Equipment predictive maintenance method and system based on multi-source data fusion and deep learning

Abstract

This invention discloses a method and system for predictive maintenance of equipment based on multi-source data fusion and deep learning, relating to the field of equipment maintenance technology. The method includes mapping multi-source heterogeneous sensor data to a probability distribution measure within a unified time window, aligning cross-physical field data using optimal transport theory, and constructing a coupled feature manifold. Further, it extracts fatigue degradation features that remain unchanged under varying operating conditions using an isovariant neural dynamics operator based on Lie group theory. This feature is then mapped to an information geometric manifold space, a Riemannian metric is constructed using the Fisher information matrix, and geodesic constraints are introduced to control the monotonic evolution of the predicted trajectory along the shortest physical path. Finally, a strictly monotonically decreasing remaining service life prediction curve is output through the Riemann arc length integral, and equipment maintenance strategies are formulated based on the prediction results. This method effectively improves multi-source data fusion capabilities and the physical consistency of prediction results, and is suitable for intelligent operation and maintenance scenarios of complex industrial equipment.

Description

Technical Field

[0001] This invention relates to the field of equipment maintenance technology, and in particular to a method and system for predictive maintenance of equipment based on multi-source data fusion and deep learning. Background Technology

[0002] As industrial equipment becomes increasingly complex, its operation typically involves the coordinated changes of multi-source heterogeneous sensor data, such as vibration, temperature, and pressure. These data exhibit significant differences in sampling frequency, time scale, and physical properties, making it difficult for traditional methods based on single signals or simple time alignment to accurately depict the true operating state of the equipment. Existing predictive maintenance methods largely rely on deep learning models to fit historical data, using time series models or feature engineering to predict remaining service life.

[0003] However, this type of method generally has two problems: First, there is a lack of a unified alignment mechanism between signals from different physical fields, and direct time synchronization can easily introduce phase distortion and information distortion; Second, deep learning models are essentially data-driven models and lack constraints on the physical laws of devices, which can easily lead to unreasonable phenomena such as non-physical oscillations or even lifespan recovery in the prediction results. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a device prediction and maintenance method based on multi-source data fusion and deep learning to solve the problems of difficulty in aligning multi-source heterogeneous data and inaccurate life prediction caused by the lack of physical constraints on prediction results.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a device prediction and maintenance method based on multi-source data fusion and deep learning, which includes acquiring multi-source asynchronous sensor data during device operation and mapping heterogeneous physical signals with different sampling frequencies that result in non-equidistant distribution on the time scale into a probability distribution measure within a unified time window.

[0008] Based on the optimal transport theory, the Wasserstein distance between various probability distribution measures is calculated, and the minimum work cost matrix across physical fields is obtained as a dynamic weight to generate an aligned coupled feature manifold graph.

[0009] By constructing an equivariant neural dynamics operator based on Lie group theory, the coupled feature manifold graph is subjected to group convolution processing to extract fatigue degradation feature vectors that remain symmetric under varying operating conditions.

[0010] The feature vector is projected onto the information geometric manifold space, and geodesic constraints that satisfy the thermodynamic dissipation law are introduced to make the model state monotonically evolve along the shortest physical path. The remaining lifetime prediction curve is output through Riemann arc length integral.

[0011] Based on the predicted curve, a maintenance plan is developed to achieve predictive maintenance of the equipment.

[0012] As a preferred embodiment of the device prediction and maintenance method based on multi-source data fusion and deep learning described in this invention, the heterogeneous physical signal includes high-frequency vibration signal and low-frequency temperature signal of the synchronously acquired device.

[0013] The probability distribution measure includes the source probability distribution measure converted from the high-frequency vibration signal. The conversion of the low-frequency temperature signal is a measure of the target probability distribution. ;

[0014] The source probability distribution measure is a probability density function generated by estimating the kernel density of the high-frequency vibration signal amplitude within a sliding time window; the target probability distribution measure is a probability density function generated by estimating the kernel density of the low-frequency temperature signal within a time window.

[0015] In the topological space formed by the sample points, a geometric cost function for measuring the physical transfer distance is constructed by mapping the discrete sample points of the source measure and the discrete sample points of the target measure through the square of the Euclidean distance.

[0016] In satisfying and In the solution space formed by all joint probability distributions that serve as constraints, the specific mass transfer scheme matrix is ​​solved using the optimal transfer integral equation.

