074]

[0075] in, This represents the total number of discrete sampling points for the steady-state signal.

[0076] The windowed local transient slice signal sets of pressure and flow are respectively represented as follows:

[0077]

[0078] Furthermore, the discrete fast Fourier transform is used to map the one-dimensional windowed time-domain pressure signal to the frequency domain.

[0079] To eliminate the masking effect of steady-state DC bias on weak high-frequency fault characteristics, the pressure slices are first subjected to zero-mean processing:

[0080] in, This refers to the original time-domain pressure signal at the nth discrete sampling point within the i-th data window. This is the zero-mean pressure signal after removing the DC bias.

[0081] A fast Fourier transform is then performed, and the true amplitude spectrum is extracted (negative frequencies are discarded and symmetrical superposition is performed):

[0082] in, For discrete frequency indexing, For the first The feature vector of the true pressure amplitude spectrum corresponding to each window.

[0083] Due to the obtained discrete frequency domain pressure sequence With discrete time domain flow sequence Since the data lengths are inconsistent, linear interpolation is used to resample them to the same target length. , respectively denoted as and .

[0084] To accelerate the convergence of the bulldozer distance in the subsequent WGAN-GP, the interpolated sequence is standardized using Z-Score:

[0085]

[0086] in, and These are the mean and standard deviation of the corresponding interpolation sequence, respectively.

[0087] After standardization, a row-first folding mapping rule is used to reshape the one-dimensional sequence into a two-dimensional feature matrix with fixed column widths, and a column width adjustment hyperparameter is defined. Calculate the number of valid rows And truncate redundant data at the tail. For matrix space coordinates Elements at that location, pressure frequency domain matrix With the flow time domain matrix The reshaping mapping is defined as:

[0088]

[0089] In the formula, .

[0090] Finally, the reshaped pressure spectrum matrix will be... With the flow time domain matrix Vertical splicing is performed along the spatial channel dimension to construct a fused feature tensor. :

[0091] In the embodiment, a strict physical boundary isolation mechanism was established before generating adversarial extensions.

[0092] First, a stratified random sampling method is used to isolate and divide the extracted pure physical seed feature tensors according to a strict ratio (e.g., 10% independent test set and 20% validation set). The isolated set will never participate in the generation of adversarial data.

[0093] Subsequently, only the remaining training seed set is fed into the conditional WGAN-GP network for data augmentation. In the traditional conditional WGAN-GP architecture, the generator... Relying on the basic adversarial loss function The driving force, whose mathematical objective is to minimize the Wasserstein distance between the generated distribution and the true distribution, is to deceive the discriminator. The dimension of the latent noise space is defined as... (These are empirical hyperparameters preset based on the network's generation capabilities; in this embodiment, they are taken as...) ), generator Receive a potential noise vector randomly sampled from a prior standard multivariate normal distribution. The mathematical expression for its origin is:

[0094]

[0095] in, Represents a multivariate normal distribution. for An identity matrix of dimension 1 For operating condition labels, The authenticity score given by the discriminator.

[0096] However, traditional purely data-driven generative adversarial networks rely solely on the above... When faced with highly complex non-Newtonian fluid flow fields, the statistical distribution of fitted data is prone to generating seemingly realistic but actually erroneous features (such as extremely high pressure but no flow rate decay) that violate the principles of fluid mechanics and mechanical dynamics.

[0097] To this end, this invention constructs a Physical Information Guided Conditional Generative Adversarial Network (PI-WGAN-GP) to hard-encode the core physical mechanism of the metering pump into the generator. The loss function forms three major physical prior constraint penalty terms, including: flow continuity constraint, meshing frequency energy constraint, and phase consistency constraint.

[0098] Specifically, as a positive displacement pump, the outlet flow rate of the metering pump must conform to the physical gap leakage model under a specific pressure difference. Let the time-domain sequence of the flow rate output by the generator be... The average pressure derived from the pressure spectrum is Flow continuity loss is defined as the difference between the mean of the generated flow and the physical theoretical displacement. Mean square error between:

[0099] This constraint effectively prevents the network from generating anti-physical data that deviates from the law of conservation of volume and displacement.

