End-to-end mechanical signal noise reduction methods, systems, equipment, and storage media

By employing an end-to-end mechanical signal denoising method, an undirected graph structure is constructed using a sliding window and the K-nearest neighbor algorithm. Combined with the SCAM ternary causal independence axiom and the differentiable causal discovery algorithm, causal decoupling of effective signal, noise, and operating condition is achieved. This solves the problem of difficulty in separating noise from effective information in existing technologies and improves the stability and accuracy of signal separation.

CN123115598APending Publication Date: 2026-11-17WUHAN UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611280457.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-22
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies cannot effectively separate noise from the original valid information in industrial equipment observation signals, resulting in distorted status determination and operation and maintenance assessment results.

Method used

An end-to-end mechanical signal denoising method is adopted. A standardized time-series sample set is constructed by sliding window segmentation and amplitude normalization. An undirected graph structure is constructed by combining the K-nearest neighbor algorithm. The ternary causal denoising model is trained by the SCAM ternary causal independence axiom and the differentiable causal discovery algorithm to achieve causal decoupling of effective signal, noise and operating condition.

Benefits of technology

It achieves high-precision signal separation under complex operating conditions, avoids the erroneous filtering of components under normal operating conditions, improves the stability and accuracy of signal separation, and ensures that the time and frequency domain characteristics of the effective signal are not distorted.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115598A_ABST
    Figure CN123115598A_ABST
Patent Text Reader

Abstract

This application proposes an end-to-end mechanical signal denoising method, system, device, and storage medium, relating to the field of signal preprocessing technology. The method includes: acquiring the original noisy observation signals and noise-free calibration signals of the monitored equipment under various operating conditions, and preprocessing them to obtain effective noisy time-domain signals; segmenting the effective noisy time-domain signals and noise-free calibration signals using a sliding window, standardizing the amplitude of the effective noisy time-domain sub-signals, matching the corresponding noise-free calibration sub-signals as denoising standards, and constructing a standardized time-series sample set; converting the standardized time-series sample set into node feature representations, and combining them with the K-nearest neighbor algorithm to obtain graph-structured data; training a ternary causal denoising model using the graph-structured data and a target loss function to obtain a target denoising model; and inputting the signal to be denoised into the target denoising model to obtain the corresponding denoised signal. This application achieves high-precision denoising of multi-point observation signals from complex equipment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of signal preprocessing technology, and in particular to an end-to-end mechanical signal noise reduction method, system, device and storage medium. Background Technology

[0002] In modern industry, the operational data of critical core equipment such as wind power equipment and aero engines are affected by the complex environment of their own operating load and speed. The collected observation signals are often coupled with a large amount of environmental noise, mechanical disturbances, electromagnetic interference, and random impact noise. If these signals are used directly for subsequent analysis without effective processing, the results of condition determination and operation and maintenance assessment will be severely distorted. Therefore, signal preprocessing and noise reduction is a core requirement in the industrial data preprocessing stage, and the fidelity of noise reduction directly determines the data quality and analytical effectiveness of all subsequent downstream tasks.

[0003] However, existing noise reduction methods still have many core flaws and cannot achieve accurate separation of noise from the original valid information at the root. Summary of the Invention

[0004] In view of this, this application proposes an end-to-end mechanical signal noise reduction method, system, device and storage medium.

[0005] In a first aspect, this application provides an end-to-end mechanical signal noise reduction method, comprising: The original noisy observation signal and noiseless calibration signal of the device under various operating conditions are acquired, and the original noisy observation signal is preprocessed to obtain an effective noisy time domain signal. The effective noisy time-domain signal and the noiseless calibration signal are segmented using a sliding window, and the amplitude of the segmented effective noisy time-domain sub-signals is normalized. The corresponding noiseless calibration sub-signals are then matched as noise reduction standards to construct a standardized time-series sample set. Using single-measurement-point time-series subsequences and single-measurement-point time-series segments in multi-measurement-point signals as nodes, the standardized time-series sample set is transformed into node feature representations, and a standardized undirected graph structure is constructed by combining the K-nearest neighbor algorithm to obtain graph structure data. The ternary causal denoising model is trained using the graph structure data and the target loss function to obtain the target denoising model. The ternary causal denoising model defines the rigid causal boundary between the effective signal, noise, and operating condition through the SCAM ternary causal independence axiom. The ternary causal independence axiom is transformed into a differentiable loss function. Based on the input graph structure data, a differentiable causal discovery algorithm is used to learn the causal adjacency matrix. The causal adjacency matrix is ​​used to modulate the attention weights of the three-branch graph attention network to drive the decoupling of the effective signal, noise, and operating condition at the causal level. The signal to be denoised is input into the target denoising model to obtain the corresponding denoised signal.

[0006] In one embodiment, the preprocessing of the original noisy observation signal to obtain an effective noisy time-domain signal includes: The invalid signals during the equipment start-up and shutdown phases in the original noisy observation signal are removed, and abnormal impact signals exceeding the measurement range are also removed to obtain the effective noisy time-domain signal.

[0007] In one embodiment, the step of segmenting the effective noisy time-domain signal and the noise-free calibration signal using a sliding window includes: The Hanning sliding window function is used to adaptively determine the window width based on the sampling frequency and the device switching frequency; Based on the window width, the same sliding window parameters are used to perform sliding window segmentation on the effective noisy time-domain signal and the noiseless calibration signal to obtain corresponding sub-signal segments.

[0008] In one embodiment, the step of using single-measurement-point time-series subsequences and single-measurement-point time-series segments in multi-measurement-point signals as nodes, transforming the standardized time-series sample set into node feature representations, and constructing a standardized undirected graph structure using the K-nearest neighbor algorithm includes: Using the single-measurement point time sequence subsequence corresponding to the sliding window and the single-measurement point time segment under the multi-measurement point signal as nodes of the graph, the time domain and frequency domain statistical features of each node are extracted to generate high-dimensional feature vectors of the nodes and form a node feature matrix. The value of K is determined based on the total number of nodes. The K-nearest neighbor algorithm is used to search the neighborhood of the nodes and complete the edge connection to generate the original adjacency matrix. The node feature matrix is ​​normalized, and the original adjacency matrix is ​​symmetrically normalized. The normalized node feature matrix and the normalized adjacency matrix are combined to form the standardized undirected graph structure.

[0009] In one embodiment, the ternary causal noise reduction model includes: an SCAM basic causal generation layer, a three-branch decoupling layer, and a causal fidelity reconstruction layer; The SCAM basic causal generation layer is used to perform SCAM causal formal reconstruction of the input graph structure data in terms of effective signal, noise, and operating condition. The rigid causal boundary between the effective signal, noise, and operating condition is defined by the ternary causal independence axiom. The three-branch decoupling layer includes a differentiable causal discovery submodule and a three-branch graph attention submodule. The differentiable causal discovery submodule is used to learn based on the hidden layer features of the nodes output by the SCAM basic causal generation layer to obtain a backpropagable causal adjacency matrix. The three-branch graph attention submodule uses the causal adjacency matrix to modulate the three-branch attention weights, extract the effective signal causal features, noise interference features, and operating condition features respectively, and drives the causal level decoupling of the three through causal independence loss. The causal fidelity reconstruction layer is used to reconstruct the signal based on the causal characteristics of the decoupled effective signal, apply multiple fidelity constraints in the time domain, frequency domain, and causal structure, and output the reconstructed effective signal.