[0017] The mass transfer scheme matrix obtained by the solution is used as a dynamic alignment weight matrix, and a dot product operation is performed with a multidimensional tensor containing the features of the original high-frequency vibration signal to generate the coupled feature manifold graph that completely eliminates the time phase deviation.

[0018] As a preferred embodiment of the device prediction and maintenance method based on multi-source data fusion and deep learning described in this invention, the construction of the isovariant neural dynamics operator based on Lie group theory includes,

[0019] For the spatial motion degrees of freedom of the core rotation axis system of the device under test, a mathematical Lie group is preset in the deep learning framework to describe continuous symmetric physical transformations; wherein, the mathematical Lie group is a set of group elements in group theory that describes the continuous rotation and translational rigid body motion of an object and has a topological geometric manifold structure.

[0020] Based on the exponential mapping of Lie algebras, the convolution operation defined in Euclidean space is reconstructed into a group convolution kernel operator defined on the Lie group manifold, so as to realize the equivariant mapping of features under group action.

[0021] The network feature extraction layer is constructed using the group convolution kernel operator, which extends parameter sharing from Euclidean space to group manifold space.

[0022] As a preferred embodiment of the equipment prediction and maintenance method based on multi-source data fusion and deep learning described in this invention, the extraction of fatigue degradation feature vectors that remain symmetrical under varying operating conditions includes:

[0023] The coupled feature manifold graph Input to contain group convolution kernel operator In deep learning networks;

[0024] Transformation operation for any physical disturbance representing a sudden change in operating conditions Force the group convolution kernel During network forward propagation, the equivalence space equation is satisfied:

[0025]

[0026] in, The mathematical Lie group is described In this context, a single group element represents the specific environmental rotational speed or load disturbance transformation; The input transform group action operator is the transform operation. The actual physical twisting and stretching matrix of the spatial coordinate system of the underlying coupled feature manifold graph; The output representation matrix is ​​the transformation operation. The isomorphic linear space transformation matrix mapped onto the feature dimension of the hidden layer of a high-level neural network;

[0027] After global pooling compression, a one-dimensional fatigue degradation feature vector is output.

[0028] As a preferred embodiment of the device prediction and maintenance method based on multi-source data fusion and deep learning described in this invention, the method involves: projecting the feature vector onto the information geometric manifold space, taking the fatigue degradation feature vector as input, and utilizing global weight parameters. The deep learning network generates a family of conditional probability distributions representing the uncertainty in the prediction of the remaining lifetime of equipment. ;

[0029] in, Global weight parameters represent the set of real numbers representing the weight matrices and bias vectors of all connection layers; The true degradation label of a given device is generated through feedforward network mapping. The conditional probability function output is given below;

[0030] Calculate the log-likelihood function of the family of conditional probability distributions with respect to the global weight parameters. The second-order partial derivatives are calculated, and their expected values ​​are obtained to derive the Fisher information matrix. ;

[0031] With the Fisher information matrix The identity matrix in the alternative Euclidean space is used as the Riemannian metric tensor to reconstruct the feature prediction mapping dimension of the deep learning network, generating a non-flat information geometric manifold space, and incorporating the fatigue-degraded feature vectors. As the origin of the dynamic coordinate system, it is mapped into the information geometric manifold space.

[0032] As a preferred embodiment of the device prediction and maintenance method based on multi-source data fusion and deep learning described in this invention, the method includes: introducing geodesic constraints that satisfy the thermodynamic dissipation law, and outputting the remaining service life prediction curve through Riemann arc length integration.

[0033] The fatigue degradation feature vector is used as the origin of the state mapping and projected onto the information geometric manifold space. Within the information geometric manifold space, a theoretically shortest smooth curve connecting the current healthy point and the endpoint of complete physical failure is solved by finding the optimal monotonic path of state evolution where the arc length of the curve becomes zero under the Fisher-Riemann metric. This curve is defined as the ideal physical dissipation geodesic. ;

[0034] Simultaneously, the fatigue degradation feature vector is tracked by continuous time step. Network evolution trajectory as depicted ;

[0035] Calculate the network inference trajectory With the ideal physical dissipation geodesic Riemannian geometric distance between ;

[0036] A nonlinear geodesic penalty is set based on the irreversible physical law of thermodynamic dissipation: based on Riemannian geometric distance. Constructing a geodesic penalty term for backpropagation Defined as:

[0037]

[0038] in, The penalty increase / decrease coefficient is a pre-set, large-scale scaling scalar constant used to exponentially dampen the amplification of the accuracy of the violation evolution.