[0100] The flow field pulsations in a gear pump are excited by the periodic meshing of the gears, thus generating a pressure spectrum tensor. Its energy must be mainly concentrated at the fundamental frequency (meshing frequency) At its higher harmonics, a physical mask vector is introduced. The forced generator must produce a peak response at the true physical characteristic frequency when generating the pressure spectrum, satisfying the following:

[0101] The instantaneous pressure surge and the instantaneous flow drop are driven by the same rotor dynamics equation and exhibit a strict phase-locked relationship. The phase delay between the pressure and flow time-domain signals generated by the cross-correlation function is constrained. This requires it to fall within the physically permissible tolerance zone. Inside:

[0102] in, It is a one-sided penalty activation function that applies a loss penalty only when the generated phase difference goes out of bounds.

[0103] In summary, by combining the Wasserstein distance and the gradient penalty term, the generator... Total optimization loss function ,satisfy:

[0104] in, In the initial optimization phase of the model, the adaptive penalty weights for physical constraints are typically set to an initial empirical range to prevent the gradient of the physical penalty term from completely drowning out the adversarial gradient of the generator. To achieve a smooth transition between data distribution exploration and physical rule exploitation in the network, dynamic annealing and adaptive weight adjustment strategies can be employed. As the number of training epochs increases, Based on the relative values ​​of each physical loss item The magnitude is adaptively amplified, thereby imposing a severe penalty on any anti-physical generation data in the later stages of training.

[0105] The aforementioned flow continuity constraints, meshing frequency energy constraints, and phase synchronization constraints are not empirical artificial penalty terms, but rather physical consistency constraints derived from the aforementioned multiphysics control equations, gear meshing dynamics, and fault cycle propagation mechanisms. Through this fusion paradigm of mechanism guidance and data exploration, the generator no longer blindly tries and fails in the vast parameter space, greatly accelerating the convergence speed of WGAN and significantly reducing the risk of mode collapse. Ultimately, it is integrated into a large-scale augmented training set with a highly balanced data distribution and strict adherence to physical mechanisms.

[0106] S3. Construct a feature extraction network with an attention mechanism and train the feature extraction network using a two-stage training paradigm. In the first stage, a supervised contrastive loss function is used to perform feature clustering learning on the training set composed of the two-dimensional feature matrix of the simulated physical field signal and / or the two-dimensional feature matrix of the physical field signal.

[0107] For the strong background flow field noise generated by the metering pump for lyocell fiber spinning when processing high-viscosity fluids, the CNN model for direct mapping classification is prone to feature aliasing at low signal-to-noise ratios, leading to metric space collapse.

[0108] To this end, an anisotropic deep convolutional neural network with a projection mechanism is constructed. The static topology includes the following core modules: anisotropic feature extraction and CBAM attention mechanism, channel attention module, spatial attention module, supervised contrastive learning metric and projection mechanism, and two-stage training and linear classification evaluation paradigm.

[0109] Because the time-frequency fusion feature tensor has drastically different physical meanings in the horizontal direction (time and frequency axes) and the vertical direction (slice axis), the encoder's first layer employs anisotropic receptive fields (such as horizontal). (Convolution kernel), specifically designed to capture weak phase pulsations within a single cycle and preserve long-range temporal correlations.

[0110] Subsequently, the intermediate feature maps are adaptively stripped of background noise through channel and spatial attention mechanisms. By automatically evaluating the weights of high-value information in the dual physical channels of "pressure spectrum" and "flow time domain", and compressing the spatial dimension using global average pooling and max pooling, a nonlinear mapping is performed through a multilayer perceptron (MLP) with shared weights. The channel weight coefficient matrix is ​​calculated as follows:

[0111] in, This represents the Sigmoid activation function. and For the shared weights of the multilayer perceptron, and These represent the channel feature descriptors after pooling.

[0112] After obtaining the channel weights, use the element-wise multiplication operation ( ), calculate the channel-weighted features The aim is to accurately locate key grid point regions in the time-frequency tensor (such as second harmonic spikes caused by pitting and local time-domain notches caused by tooth breakage). This involves weighting the features after channel processing. Max pooling and average pooling are performed along the channel axis, and the resulting images are concatenated before large-size convolution operations are performed to extract the spatial response map.

[0113] in, This represents the feature map after being stitched together along the channel axis. Indicates the sensory field as The convolution operation ultimately results in the weighted features being... .

[0114] After the above multi-layer attention extraction, the feature tensor output by the network terminal has a dimension of . (The number of backend channels is set to) ).