[0010] In one embodiment, the causal fidelity reconstruction layer includes multiple stacked graph convolutional decoding layers, as well as temporal fidelity constraint branches, frequency-domain fidelity constraint branches, and causal structure fidelity constraint branches. Multiple graph convolutional decoding layers are used to sequentially perform graph domain signal reconstruction based on the causal features of the input valid signal, and output the reconstructed valid signal; The time-domain fidelity constraint branch, frequency-domain fidelity constraint branch, and causal structure fidelity constraint branch are used to apply time-domain, frequency-domain, and causal structure fidelity constraints, respectively, during model training.

[0011] In one embodiment, the target loss function includes triple-fidelity reconstruction loss, ternary causal decoupling loss, invariance regularization term, and acyclic constraint term; The triple-fidelity reconstruction loss includes time-domain fidelity loss, frequency-domain fidelity loss, and causal structure fidelity loss, which are used to apply time-domain, frequency-domain, and causal structure constraints to the reconstructed signal. The ternary causal decoupling loss is used to constrain the causal characteristics of the effective signal, the noise interference characteristics, and the operating condition characteristics to be independent of each other. The invariance regularization term is used to constrain the operating condition characteristics to have environmental invariance. The acyclic constraint term is used to constrain the causal adjacency matrix to satisfy the directed acyclic requirement.

[0012] Secondly, this application also provides an end-to-end mechanical signal noise reduction system, comprising: The acquisition module is used to acquire the original noisy observation signal and the noiseless calibration signal of the device under various operating conditions, and to preprocess the original noisy observation signal to obtain the effective noisy time domain signal. The sample construction module is used to segment the effective noisy time-domain signal and the noiseless calibration signal using a sliding window, and to perform amplitude normalization on the segmented effective noisy time-domain sub-signals, and match the corresponding noiseless calibration sub-signals as noise reduction standards to construct a standardized time-series sample set. The conversion module is used to convert the standardized time series sample set into node feature representations using single-measurement-point time series subsequences and single-measurement-point time series segments in multi-measurement-point signals as nodes, and to construct a standardized undirected graph structure by combining the K-nearest neighbor algorithm to obtain graph structure data. The training module is used to train the ternary causal denoising model using the graph structure data and the target loss function to obtain the target denoising model. The ternary causal denoising model defines the rigid causal boundary between the effective signal, noise, and operating condition through the SCAM ternary causal independence axiom, transforms the ternary causal independence axiom into a differentiable loss function, learns the causal adjacency matrix based on the input graph structure data using a differentiable causal discovery algorithm, and uses the causal adjacency matrix to modulate the attention weights of the three-branch graph attention network to drive the decoupling of the effective signal, noise, and operating condition at the causal level. The noise reduction module is used to input the signal to be denoised into the target noise reduction model to obtain the corresponding noise reduction signal.

[0013] Thirdly, this application also provides an electronic device, including a processor and a memory; the memory has a computer program stored thereon, wherein the computer program, when executed by the processor, implements the end-to-end mechanical signal noise reduction method as described in the first aspect.

[0014] Fourthly, this application also provides a non-transitory computer storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the end-to-end mechanical signal noise reduction method as described in the first aspect.

[0015] The end-to-end mechanical signal noise reduction method of this application has the following advantages over related technologies: 1. The end-to-end mechanical signal denoising method of this application preprocesses the original observed signal, performs sliding window segmentation and amplitude standardization, and matches the corresponding noiseless calibration sub-signal as the denoising standard to construct a standardized time series sample set. Based on this, the standardized time series sample set is transformed into node feature representations, and a standardized undirected graph structure is constructed using the K-nearest neighbor algorithm to obtain graph structure data. This allows one-dimensional time series signals to be converted into graph domain data that can characterize the correlation between measurement points, overcoming the limitation of traditional time series denoising that only utilizes single-point time domain information and improving the ability to mine signal correlation features.

[0016] 2. A ternary causal denoising model based on the SCAM ternary causal independence axiom is adopted, and semi-supervised training is implemented with a target loss function. The model learns the causal adjacency matrix using a differentiable causal discovery algorithm, and uses this matrix to modulate the attention weights of the three-branch graph, achieving causal decoupling of effective signal, noise, and operating condition. Compared with conventional graph denoising models that rely solely on data distribution to drive feature separation, this scheme introduces rigid causal boundary constraints to avoid feature confusion caused by data overfitting. It can distinguish between operating condition changes and real noise interference, avoids the erroneous filtering of normal operating condition components, and improves the stability of signal separation under complex operating conditions. After training to obtain the target denoising model, the denoised signal can be output directly by inputting the signal to be denoised. The entire process realizes an end-to-end closed loop for time-series signals, from sample structure construction and causal decoupling model training to inference denoising. While preserving the inherent characteristics of the effective signal in the time and frequency domains, it suppresses noise, ultimately achieving high-precision denoising of multi-point observation signals from complex equipment. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an end-to-end mechanical signal noise reduction method according to an embodiment of this application; Figure 2 This is a schematic diagram of the structure of the ternary causal noise reduction model and the causal fidelity reconstruction layer in one embodiment of this application; Figure 3 This is a schematic diagram of the network architecture of a ternary causal noise reduction model in one embodiment of this application; Figure 4 This is a schematic diagram of the structure of the three-branch decoupling layer in one embodiment of this application; Figure 5 This is a comparative schematic diagram of the planetary gearbox before and after normal signal noise reduction in one embodiment of this application; Figure 6 This is a schematic diagram comparing the wear signal of the planetary gearbox before and after noise reduction in one embodiment of this application; Figure 7 This is a schematic diagram showing the comparison of the planetary gearbox tooth root fracture signal before and after noise reduction in one embodiment of this application; Figure 8 This is a schematic diagram of the structure of an end-to-end mechanical signal noise reduction system in one embodiment of this application. Detailed Implementation

[0019] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0020] In some embodiments, such as Figure 1 As shown, this application provides an end-to-end mechanical signal noise reduction method, which includes the following steps S101 to S105.

[0021] S101: Acquire the original noisy observation signal and noiseless calibration signal of the device under various operating conditions, and preprocess the original noisy observation signal to obtain the effective noisy time domain signal.

[0022] It is understandable that collecting noisy observation data and noiseless benchmark calibration data for equipment under varying operating conditions establishes a pair of benchmark samples for subsequent model learning of noise distribution and effective signal characteristics. At the same time, preprocessing removes invalid abnormal components introduced by equipment failure, instantaneous pulse interference, etc. from the original observation signal, retains the time-domain waveform components that can reflect the actual operating state of the equipment, and outputs a standardized effective noisy time-domain signal.