[0039] The Riemann arc length integral is calculated along the state evolution trajectory constrained by the geodesic penalty term, and the remaining lifetime prediction curve with absolute monotonically decreasing is output.

[0040] As a preferred embodiment of the equipment predictive maintenance method based on multi-source data fusion and deep learning described in this invention, the predictive maintenance includes: a preset dynamic physical life safety red line threshold that is deeply coupled with the current equipment's ultimate physical fatigue fracture resistance process;

[0041] When the real-time polling program detects that the value of the remaining useful life prediction curve reaches the dynamic physical life safety threshold, a maintenance command is triggered.

[0042] Secondly, the present invention provides a device prediction and maintenance system based on multi-source data fusion and deep learning, including a data acquisition unit, which acquires multi-source asynchronous sensor data during the operation of the device and maps heterogeneous physical signals with different sampling frequencies that result in non-equidistant distribution on the time scale into a probability distribution measure within a unified time window.

[0043] The computational unit calculates the Wasserstein distance between various probability distribution measures based on optimal transport theory, obtains the minimum work cost matrix across physical fields as dynamic weights, and generates an aligned coupled feature manifold graph.

[0044] The extraction unit extracts fatigue degradation feature vectors that remain symmetric under varying operating conditions by constructing an equivariant neural dynamics operator based on Lie group theory and performing group convolution on the coupled feature manifold.

[0045] The evolution unit projects the feature vectors into the information geometric manifold space, introduces geodesic constraints that satisfy the thermodynamic dissipation law, and makes the model state monotonically evolve along the shortest physical path. The remaining lifetime prediction curve is output through Riemann arc length integral.

[0046] The maintenance unit formulates equipment maintenance plans based on the predicted curves to achieve predictive maintenance of the equipment.

[0047] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the device prediction and maintenance method based on multi-source data fusion and deep learning as described in the first aspect of the present invention.

[0048] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the device prediction and maintenance method based on multi-source data fusion and deep learning as described in the first aspect of the present invention.

[0049] The beneficial effects of this invention are as follows: By introducing optimal transmission theory, unified alignment of multi-source heterogeneous data in the probability measure space is achieved, effectively avoiding the phase distortion problem caused by traditional time synchronization; by constructing an isovariant neural dynamics operator based on Lie group theory, equipment operating condition disturbances are modeled as continuous symmetric transformations, achieving stable extraction of features under complex operating condition changes; further, by combining information geometry methods to construct a non-Euclidean manifold space and introducing geodesic constraints, the predicted trajectory satisfies the thermodynamic dissipation law, fundamentally avoiding non-physical oscillation problems in lifetime prediction; by obtaining strictly monotonically decreasing remaining lifetime prediction results through Riemann arc length integration, the physical consistency and reliability of the prediction results are improved; simultaneously, by combining dynamic safety thresholds to achieve closed-loop linkage between prediction results and equipment control, the safety of equipment operation and the accuracy of maintenance decisions are significantly improved. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart of a device predictive maintenance method based on multi-source data fusion and deep learning.

[0052] Figure 2 A computer equipment diagram for a device predictive maintenance method based on multi-source data fusion and deep learning. Detailed Implementation

[0053] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0054] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0055] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0056] Reference Figure 1 and Figure 2This is one embodiment of the present invention, which provides a device prediction and maintenance method based on multi-source data fusion and deep learning, including the following steps:

[0057] S1: Acquire multi-source asynchronous sensor data during equipment operation, and map heterogeneous physical signals with different sampling frequencies that result in non-equidistant distribution on the time scale into a probability distribution measure within a unified time window.

[0058] Multi-source asynchronous sensor data (heterogeneous physical signals) refers to sensor data originating from different physical fields and having different sampling frequencies and dynamic characteristics. Among them, high-frequency vibration signals are used to characterize the transient impact and local structural response of the equipment, while low-frequency temperature signals are used to characterize the energy dissipation and long-term degradation state of the equipment. Various signal data are acquired under a unified time reference system to ensure that different physical signals have a correlated time reference.