[0115] Furthermore, to address the issue of different sliding step sizes in the preceding sequence... The resulting abrupt change in input dimension scaling is addressed by introducing adaptive average pooling to force compression of the spatial dimension. Then, a flattening operation is performed to output a 512-dimensional high-order feature representation vector with a high signal-to-noise ratio. Its spatial dimension reduction mapping is defined as:

[0116] To prevent the network from overfitting to spurious features with a noisy background, a projection head module is introduced to project the 512-dimensional feature vector. Mapping to a lower-dimensional (128-dimensional) metric clustering space .

[0117] Within this space, a supervised contrastive loss function (SupConLoss) is used instead of cross-entropy loss for feature representation learning. Its core physical mechanism is: constructing a homogeneous mask using real physical labels. On the normalized hypersphere, samples with the same fault conditions are forced to be pulled closer together (attracted), while samples with different conditions are pushed apart (repelled).

[0118] For any anchor point sample within a batch Its loss function is mathematically expressed as:

[0119] in, This represents the supervised contrastive loss function. This is the set of indices for all samples within the current training micro-batch. For batch size, For anchor point A set of positive samples with the same working condition labels. The two-dimensional feature matrices represent the two-dimensional feature matrices of the anchor point samples respectively. Positive samples and arbitrary comparison sample two-dimensional feature matrix Embedding features of a two-dimensional feature matrix in the projection space, and satisfying of Normalization conditions Temperature coefficient (preferred) This is used to amplify small similarity differences between samples and control the sensitivity to penalties for difficult negative samples.

[0120] Based on the above architecture, a two-stage decoupled training paradigm of feature clustering-linear evaluation is designed to effectively block noise coupling interference in gradient backpropagation: Phase 1: Supervised contrastive representation learning based on physically augmented data.

[0121] To prevent deep convolutional networks from relearning spurious features that violate physical laws under complex background flow field noise, this stage directly relies on the augmented dataset generated by the aforementioned PI-WGAN-GP, which has been rigorously injected with multi-physical mechanism constraints, for training.

[0122] At this stage, the network's supervised contrast loss, as defined in the previous section, is used. The sole optimization objective is to ensure that the dataset fed to the network internalizes physical laws such as flow continuity, meshing frequency energy, and phase synchronization. By forcing the network to perform comparative clustering in the data distribution with this real physical variance, the deep network is always subject to data-driven constraints based on real mechanical dynamics during the feature representation learning process.

[0123] Phase Two: Linear Classification Assessment.

[0124] By implementing a tail-cutting survival strategy, the weight parameters of the feature encoder and the projection head are frozen, the projection head is truncated, and only the thoroughly optimized 512-dimensional feature interface is exposed. A single-layer linear fully connected layer is connected to the network, and cross-entropy loss is used for fine-tuning training. Since the features have achieved high separability in the multi-dimensional space in the early stage, the linear classifier can converge with very few training rounds, achieving extremely high fault diagnosis accuracy on the blind test set.

[0125] Cross-entropy loss Defined as:

[0126] in, This represents the number of samples in the micro-batch (e.g., 32). This represents the total number of fault categories (with a value of 4). As an indicator variable, if the sample The true category is Then take 1, otherwise take 0. The sample is the output of a fully connected classifier and after being activated by Softmax. Category The predicted probability.

[0127] S4. Using the two-dimensional feature matrix to be diagnosed, obtain the fault diagnosis result through the trained feature extraction network, and locate the physical component where the fault occurs based on the attention mechanism.

[0128] After training, the model weights that achieve the best accuracy on the validation set are loaded, and blind testing is performed on a completely unseen, independent test set. Given that this is geared towards... For multi-class fault diagnosis tasks, based on the true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN) in the confusion matrix, a macro-averaging multi-dimensional index is introduced to rigorously quantify the diagnostic performance:

[0129]

[0130]

[0131]

[0132] Among them, the F1 score, as the harmonic mean of precision and recall, is used as the core criterion for measuring the model's comprehensive recognition performance in complex working conditions and strong interference backgrounds, overcoming class misclassification and missed classification.

[0133] The trained lightweight Phase 2 model (with the projection head removed and only the CBAM encoder and classifier retained) is deployed in the industrial edge computing node of Lyocell metering pump to achieve closed-loop control from physical perception to intelligent decision-making: online real-time inference of high-frequency signals collected by edge nodes from sensors, input into the model after local transient slicing and time-frequency feature folding, and output the current device's health probability state distribution in milliseconds.

[0134] When "broken tooth" or "wear" is detected, the two-dimensional weight matrix output by the spatial attention module is automatically extracted. Since this matrix has a strict spatial mapping relationship with the original time-frequency tensor, the response peak coordinates of the weight matrix can be searched. By reverse calculation, the corresponding time-domain window time can be determined. .