[0023] S102: A sliding window is used to segment the effective noisy time-domain signal and the noiseless calibration signal respectively. The amplitude of the segmented effective noisy time-domain sub-signals is standardized, and the corresponding noiseless calibration sub-signals are matched as noise reduction standards to construct a standardized time-series sample set.

[0024] It is understandable that using a sliding window to segment long time-series signals yields time-series sub-signals of uniform length, adapting to the fixed-length input requirements of subsequent models. Amplitude normalization eliminates the inconsistency in amplitude scale caused by differences in signal acquisition gain. Then, a one-to-one matching relationship is established between noisy sub-signals and noiseless calibration sub-signals in the same time period, thereby forming a standardized and aligned set of time-series samples. This reduces the interference of amplitude deviation on model feature learning and provides paired, aligned, and scaled basic samples for subsequent model training.

[0025] S103: Using single-measurement-point time series subsequences and single-measurement-point time series segments in multi-measurement-point signals as nodes, the standardized time series sample set is transformed into node feature representations, and a standardized undirected graph structure is constructed by combining the K-nearest neighbor algorithm to obtain graph structure data.

[0026] In current applications, time-series analysis using related technologies focuses only on the time-series evolution of a single signal, resulting in significant storage limitations. This application, however, maps each independent single-measurement-point time-series segment to feature nodes in a graph structure. It transforms the one-dimensional time-series information in a standardized time-series sample set into high-dimensional node feature information, fully preserving the time-domain characteristics and operational status information of each measurement point signal. Simultaneously, it uses the K-nearest neighbor algorithm to mine potential correlations and coupling relationships between different measurement point nodes, adaptively establishing topological connections between nodes. This constructs a standardized undirected graph structure that aligns with the operational correlation characteristics of multiple measurement points in the equipment, achieving an effective conversion from time-series data to graph structure data. This significantly improves the model's ability to model coupled signals from multiple measurement points in the equipment.

[0027] S104: The ternary causal denoising model is trained using graph-structured data and a target loss function to obtain the target denoising model. Specifically, the ternary causal denoising model defines the rigid causal boundary between the effective signal, noise, and operating condition using the SCAM ternary causal independence axiom. This axiom is then transformed into a differentiable loss function. Based on the input graph-structured data, a differentiable causal discovery algorithm is used to learn the causal adjacency matrix. This causal adjacency matrix is ​​then used to modulate the attention weights of the three-branch graph attention network, driving the decoupling of the effective signal, noise, and operating condition at the causal level.

[0028] Among them, the ternary causal denoising model can be constructed based on the Structural Causal Additive Model (SCAM) and Graph Convolutional Neural Network (GCN).

[0029] It is understandable that this embodiment uses standardized graph structure data for supervised model training, which differs from traditional pure data-driven denoising models. By introducing the SCAM ternary causal independence axiom, it defines the rigid causal boundary between effective signal, noise, and operating condition, constraining the model's learning logic from a physical causal perspective. This overcomes the technical bottleneck of traditional models, which struggle to distinguish between operating condition fluctuations and real noise interference. Simultaneously, the static causal axiom is transformed into a differentiable loss function, adapted to the deep learning backpropagation iterative optimization mechanism. A differentiable causal discovery algorithm adaptively mines the real causal relationships between graph nodes and generates a backpropagable causal adjacency matrix. Furthermore, this causal topological information is used to dynamically modulate the weight distribution of the three-branch graph attention network, achieving targeted extraction of three types of features and precise decoupling at the causal level. Combined with a multi-constraint objective loss function, iterative optimization of model parameters is completed, ultimately training a high-precision objective denoising model with causal cognition capabilities, strong anti-interference ability, and adaptability to complex and variable operating conditions.

[0030] S105: Input the signal to be denoised into the target denoising model to obtain the corresponding denoised signal.

[0031] It is understandable that when the signal to be denoised is input into the target denoising model, the model will automatically distinguish the effective signal component, noise component and operating condition disturbance component in the signal to be denoised. Thus, the effective signal component can be separated while ensuring that the effective signal is not distorted in the time domain, frequency domain and inherent causal topology. Based on the effective signal component, the restored denoised signal is output to obtain the denoised signal corresponding to the signal to be denoised.

[0032] The aforementioned end-to-end mechanical signal denoising method preprocesses the original observed signal, performs sliding window segmentation and amplitude standardization, and matches the corresponding noiseless calibration sub-signal as the denoising standard to construct a standardized time-series sample set. Based on this, the standardized time-series sample set is transformed into node feature representations, and a standardized undirected graph structure is constructed using the K-nearest neighbor algorithm to obtain graph-structured data. This allows one-dimensional time-series signals to be converted into graph-domain data that can characterize the correlation between measurement points, overcoming the limitation of traditional time-series denoising that only utilizes single-point time-domain information and improving the ability to mine signal correlation features. Furthermore, a ternary causal denoising model based on the SCAM ternary causal independence axiom is adopted, and semi-supervised training is implemented using a target loss function. The model learns the causal adjacency matrix using a differentiable causal discovery algorithm and uses this matrix to modulate the attention weights of the three-branch graph, achieving effective causal decoupling of signal, noise, and operating condition. Compared to conventional graph denoising models that rely solely on data distribution to drive feature separation, this approach introduces rigid causal boundary constraints to avoid feature confusion caused by data overfitting. This allows for the differentiation between changing operating conditions and real noise interference, preventing the erroneous filtering of normal operating condition components and improving the stability of signal separation under complex conditions. After training the target denoising model, the denoised signal can be output directly by inputting the signal to be denoised. The entire process achieves an end-to-end closed loop for time-series signals, from sample structure construction and causal decoupling model training to inference denoising. It suppresses noise while preserving the inherent time and frequency domain characteristics of the effective signal, ultimately achieving high-precision denoising of multi-point observation signals from complex equipment.

[0033] In some embodiments, preprocessing the original noisy observation signal to obtain an effective noisy time-domain signal includes: removing invalid signals during the device start-up and shutdown phases from the original noisy observation signal and removing abnormal impact signals that exceed the measurement range to obtain an effective noisy time-domain signal.

[0034] In applications, to meet the preprocessing needs of rotating machinery operation signals in large industrial settings, a signal acquisition system can be built targeting the core monitoring parts of the equipment under test. This system can include at least one of the following: vibration acceleration sensors, acoustic sensors, data acquisition instruments, and speed encoders. The acquired signals are the raw observation signals during equipment operation. For example, a vibration acceleration sensor can be installed at a key monitoring location on the equipment under test, and an encoder can be installed at the speed detection location. The sensor sampling frequency is set according to the equipment's operating characteristics.