[0059] The probability distribution measure includes the source probability distribution measure converted from the high-frequency vibration signal. The conversion of the low-frequency temperature signal is a measure of the target probability distribution. .

[0060] The source probability distribution measure is a probability density function generated by estimating the kernel density of the high-frequency vibration signal amplitude within a sliding time window; the target probability distribution measure is a probability density function generated by estimating the kernel density of the low-frequency temperature signal within a time window.

[0061] It should be noted that the source probability distribution measure and the target probability distribution measure are used to represent the initial distribution and target distribution in the optimal transmission process, respectively. The source probability distribution measure provides the probabilistic quality to be transmitted, while the target probability distribution measure characterizes the distribution structure after transmission, thus establishing a mapping relationship between different physical fields. The constructed mapping relationship essentially matches the energy and probability dimensions of the high-frequency vibration signal (as the initial source, providing transient microscopic characteristics of equipment degradation) to the low-frequency temperature signal (as the target, providing macroscopic and stable degradation trend constraints). This mapping utilizes the stable and monotonic macroscopic physical laws of the temperature signal to regulate and constrain the chaotic and easily disturbed high-frequency vibration characteristics, thereby completely eliminating the phase deviation between the two. The mathematical manifestation of this mapping relationship is the mass transfer scheme matrix obtained through optimal transmission theory.

[0062] At the level of physical mechanism, when equipment undergoes fatigue degradation, the first thing that is triggered is microscopic or macroscopic mechanical damage, which is directly manifested as transient impact of high-frequency vibration signals. Therefore, vibration signals are the "cause" and source of degradation energy and are suitable as the initial distribution to be transmitted. On the other hand, the rise in temperature signals is essentially the "effect" of the further accumulation of mechanical friction and structural impact in the physical field and the irreversible heat dissipation. It has a natural hysteresis in response and represents the final state after energy conversion, so it is suitable as the target distribution of the receiving topology.

[0063] S2: Based on the optimal transport theory, calculate the Wasserstein distance between each probability distribution measure, obtain the minimum work cost matrix across physical fields as dynamic weights, and generate an aligned coupled feature manifold graph.

[0064] It should be noted that during continuous operation, the equipment generates a one-dimensional high-frequency mechanical vibration time-series signal and a low-frequency ambient temperature time-series signal. The system sets a sliding time window with a fixed length and sliding step size to continuously capture the time-series signals. For the local time series data within each sliding time window, the system uses a feature extraction module to condense it into a set of feature values ​​characterizing the current transient state of the equipment (the feature values ​​include, but are not limited to, time-domain statistical features such as root mean square, local peak value, and kurtosis, or frequency-domain features such as spectral energy). The set of these feature values ​​constitutes a multi-dimensional feature vector. When this multi-dimensional feature vector is projected onto a predefined multi-dimensional feature topology space, it is represented as a discrete coordinate point within that topology space, i.e., the "sample point" described in this invention. The discrete sample points collectively constitute the terrain of the probability distribution.

[0065] In the topological space formed by the sample points, the j-th discrete sample point through the source measure The i-th discrete sample point of the target measure The squared Euclidean distance between them is used to construct a geometric cost function to measure the physical transfer distance. In this process, discrete sample points of each source measure and discrete samples of each target measure participate in the transfer.

[0066] In satisfying and In the solution space formed by all joint probability distributions as constraints, the optimal transfer integral equation is used to solve for a specific mass transfer scheme matrix; this is used to evaluate the scalar value of the resistance to energy topological transfer from the vibrational physical field to the temperature physical field.

[0067] The optimal transport integral equation is defined as:

[0068]

[0069] in, denoted as the minimum cost of work done across physical fields, i.e., the minimum cost of optimal transmission from the source probability distribution measure to the target probability distribution measure; inf represents taking the infimum among all feasible solutions, i.e., selecting the minimum transmission cost; This represents the solution space comprised of all joint probability distributions that satisfy the marginal distribution constraint. The marginal distribution constraint means that, during the optimal transmission solution process, the marginal distributions of the constructed joint probability distribution on the source and target sides remain consistent with the preset source probability distribution measure and target probability distribution measure, respectively. In other words, during quality transfer, the total probability quality of each source measure sample point remains unchanged, while the total probability quality received by each target measure sample point remains consistent with its original distribution. This ensures that the entire transmission process satisfies the quality conservation constraint, reconstructing only the correspondence between different physical fields without altering the original probability distribution itself.