[0135]

[0136] Furthermore, combining the current pump shaft rotation absolute phase angle equation... With the total number of teeth of the gear Calculate the specific tooth number where the physical defect occurred. :

[0137] In the formula, The start time of the current data window, The pump shaft angular velocity, As an up-rounding function, the above-mentioned physical reverse mapping mechanism can accurately trace and lock the specific gear tooth number where the physical defect occurred, realizing a leap from "black box diagnosis" to "deep process traceability".

[0138] Furthermore, the predicted probability can be used as an evolutionary indicator of equipment health and fed back to the Manufacturing Execution System (MES) via industrial Ethernet. When the "wear" probability exceeds a preset safety threshold, the MES automatically issues instructions to dynamically fine-tune the pumping frequency of the main motor inverter to compensate for internal leakage losses caused by increased tooth tip clearance due to wear; in the event of a sudden and severe "tooth breakage," an emergency shutdown work order is generated. This ensures extremely uniform lyocell fiber output from the source, avoiding large-scale waste caused by uneven fineness or tooth breakage.

[0139] It should be noted that the specific implementation methods described above, such as image processing, numerical simulation, and the construction and training of machine learning models, can all be accomplished by the processor by calling the corresponding computer program instructions stored in memory. Those skilled in the art can implement the above functions using algorithms and tools known in the prior art, according to actual needs.

[0140] Please see Figure 2 In an embodiment, to efficiently execute the multi-fault diagnosis method for a lyocell fiber spinning metering pump provided by the present invention, the present invention also provides a multi-fault diagnosis system for a lyocell fiber spinning metering pump, comprising: an input device 1, an output device 2, a processor 3, and a memory 4, wherein the input device 1, output device 2, processor 3, and memory 4 are interconnected, and the memory 4 stores program instructions for executing the steps of the multi-fault diagnosis method for the lyocell fiber spinning metering pump. The multi-fault diagnosis system for a lyocell fiber spinning metering pump of the present invention has a compact structure and stable performance, and can stably execute the multi-fault diagnosis method for a lyocell fiber spinning metering pump of the present invention, further improving the overall applicability and practical application capability of the present invention.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the present invention.

Claims

1. A method for diagnosing multiple faults in a metering pump for lyocell fiber spinning, characterized in that, Includes the following steps: Based on the structure and fluid parameters of the metering pump, a computational fluid dynamics digital twin model is constructed to generate simulated physical field signals under various fault conditions. The physical field signal generated by the metering pump during operation is acquired, and the physical field signal and the simulated physical field signal are fused in the time domain and frequency domain respectively to construct their respective two-dimensional feature matrices. A feature extraction network with an attention mechanism is constructed, and the feature extraction network is trained using a two-stage training paradigm. In the first stage, a supervised contrastive loss function is used to perform feature clustering learning on the training set composed of the two-dimensional feature matrix of the simulated physical field signal and / or the two-dimensional feature matrix of the physical field signal. The fault diagnosis result is obtained by using the two-dimensional feature matrix to be diagnosed through a trained feature extraction network, and the physical component where the fault occurs is located based on the attention mechanism.

2. The method for diagnosing multiple faults in a lyocell fiber spinning metering pump according to claim 1, characterized in that, The computational fluid dynamics digital twin model is constructed to generate simulated physical field signals under various fault conditions, including: Construct the computational fluid dynamics digital twin model, and by adjusting the gear geometric boundary parameters in the digital twin model, inject and generate simulated physical field signals that include at least one of the faults: wear, tooth breakage, and pitting. The process involves adjusting the gear geometric boundary parameters in the digital twin model to inject and generate simulated physical field signals that include at least one of the following faults: wear, tooth breakage, and pitting. Wear failure was simulated by increasing radial clearance by reducing tooth thickness. A broken tooth failure is simulated by cutting off the tooth tip to reduce the tooth height. Pitting corrosion is simulated by creating localized pits in the tooth surface contact area.

3. The method for diagnosing multiple faults in a lyocell fiber spinning metering pump according to claim 1, characterized in that, The physical field signal is fused in both the time and frequency domains to construct the two-dimensional feature matrix, including: Extracting the frequency domain feature sequence of the pressure signal and the time domain feature sequence of the flow signal from the physical field signal; After standardizing the frequency domain feature sequence and the time domain feature sequence, they are folded and recombined into a two-dimensional feature matrix.