[0035] The control equipment operates within a full range of operating conditions, covering typical operating conditions such as constant speed, variable speed, and variable load in actual equipment operation. Under each operating condition, continuous noisy observation signals are collected for a preset time (e.g., at least 5 minutes), and speed pulse data is recorded by an encoder to ensure that the observation signals correspond one-to-one with the operating conditions. Simultaneously, noiseless calibration signals under the stable no-load operating state of the equipment are collected as standard data for model training.

[0036] The raw signals acquired by the sensors are initially cleaned to remove invalid signals during equipment start-up and shutdown, abnormal impact signals exceeding the measurement range, and retain valid data during stable equipment operation, providing a basic source for subsequent noise reduction processing.

[0037] In some embodiments, a sliding window is used to segment the effective noisy time-domain signal and the noiseless calibration signal respectively, including: using the Hanning sliding window function to adaptively determine the window width based on the sampling frequency and the device switching frequency; and using the same sliding window parameters to perform sliding window segmentation on the effective noisy time-domain signal and the noiseless calibration signal respectively based on the window width to obtain the corresponding sub-signal segments.

[0038] In applications, the Hanning sliding window function can be used to segment the pre-processed effective noisy time-domain signal. The window width W is determined by the sampling frequency. With equipment frequency conversion (Acquired synchronously from the encoder) Adaptive settings ensure that the signal within each window fully characterizes the periodic features of the device's operation:

[0039] Where K is the number of complete rotation cycles contained in each window, and its specific value is determined by grid search optimization within the range of [2,8] based on the equipment type and operating conditions: for low-frequency rotating equipment (rotation frequency <10Hz), K∈[4,8], and for high-frequency rotating equipment (rotation frequency ≥10Hz), K∈[2,4]. The window overlap rate is fixed at 50%, and the step size of adjacent windows is W / 2.

[0040] Then, amplitude normalization is performed on all sub-signals after sliding window segmentation, normalizing the signal amplitude of all samples to the [-1, 1] interval to eliminate model training bias caused by dimensional differences. Next, a corresponding noiseless calibration signal is matched to each noisy sample as the gold standard for denoising, generating a standardized time-series sample set. It should be noted that the noisy samples and noiseless calibration signals correspond one-to-one with the same window index, and the window segmentation parameters are completely identical. The expression for the corresponding standardized time-series sample set is:

[0041]

[0042] In the formula, For the first One noisy observation time series sample, For the first The noise-free standard effective signal corresponding to each sample The total number of samples.

[0043] For example, the standardized time series sample set can be randomly divided into a training set, a validation set, and a test set in a 6:2:2 ratio.

[0044] In some embodiments, using single-point time-series subsequences and single-point time-series segments in multi-point signals as nodes, the standardized time-series sample set is transformed into node feature representations, and a standardized undirected graph structure is constructed by combining the K-nearest neighbor algorithm. This includes: using the single-point time-series subsequences corresponding to the sliding window and the single-point time-series segments in multi-point signals as nodes of the graph, extracting the time-domain and frequency-domain statistical features of each node, generating high-dimensional feature vectors for the nodes, and forming a node feature matrix; determining the K value based on the total number of nodes, using the K-nearest neighbor algorithm to perform neighborhood search and edge connection for the nodes, generating an original adjacency matrix; normalizing the node feature matrix and performing symmetric normalization on the original adjacency matrix, and combining the normalized node feature matrix with the normalized adjacency matrix to form a standardized undirected graph structure.

[0045] In the application, after dividing the training, validation, and test sets, graph node partitioning and feature construction are performed on the time-series samples of the divided training, validation, and test sets: single-measurement-point time-series subsequences within a sliding window are used as single nodes, and multi-measurement-point signals are used as independent nodes for single-measurement-point time-series segments. Basic time-domain and frequency-domain statistical features of each node are extracted, and a high-dimensional feature vector with a dimension of, for example, 2048 is output as the node features of the graph structure. These node feature vectors are stacked in node order to form the node feature matrix. ,in This represents the total number of nodes.

[0046] An undirected graph structure is constructed using the K-Nearest Neighbors (KNN) algorithm, with each node as the center of the search neighborhood. The value of K is dynamically set according to the total number of nodes N.

[0047] in, K is a scaling factor, optimized within the range [2,6] by combining grid search with graph reconstruction accuracy on the validation set: for small-scale graph structures with N<100, K is typically 5~8; for medium-scale graphs with 100≤N<500, K is typically 8~15; for large-scale graphs with N≥500, K is typically 15~25. Nodes within the neighborhood are connected by edges to generate the original adjacency matrix A of the graph. At the beginning of each training round, the value of K is dynamically fine-tuned based on the node feature distribution of the current batch: if the average clustering coefficient of the graph under the current K value is lower than the threshold of 0.3, the K value is appropriately increased (step size of 1) to enhance graph connectivity; if the average node degree exceeds N / 3, the K value is appropriately decreased, with a step size of 1 when decreasing the K value. The fine-tuning range of the K value is kept no less than 3 to avoid the graph becoming too dense and overly smooth.

[0048] Then, the node feature vectors are normalized to eliminate model training bias caused by differences in dimensions and values; simultaneously, the original adjacency matrix is ​​symmetrically normalized to balance the weights of nodes with a large number of edges, as shown in the formula:

[0049] In the formula, Original adjacency matrix The degree matrix, It is the symmetric normalized adjacency matrix.

[0050] Finally, the normalized node feature matrix is... With normalized adjacency matrix Combine to construct a complete graph structure data The datasets were divided into three categories according to the established proportions to provide standardized data for subsequent training.

[0051] In some embodiments, such as Figures 2 to 4 As shown, the ternary causal noise reduction model includes: the SCAM basic causal generation layer, the three-branch decoupling layer, and the causal fidelity reconstruction layer.

[0052] The SCAM basic causal generation layer is used to perform SCAM causal formal reconstruction of the input graph structure data in terms of effective signal, noise, and operating condition. It uses the ternary causal independence axiom to define the rigid causal boundary between the effective signal, noise, and operating condition.

[0053] The three-branch decoupling layer includes a differentiable causal discovery submodule and a three-branch graph attention submodule. The differentiable causal discovery submodule is used to learn the hidden features of the nodes based on the output of the SCAM basic causal generation layer to obtain a backpropagable causal adjacency matrix. The three-branch graph attention submodule uses the causal adjacency matrix to modulate the three-branch attention weights, extract the effective signal causal features, noise interference features and operating condition features respectively, and drive the causal level decoupling of the three through causal independence loss.

[0054] The causal fidelity reconstruction layer is used to reconstruct signals based on the causal characteristics of the decoupled effective signals. It applies multiple fidelity constraints in the time domain, frequency domain, and causal structure, and outputs the reconstructed effective signal.

[0055] The model uses graph-structured data. The input consists of a SCAM basic causal generation layer, a three-branch decoupling layer, and a causal fidelity reconstruction layer connected in series. With the help of the forward propagation mechanism, it completes the causal mechanism modeling of effective signal and noise, feature decoupling, and clean signal reconstruction.