[0070] The obtained mass transfer scheme matrix As a dynamic alignment weight matrix, it is multiplied with a multidimensional tensor containing the features of the original high-frequency vibration signal to generate the coupled feature manifold graph that completely eliminates the time phase deviation. .

[0071] It should be noted that, in the specific embodiments of the present invention, the "multidimensional tensor containing the characteristics of the original high-frequency mechanical vibration signal" refers to a high-dimensional data structure generated by expanding and reconstructing the one-dimensional discrete-time sequence signal acquired at the bottom layer through a specific dimension, which is specifically used for manifold matrix operations by advanced neural dynamics operators. Specifically, this multidimensional tensor deeply integrates at least three dimensions of physical information in its topological structure: the first is the time dimension, which represents multiple sliding time window sequences extracted continuously according to the time series; the second is the spatial dimension, which represents the multi-channel heterogeneous sensor array nodes deployed on different mechanical degrees of freedom of the tested device; and the third is the feature dimension, which represents the time-domain statistical characteristics (such as root mean square, skewness, etc.) and frequency-domain energy characteristics calculated for the data locally extracted for each sliding time window. By structuring and reconstructing the original and messy one-dimensional time-series signals into this multi-dimensional tensor format, not only is the spatiotemporal coupling evolution law of the device's nonlinear dynamics fully preserved, but also the data entities across physical fields can perform precise Hadamard dot product tensor operations with the dynamic weight matrix solved by the optimal transmission algorithm under a unified geometric dimension. This provides a standard high-dimensional manifold input base that is size-compatible and phase-distortion-free for subsequent variable group convolution operators such as Lie groups.

[0072] It's important to understand that the Lie group isovariant neurodynamic operator in S3 is a mathematically rigorous tool for extracting geometric features. It requires the input data to be on a clean, phase-distortion-free geometric manifold. If the front end uses traditional "hard interpolation" to align heterogeneous data, it will inevitably create a large number of "frequency artifacts" on the manifold surface. These high-frequency artifacts may be just ordinary noise for traditional CNNs, but they are fatal to Lie group networks that have rotational symmetry awareness. The Lie group network will misinterpret these false phase distortions as real physical rigid body rotation signals, thus performing incorrect group matrix transformations within the network, ultimately causing the "computational collapse" of the entire isovariant feature extraction process. Therefore, S2 safely and stably unleashes the mathematical power of the Lie group isovariant operator on a geometric manifold base that has been cleaned of all frequency artifacts by S1 (optimal transmission) and provides absolute purity and strict alignment.

[0073] In industrial settings, rotating machinery (such as gearboxes and bearings) constantly operates under complex conditions of rapidly increasing speeds or sudden load changes. These changes cause nonlinear stretching or compression of the waveforms captured by sensors. Traditional deep learning models, lacking an understanding of the rigid body motion laws of the physical world, often misinterpret these changes as serious malfunctions. The nature of rotating machinery is governed by Lie group symmetry. Introducing Lie group-equalizing operators injects "physical mechanics genes" into deep learning. By forcibly setting equivalent constraint equations, the network establishes a defense mechanism: when the input signal undergoes physical distortion due to sudden changes in operating conditions, the deep hidden features of the operators are forcibly driven, resulting in a completely proportional mathematical inverse rotation. This design, using "mathematical isomorphic mapping" to directly counteract "physical environment interference," completely neutralizes the spatial drift caused by changing operating conditions.

[0074] S3: By constructing an equivariant neural dynamics operator based on Lie group theory, group convolution processing is performed on the coupled feature manifold graph to extract fatigue degradation feature vectors that remain symmetric under varying operating conditions.

[0075] For the spatial motion degrees of freedom of the core rotation axis system of the device under test, a mathematical Lie group is preset in the deep learning framework to describe continuous symmetric physical transformations; wherein, the mathematical Lie group is a set of group elements in group theory that describes the continuous rotation and translational rigid body motion of an object and has a topological geometric manifold structure.