4. The method for diagnosing multiple faults in a Lyocell fiber spinning metering pump according to claim 1, characterized in that, After constructing their respective two-dimensional feature matrices, and before inputting them into the feature extraction network, the process further includes: A generative adversarial network guided by physical information is used to augment the data of the two-dimensional feature matrix; The generator loss function of the generative adversarial network incorporates a physical consistency penalty term to constrain the generated data to conform to the laws of fluid dynamics. The physical consistency penalty includes at least one of the following: Flow continuity constraints are used to penalize the error between the flow rate in the generated data and the theoretical displacement. Meshing frequency energy constraint is used to force the energy of the pressure frequency domain in the generated data to be concentrated at the meshing frequency and higher harmonics. Phase consistency constraints are used to penalize the phase difference between pressure and flow signals in the generated data that exceeds a preset physical tolerance.

5. The method for diagnosing multiple faults in a lyocell fiber spinning metering pump according to claim 1, characterized in that, The feature extraction network with attention mechanism includes: The first layer uses an anisotropic convolution kernel to differentially extract the features of the two-dimensional feature matrix in the time and frequency dimensions; Following the first layer, a dual mechanism of channel attention and spatial attention is integrated to adaptively weight the features of different channels and spatial regions; The two-stage training paradigm includes: In the first stage, based on the training set, the structure other than the projection head connected after the feature extraction network is frozen, and the feature extraction network and the projection head are trained using a supervised contrastive loss function to achieve clustering of similar features in a high-dimensional space. In the second stage, the projection head is removed and the feature extraction network is frozen. A linear classifier is then connected, and the linear classifier is trained using the cross-entropy loss function.

6. The method for diagnosing multiple faults in a lyocell fiber spinning metering pump according to claim 5, characterized in that, The supervised contrastive loss function used in the first stage satisfies: in, This represents the supervised contrastive loss function. This is the set of indices for all samples within the current training micro-batch. For batch size, For anchor point A set of positive samples with the same working condition labels. The two-dimensional feature matrices represent the two-dimensional feature matrices of the anchor point samples respectively. Positive samples and arbitrary comparison sample two-dimensional feature matrix Embedding features of a two-dimensional feature matrix in the projection space, and satisfying of Normalization conditions It is a temperature coefficient used to amplify small similarity differences between samples and control the sensitivity to penalties for difficult negative samples.

7. The method for diagnosing multiple faults in a Lyocell fiber spinning metering pump according to claim 1, characterized in that, Before performing the time-domain and frequency-domain fusion processing, the following steps are also included: Based on the rotation cycle of the metering pump, the window length and sliding step size of the synchronous sliding window are determined. The physical field signal is sliced ​​using the synchronous sliding window to obtain multiple local transient signal slices; The time-domain and frequency-domain fusion processing of the physical field signal is performed on each of the local transient signal slices.

8. The method for diagnosing multiple faults in a lyocell fiber spinning metering pump according to claim 1, characterized in that, After acquiring the physical field signals generated by the metering pump during operation, the process also includes: Noise is injected into the physical field signal using a phase modulation function based on the gear meshing frequency to generate a physical seed signal with real physical variance. The time-domain and frequency-domain fusion processing of the physical field signal is performed on the physical seed signal.

9. The method for diagnosing multiple faults in a Lyocell fiber spinning metering pump according to claim 1, characterized in that, The attention mechanism includes: Channel attention processing is performed on the input feature map to generate channel weights, which are then multiplied with the input feature map to obtain channel-weighted features; Spatial attention processing is applied to the channel weighted features to generate spatial weights, which are then multiplied with the channel weighted features to obtain the final weighted features. The input feature map is compressed using global average pooling and max pooling, and then nonlinearly mapped using a multilayer perceptron with shared weights to obtain the channel attention weights and the channel weight coefficient matrix. ,satisfy: in, This represents the Sigmoid activation function. and For the shared weights of the multilayer perceptron, and These represent the channel feature descriptors after pooling.

10. A multi-fault diagnosis system for a lyocell fiber spinning metering pump, characterized in that, The lyocell fiber spinning metering pump multi-fault diagnosis system includes: an input device, an output device, a processor, and a memory. The input device, output device, processor, and memory are interconnected. The memory stores program instructions, which are used to execute the lyocell fiber spinning metering pump multi-fault diagnosis method according to any one of claims 1-9.

Images