[0056] The SCAM basic causal generation layer performs causal formal reconstruction of noisy observation signals, defining a rigid causal boundary between effective signals and noise at the signal generation mechanism level. It is important to emphasize that the fundamental difference between the SCAM basic causal generation layer and existing causal denoising methods is that existing denoising methods based on causal backdoor paths construct a structural causal model with homogeneous noise as a confounding factor. Its core mathematical form is:

[0057]

[0058] in: for right causal effect for right causal effect for right causal effect for right causal effect for right causal effect for right The causal effect; These are observable intermediate variables of the operating condition, used as adjustment conditions for the backdoor path to achieve... and The model achieves conditional independence. The core idea is to introduce observable intermediate variables. As a conditioning condition for the backdoor path, non-homogeneous noise is eliminated through residual regression. This method is essentially a system separation at the level of independently adjusting statistical conditions, rather than at the level of causal generation mechanisms.

[0059] In contrast, the ternary causal denoising model constructed in this invention defines three completely independent causal systems at the signal generation mechanism level: the effective signal causal system, the noise causal system, and the operating condition causal system. Its core lies in defining a rigid causal independence boundary between the three systems (based on the definition of the ternary causal independence axiom), transforming the denoising problem from "statistical decorrelation" to "causal system separation." This difference, mathematically speaking, is unconditional causal independence that does not depend on any observable intermediate variables, rather than statistical independence conditioned on intermediate variables.

[0060] Among them, each node of the valid signal Its generation follows its own inherent causal mechanism and has no causal relationship with noise. The causal generation equation of the effective signal is:

[0061] In the formula, The node set after topological transformation of the signal graph; For the nonlinear causal generation function of the effective signal, For nodes The set of parent nodes in a valid causal subgraph; The intrinsic independent noise of the effective signal is related only to the physical properties of the signal itself and satisfies the following conditions: , .

[0062] New state parameter The generating equation is:

[0063] in, For nodes The corresponding hidden variables of the operating condition. This refers to the exogenous independent disturbance term of the state variable.

[0064] The noise generation mechanism is completely independent of the causal system of the effective signal; the causal generation equation of the noise signal is:

[0065] In the formula, is the noise generation function; Nodes in the noise subgraph The set of parent nodes; This represents the exogenous independent interference term of the noise.

[0066] In the preceding content, the core ternary causal independence axiom is proposed as a rigid mathematical criterion for noise reduction and separation. This clarifies the essential causal boundaries of noise, effective signal, and operating condition, serving as the core criterion for their separation. The formula is as follows:

[0067] That is, the exogenous interference terms among the effective signal S, noise N, and operating condition Z are mutually independent.

[0068] This theorem also satisfies the observation additive constraint: .

[0069] In contrast to existing causal noise reduction methods, which can only achieve statistical independence conditioned on intermediate variables, this approach... and In a given Under the condition of mutual independence ( This is a conditional, weakened independence constraint, the effectiveness of which is highly dependent on intermediate variables. The observability and completeness of the data. Furthermore, while existing deep learning-based denoising methods achieve this through statistical decorrelation... and The separation is possible, but its independence constraint only holds at the data distribution level and cannot guarantee independence. and True decoupling at the causal generation mechanism level, when operating conditions change, and Will be due to common suffering The influence of this can create spurious correlations, causing the binary independence hypothesis to fail.

[0070] In contrast, the ternary causal independence axiom proposed in this invention is an unconditional and reinforced constraint on causal independence. It does not rely on any intermediate variables or environmental labels, and directly defines the effective signal from the level of the signal generation mechanism. ,noise Operating conditions A rigid causal boundary exists between the three. Specifically, by introducing... As a third variable, this invention achieves Conditional causal independence—that is, under any operating condition, the generation mechanisms of effective signals and noise remain causally independent, fundamentally solving the problem in variable operating condition scenarios. and Because of common suffering This addresses the challenge of spurious correlations and incomplete separation caused by external influences. This axiom provides, for the first time, a rigorous ternary causal separation criterion for signal denoising, achieving a paradigm shift from "numerical fitting at the statistical level" in traditional methods to "decoupling and separating three independent causal systems," providing a core mathematical foundation for subsequent denoising processes.

[0071] The three-branch decoupling layer uses SCAM to define causal boundaries and embeds a differentiable causal discovery algorithm based on the Non-combinatorial Optimization via Trace Exponential and Augmented Lagrangian (NOTEALS) framework. Iteratively, it learns and separates three mutually exclusive subgraphs: the effective signal causal subgraph. Noise interference subgraph and operating condition sub-diagram .

[0072] Existing causal denoising methods employ traditional residual regression techniques, where the modeling of the causal structure and the denoising process are disconnected, failing to achieve end-to-end joint optimization. Furthermore, most existing algorithms are discrete combinatorial optimization forms, non-differentiable, and cannot be embedded into neural networks for gradient propagation. This invention, for the first time, deeply integrates the differentiable causal discovery algorithm of the NOTEALS framework with the graph attention mechanism of GCN, enabling the causal adjacency matrix to... gradient It can backpropagate to the parameters of each layer of GCN through the attention weight calculation chain, realizing end-to-end joint parameter space optimization of the causal discovery module and the GCN feature extraction module.

[0073] The core optimization problem of differentiable causal discovery algorithms is formulated as follows: learning a weighted adjacency matrix. , making The corresponding directed graph satisfies the directed acyclic constraint and maximizes the observed data. The likelihood. Specifically, the objective function to be optimized is:

[0074]

[0075] in, For the sample size, The node feature matrix, It is the Frobenius norm. For L1 sparse regularization terms, [0.001,0.1] For sparse regularization coefficients, The acyclic constraint function ensures that the learned causal graph satisfies the directed acyclic graph (DAG) condition. The structure is as follows: The augmented Lagrangian method is used to transform the constrained optimization into an unconstrained form:

[0076]

[0077] In the formula As a penalty factor, it gradually increases by a factor of 2 during the iteration process to force the acyclic constraint to be satisfied.

[0078] The coupling mechanism between this differentiable causality discovery algorithm and the GCN backpropagation mechanism is as follows: In each training iteration, the hidden layer representation of the node is first determined based on the output of the current GCN encoding layer. As input, calculate the weighted adjacency matrix. ; then With the original adjacency matrix By fusing differentiable gating mechanisms, graph attention weights for a three-branch decoupling layer are generated:

[0079]

[0080]

[0081] in For the Sigmoid activation function, Mapping to the [0,1] interval serves as a probabilistic interpretation of the strength of causal connections; , , These are the attention weight coefficients for the effective signal subgraph, the noise subgraph, and the operating condition subgraph, respectively. The sum of these three is 1, forming a mutually exclusive soft assignment. gradient By backpropagating the aforementioned attention weight calculation chain to the parameters of each layer of GCN, end-to-end joint optimization of the differentiable causal discovery submodule and the three-branch graph attention submodule is achieved.

[0082] Existing Graph Attention Networks (GATs) determine attention weights solely by the similarity between node features, while the three-branch attention weights of this invention are additionally multiplied by the causal adjacency matrix. The Sigmoid mapping. This design allows causal priors to directly participate in the calculation of attention weights, and the update direction of the weight coefficients is dominated by causal criteria rather than driven solely by data fitting error.

[0083] A three-branch decoupling layer is used to learn features from the two subgraphs, extracting the causal features FS of the effective signal and the interference features FN of the noise. Causal independence loss is then applied. Defined as the Hilbert-Schmidt Independence Criterion (HSIC) between FS and FN:

[0084]

[0085]

[0086]

[0087]

[0088] in, , , They are respectively , , The Gram matrix is ​​denoted by H, where H is the centered matrix. This loss function forces... , Complete decoupling in the feature space eliminates causal dependencies and avoids residual coupling of effective information and noise.

[0089] We introduce HSIC into the field of signal denoising as a loss of causal independence. HSIC measures the dependence of two feature sets in the reproducing kernel Hilbert space. When the HSIC of all three approaches 0, it means that the effective signal features, noise features, and state features are completely independent in the kernel space. This is a stronger causal constraint than statistical decorrelation because it can capture nonlinear dependencies, not just linear correlations.

[0090] Causal fidelity reconstruction layer is based on the causal features of the decoupled effective signal. The algorithm reconstructs the signal with high fidelity through a graph convolutional decoding layer, outputting a denoised result that is distortion-free and identical to the original valid signal. Simultaneously, based on the SCAM causal generation mechanism, a causal structure fidelity constraint is constructed to force the denoised signal to have a completely consistent causal dependency structure with the noise-free standard signal. This ensures that the inherent physical characteristics and nonlinear coupling relationships of the valid signal are fully preserved, preventing the accidental deletion of valid information. The final output is a high-fidelity, clean, and valid signal after denoising, providing high-quality data input for subsequent downstream tasks such as state recognition and lifetime prediction.

[0091] In some embodiments, such as Figure 2 As shown, the causal-fidelity reconstruction layer includes multiple stacked graph convolutional decoding layers, as well as temporal-domain fidelity constraint branches, frequency-domain fidelity constraint branches, and causal structure fidelity constraint branches.

[0092] Multiple graph convolutional decoding layers are used to sequentially perform graph domain signal reconstruction based on the causal features of the input valid signal, and output the reconstructed valid signal. Temporal fidelity constraint branches, frequency-domain fidelity constraint branches, and causal structure fidelity constraint branches are used to apply temporal, frequency, and causal structure fidelity constraints respectively during model training.

[0093] It is understandable that the causal fidelity reconstruction layer uses stacked graph convolutional decoding layers as the core reconstruction pathway. It can fully rely on the topological awareness capability of graph convolution to perform layer-by-layer feature restoration and graph domain signal reconstruction of the causal features of the decoupled clean and effective signal. It accurately reproduces the feature distribution and waveform information of the original signal, ensuring the feature integrity of the forward reconstruction process. At the same time, it is equipped with three parallel fidelity constraint branches of time domain, frequency domain, and causal structure. During the model training stage, it performs joint constraint correction on the reconstruction results from multiple dimensions, respectively ensuring the consistency of the waveform amplitude in the time domain, the integrity of the effective features in the frequency domain, and the rationality of the causal topological structure of the reconstructed signal. It avoids signal distortion, feature loss, or structural offset problems caused by single constraints, so that the model can retain the effective signal features of the actual operation of the device to the maximum extent while filtering out noise, significantly improving the accuracy and reliability of noise reduction and reconstruction.

[0094] In some embodiments, the target loss function includes triple-fidelity reconstruction loss, ternary causal decoupling loss, invariance regularization term, and acyclic constraint term.

[0095] The triple-fidelity reconstruction loss includes time-domain fidelity loss, frequency-domain fidelity loss, and causal structure fidelity loss, which are used to impose time-domain, frequency-domain, and causal structure constraints on the reconstructed signal; the ternary causal decoupling loss is used to constrain the causal features, noise interference features, and operating condition features of the effective signal to be independent of each other; the invariance regularization term is used to constrain the operating condition features to have environmental invariance; and the acyclic constraint term is used to constrain the causal adjacency matrix to meet the directed acyclic requirement.

[0096] The training process constructs the total loss function, i.e., the target loss function, based on the causal constraints of SCAM, with the following formula:

[0097] In the formula The balancing hyperparameters for each loss term, The loss is used for causal structure fidelity reconstruction; it includes time-domain fidelity loss, frequency-domain fidelity loss and causal structure fidelity loss, ensuring that the effective signal is undistorted after denoising. These are the weighting coefficients for the three terms. This can be adjusted based on the validation set. The time-domain fidelity loss is:

[0098]

[0099] In the formula, The total number of nodes in the graph. For node feature dimensions; For the reconstruction output of the first Each node has a valid signal. To correspond to the noise-free calibration signal, the amplitude and shape of the time-domain waveform are constrained to be consistent. The frequency-domain fidelity loss is:

[0100]

[0101] In the formula, For Fast Fourier Transform, The Frobenius norm is used to constrain the consistency between the spectral amplitude and the characteristic frequency components, preventing the false filtering of effective high-frequency information. The causal structure fidelity loss is:

[0102]

[0103] In the formula, To reconstruct the causal adjacency matrix corresponding to the valid signal, This is the standard causal adjacency matrix corresponding to the noiseless calibration signal. The inherent causal topology of the constraint signal is consistent, ensuring that the physical coupling relationship and generation mechanism are not disrupted.

[0104] This is the ternary causal decoupling loss, which is the causal independence loss based on the intervention distribution, ensuring that noise and effective signals are decoupled at the causal level.

[0105] It is a ternary causal invariant regularization term, in which the training samples are classified according to the latent variables of the working conditions. The clustering results are divided into A subset of operating conditions; For the first The effective signal causal adjacency matrix learned under each operating condition subset. This term ensures that the intrinsic generation mechanism of the effective signal has operating condition invariance by minimizing the difference in causal topology under different operating conditions.

[0106] For the acyclic constraint term of the effective causal subgraph, in the formula... The learned causal adjacency matrix; This is the Hadamard product of matrices (element-wise multiplication). It is a matrix exponential function; Let B be the total number of nodes in the graph. The constraint function is true if and only if the directed graph corresponding to B is a directed acyclic graph. This ensures that the causal structure conforms to the definition of a causal mechanism.

[0107] For example, during training, an adaptive moment estimation optimization algorithm is used to update the model parameters, and hyperparameters are set according to the noise reduction scenario: the initial learning rate is set to 6×10. -4The dropout rate is 0.5, the learning rate decay coefficient is 0.5, and the number of training epochs is 300. During training, the validation set data is used as the object for real-time monitoring of model performance. When the validation set noise reduction loss does not decrease for 10 consecutive epochs, the learning rate decay strategy is executed to prevent the model from overfitting. The training continues iteratively until the loss function converges, and the training ends.

[0108] The training model was tested for denoising performance using graph structure data from the test set. The testing process was completely consistent with the forward propagation process of the training set. After graph structure construction, SCAM causal generation layer modeling, three-branch causal decoupling, and causal fidelity reconstruction, the denoised clean and effective signal of each test sample was finally output.

[0109] The noise reduction effect of the model is quantified by four core indicators: signal-to-noise ratio, root mean square error, spectral fidelity, and waveform similarity coefficient. At the same time, time-domain waveform comparison and frequency-domain comparison before and after noise reduction are plotted to intuitively verify the noise reduction fidelity performance of the model.

[0110] To verify the effectiveness of this scheme, the proposed method was used to denoise the corresponding signal on the publicly available planetary gearbox dataset from Beijing Jiaotong University, with an input signal-to-noise ratio of 0dB as the baseline. The signals before and after denoising were then compared. Figure 5 A comparative diagram showing the difference between normal signal noise reduction and noise reduction in a planetary gearbox. Figure 6 A comparative diagram showing the difference between before and after noise reduction of wear signals in a planetary gearbox; Figure 7 This is a schematic diagram comparing the noise reduction of the planetary gearbox tooth root fracture signal before and after. The proposed method was then compared with conventional noise reduction methods, and the results are shown in Table 1.

[0111] Table 1 Evaluation indicators for various noise reduction methods

[0112] Under strong noise conditions (input SNR=-5dB), the output signal-to-noise ratio of this invention can reach 12.3dB, which is about 5.8dB higher than the conventional GCN method and about 9.1dB higher than the traditional wavelet method, verifying the significant advantage of the causal-level decoupling strategy in extremely low signal-to-noise ratio scenarios. To leverage the generalization advantage brought by causal invariance, this invention designed out-of-distribution (OOD) generalization test scenarios: the training set covers two noise types under constant speed conditions: Gaussian white noise and mechanical impact noise; the test set is set as follows: (1) variable speed conditions (speed fluctuation ±20%); (2) unseen noise types; (3) mixed noise scenarios. Under the above OOD test conditions, the noise reduction results for the corresponding scenarios are shown in Table 2.

[0113] Table 2 Noise reduction results in different scenarios

[0114] Experimental results show that the effective signal subgraph extracted by the present invention based on causal invariance learning has significant invariance to operating conditions and noise type. When the distribution offset increases, the performance advantage over conventional methods actually expands, effectively verifying the fundamental advantage of the causal driving mechanism in improving the model's OOD generalization ability.

[0115] Compared with conventional GCN denoising methods: This invention, through HSIC and causality-invariant regularization, enables the model to maintain stable denoising performance in scenarios with no known noise type or varying operating conditions. Under strong noise conditions with an input SNR of -5dB, the output SNR of this invention can reach 12.3dB, which is an improvement of approximately 5.8dB compared to conventional GCN methods and approximately 9.1dB compared to traditional wavelet methods.

[0116] In terms of engineering practicality, this invention achieves end-to-end, differentiable fusion of causal inference and graph neural networks. The time required for a single forward inference iteration is only about 12% longer than that of traditional GCN, while the noise reduction performance is improved by more than 30%. Furthermore, in variable speed scenarios (speed fluctuation ±20%), the output SNR of traditional binary noise reduction methods decreases by more than 25%, while the ternary causal model of this invention, which introduces the operating state Z as a third variable, experiences a decrease of less than 10%, verifying the significant advantage of the ternary causal independence axiom in variable operating conditions. Considering overall cost-effectiveness, it is significantly superior to existing solutions, meets the efficiency requirements of real-time preprocessing of industrial signals, and possesses good engineering implementation capabilities.

[0117] In some embodiments, please refer to Figure 8 This application provides an end-to-end mechanical signal noise reduction system 80, including: an acquisition module 81, a sample construction module 82, a conversion module 83, a training module 84, and a noise reduction module 85.

[0118] The acquisition module 81 is used to acquire the original noisy observation signal and the noiseless calibration signal of the device under various operating conditions, and to preprocess the original noisy observation signal to obtain the effective noisy time domain signal.

[0119] The sample construction module 82 is used to segment the effective noisy time-domain signal and the noiseless calibration signal using a sliding window, and to perform amplitude standardization on the segmented effective noisy time-domain sub-signals, and match the corresponding noiseless calibration sub-signals as noise reduction standards to construct a standardized time-series sample set.

[0120] The conversion module 83 is used to convert the standardized time series sample set into node feature representations using single-measurement-point time series subsequences and single-measurement-point time series segments in multi-measurement-point signals as nodes, and to construct a standardized undirected graph structure by combining the K-nearest neighbor algorithm to obtain graph structure data.

[0121] Training module 84 is used to train the ternary causal denoising model using graph structure data and target loss function to obtain the target denoising model. The ternary causal denoising model defines the rigid causal boundary between effective signal, noise, and operating condition through the SCAM ternary causal independence axiom, transforms the ternary causal independence axiom into a differentiable loss function, learns the causal adjacency matrix based on the input graph structure data using a differentiable causal discovery algorithm, and uses the causal adjacency matrix to modulate the attention weights of the three-branch graph attention network to drive the decoupling of effective signal, noise, and operating condition at the causal level.

[0122] The noise reduction module 85 is used to input the signal to be denoised into the target noise reduction model to obtain the corresponding noise reduction signal.

[0123] It should be noted that the end-to-end mechanical signal noise reduction system 80 provided in this application embodiment and the end-to-end mechanical signal noise reduction method provided in this application embodiment are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned end-to-end mechanical signal noise reduction method, and the repeated parts will not be described again.

[0124] In some embodiments, an electronic device provided in this application includes a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the above-described end-to-end mechanical signal noise reduction method.

[0125] This application also provides a non-transitory computer storage medium storing a computer program thereon, which, when executed by a processor, implements the end-to-end mechanical signal noise reduction method described above. This computer-readable medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into that device / apparatus / system. The aforementioned computer-readable medium carries one or more programs, which, when executed, implement the method as described in the embodiments of this application.

[0126] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0127] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application, even if such combinations or combinations are not explicitly described in this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An end-to-end mechanical signal noise reduction method, characterized in that, include: The original noisy observation signal and noiseless calibration signal of the device under various operating conditions are acquired, and the original noisy observation signal is preprocessed to obtain an effective noisy time domain signal. The effective noisy time-domain signal and the noiseless calibration signal are segmented using a sliding window, and the amplitude of the segmented effective noisy time-domain sub-signals is normalized. The corresponding noiseless calibration sub-signals are then matched as noise reduction standards to construct a standardized time-series sample set. Using single-measurement-point time-series subsequences and single-measurement-point time-series segments in multi-measurement-point signals as nodes, the standardized time-series sample set is transformed into node feature representations, and a standardized undirected graph structure is constructed by combining the K-nearest neighbor algorithm to obtain graph structure data. The ternary causal denoising model is trained using the graph structure data and the target loss function to obtain the target denoising model. The ternary causal denoising model defines the rigid causal boundary between the effective signal, noise, and operating condition through the SCAM ternary causal independence axiom. The ternary causal independence axiom is transformed into a differentiable loss function. Based on the input graph structure data, a differentiable causal discovery algorithm is used to learn the causal adjacency matrix. The causal adjacency matrix is ​​used to modulate the attention weights of the three-branch graph attention network to drive the decoupling of the effective signal, noise, and operating condition at the causal level. The signal to be denoised is input into the target denoising model to obtain the corresponding denoised signal.

2. The end-to-end mechanical signal noise reduction method as described in claim 1, characterized in that, The preprocessing of the original noisy observation signal to obtain an effective noisy time-domain signal includes: The invalid signals during the equipment start-up and shutdown phases in the original noisy observation signal are removed, and abnormal impact signals exceeding the measurement range are also removed to obtain the effective noisy time-domain signal.

3. The end-to-end mechanical signal noise reduction method as described in claim 1, characterized in that, The step of segmenting the effective noisy time-domain signal and the noise-free calibration signal using a sliding window includes: The Hanning sliding window function is used to adaptively determine the window width based on the sampling frequency and the device switching frequency; Based on the window width, the same sliding window parameters are used to perform sliding window segmentation on the effective noisy time-domain signal and the noiseless calibration signal to obtain corresponding sub-signal segments.

4. The end-to-end mechanical signal noise reduction method as described in claim 1, characterized in that, The process involves using single-measurement-point time-series subsequences and single-measurement-point time-series segments from multi-measurement-point signals as nodes, transforming the standardized time-series sample set into node feature representations, and constructing a standardized undirected graph structure using the K-nearest neighbor algorithm, including: Using the single-measurement point time sequence subsequence corresponding to the sliding window and the single-measurement point time segment under the multi-measurement point signal as nodes of the graph, the time domain and frequency domain statistical features of each node are extracted to generate high-dimensional feature vectors of the nodes and form a node feature matrix. The value of K is determined based on the total number of nodes. The K-nearest neighbor algorithm is used to search the neighborhood of the nodes and complete the edge connection to generate the original adjacency matrix. The node feature matrix is ​​normalized, and the original adjacency matrix is ​​symmetrically normalized. The normalized node feature matrix and the normalized adjacency matrix are combined to form the standardized undirected graph structure.

5. The end-to-end mechanical signal noise reduction method as described in claim 1, characterized in that, The ternary causal noise reduction model includes: an SCAM basic causal generation layer, a three-branch decoupling layer, and a causal fidelity reconstruction layer; The SCAM basic causal generation layer is used to perform SCAM causal formal reconstruction of the input graph structure data in terms of effective signal, noise, and operating condition. The rigid causal boundary between the effective signal, noise, and operating condition is defined by the ternary causal independence axiom. The three-branch decoupling layer includes a differentiable causal discovery submodule and a three-branch graph attention submodule. The differentiable causal discovery submodule is used to learn based on the hidden layer features of the nodes output by the SCAM basic causal generation layer to obtain a backpropagable causal adjacency matrix. The three-branch graph attention submodule uses the causal adjacency matrix to modulate the three-branch attention weights, extract the effective signal causal features, noise interference features, and operating condition features respectively, and drives the causal level decoupling of the three through causal independence loss. The causal fidelity reconstruction layer is used to reconstruct the signal based on the causal characteristics of the decoupled effective signal, apply multiple fidelity constraints in the time domain, frequency domain, and causal structure, and output the reconstructed effective signal.

6. The end-to-end mechanical signal noise reduction method as described in claim 5, characterized in that, The causal-fidelity reconstruction layer includes multiple stacked graph convolutional decoding layers, as well as temporal-domain fidelity constraint branches, frequency-domain fidelity constraint branches, and causal structure fidelity constraint branches. Multiple graph convolutional decoding layers are used to sequentially perform graph domain signal reconstruction based on the causal features of the input valid signal, and output the reconstructed valid signal; The time-domain fidelity constraint branch, frequency-domain fidelity constraint branch, and causal structure fidelity constraint branch are used to apply time-domain, frequency-domain, and causal structure fidelity constraints, respectively, during model training.

7. The end-to-end mechanical signal noise reduction method as described in claim 6, characterized in that, The target loss function includes triple-fidelity reconstruction loss, ternary causal decoupling loss, invariance regularization term and acyclic constraint term; The triple-fidelity reconstruction loss includes time-domain fidelity loss, frequency-domain fidelity loss, and causal structure fidelity loss, which are used to apply time-domain, frequency-domain, and causal structure constraints to the reconstructed signal. The ternary causal decoupling loss is used to constrain the causal characteristics of the effective signal, the noise interference characteristics, and the operating condition characteristics to be independent of each other. The invariance regularization term is used to constrain the operating condition characteristics to have environmental invariance. The acyclic constraint term is used to constrain the causal adjacency matrix to satisfy the directed acyclic requirement.

8. An end-to-end mechanical signal noise reduction system, characterized in that, include: The acquisition module is used to acquire the original noisy observation signal and the noiseless calibration signal of the device under various operating conditions, and to preprocess the original noisy observation signal to obtain the effective noisy time domain signal. The sample construction module is used to segment the effective noisy time-domain signal and the noiseless calibration signal using a sliding window, and to perform amplitude normalization on the segmented effective noisy time-domain sub-signals, and match the corresponding noiseless calibration sub-signals as noise reduction standards to construct a standardized time-series sample set. The conversion module is used to convert the standardized time series sample set into node feature representations using single-measurement-point time series subsequences and single-measurement-point time series segments in multi-measurement-point signals as nodes, and to construct a standardized undirected graph structure by combining the K-nearest neighbor algorithm to obtain graph structure data. The training module is used to train the ternary causal denoising model using the graph structure data and the target loss function to obtain the target denoising model. The ternary causal denoising model defines the rigid causal boundary between the effective signal, noise, and operating condition through the SCAM ternary causal independence axiom, transforms the ternary causal independence axiom into a differentiable loss function, learns the causal adjacency matrix based on the input graph structure data using a differentiable causal discovery algorithm, and uses the causal adjacency matrix to modulate the attention weights of the three-branch graph attention network to drive the decoupling of the effective signal, noise, and operating condition at the causal level. The noise reduction module is used to input the signal to be denoised into the target noise reduction model to obtain the corresponding noise reduction signal.

9. An electronic device, characterized in that, It includes a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the end-to-end mechanical signal noise reduction method as described in any one of claims 1 to 7.

10. A non-transitory computer storage medium, characterized in that, It stores a computer program, wherein the computer program, when executed by a processor, implements the end-to-end mechanical signal noise reduction method as described in any one of claims 1 to 7.