[0076] Specifically, traditional deep learning feature extraction operators (such as conventional convolutional neural networks (CNNs) are typically built on flat Euclidean space, possessing only linear translation invariance and lacking analytical capabilities for the inherent spatial rotational symmetry of rotating machinery. When industrial equipment experiences nonlinear abrupt changes in operating speed or load, the multi-source signals collected by sensors undergo physical distortions in the time or frequency domains (such as waveform compression, stretching, or phase distortion). Traditional network models often misjudge such fluctuations as unknown serious faults, leading to a complete failure of cross-condition diagnosis and prediction capabilities. Lie groups, as a mathematical set combining group algebraic structure and smooth differential manifold structure, can rigorously and continuously describe rigid body motion transformations in the physical world. In specific engineering implementations, the system selects a matching set of Lie groups based on the spatial topological degrees of freedom of the core rotating components of the tested equipment: for example, for a single-axis rotor system continuously operating in a plane, a two-dimensional special orthogonal group is preset to constrain the physical transformations of planar rotation; for a multi-stage gearbox with complex spatial off-center loading modes, a three-dimensional special Euclidean group is preset to constrain the hybrid rigid body transformations involving three-dimensional translation and rotation. The conventional feature extraction network layers are reconstructed into group convolution kernels defined on the surface of this Lie group manifold. Through this underlying pre-set mechanism, the deep model is forcibly injected with symmetric physical prior knowledge of mechanics, enabling the high-dimensional hidden layer features inside the network to maintain strict isomorphism with the working condition distortion of the external physical input signal during forward dynamic evolution. This allows the algorithm to spontaneously neutralize and cancel the transient interference caused by the changing working conditions, ultimately decoupling and extracting pure fatigue degradation features with absolute generalization ability.

[0077] Based on the exponential mapping of Lie algebras, the convolution operation defined in Euclidean space is reconstructed into a group convolution kernel operator defined on the Lie group manifold, thereby achieving equivariant mapping of features under group action. Considering the highly curved and nonlinear characteristics of the Lie group manifold formed by the rigid body motion of the tested mechanical equipment (such as nonlinear rotation of the principal shaft), directly solving for the convolution kernel parameters in this continuous manifold space faces extremely high computational complexity. Therefore, Lie algebras are introduced as the local flat tangent space of the Lie group at the identity element, abstracting, reducing the dimension of complex geometric transformations, and locally linearizing them into feature vectors in the tangent space. Subsequently, the system calls the exponential mapping operator to precisely "wrap" the linear Lie algebra vector defined in the flat tangent space along the geodesic and map it back to the surface of the non-Euclidean geometric Lie group manifold, generating continuous real physical transformation group elements.

[0078] Furthermore, the equivariant mapping implemented in this embodiment mandates that the high-dimensional hidden layer features extracted from the network must maintain strict "isomorphic linkage" with the dynamic distortion of the external physical input signal. Specifically, when the device experiences physical spatial or phase distortion of the underlying input signal due to drastic changes in operating conditions, the equivariant mapping mechanism can drive the feature map within the network to spontaneously generate a completely equivalent proportional geometric transformation, ensuring that its internal topological structure remains absolutely stable. This mechanism prevents deep learning models from misjudging transient disturbances as disordered noise or unknown faults, instead transforming them into mathematically analyzable physical symmetry constraints. This allows for the accurate removal of transient disturbances under complex operating conditions with strong interference and multiple variables, uniquely identifying and outputting the device's underlying fatigue degradation features.

[0079] The network feature extraction layer is constructed using the group convolution kernel operator, which extends parameter sharing from Euclidean space to group manifold space.

[0080] Furthermore, the extraction of fatigue degradation feature vectors that remain symmetric and invariant under varying operating conditions includes:

[0081] The coupled feature manifold graph Input to contain group convolution kernel operator In deep learning networks.

[0082] Transformation operation for any physical disturbance representing a sudden change in operating conditions Force the group convolution kernel During network forward propagation, the equivalence space equation is satisfied:

[0083]

[0084] in, The mathematical Lie group is described In this context, a single group element represents a specific environmental change in rotational speed or load disturbance; it describes the specific state of the machine's current "variable operating condition" (e.g., the rotational speed suddenly increases from 1000 rpm to 1500 rpm). It is directly parameterized from the "physical changes in the environment." For example, if the rotational speed increases by 50%, in the two-dimensional rotational group SO(2), this... This can be directly quantified using a rotation angle θ (such as a scalar representing phase shift or waveform compression). It is the "disturbance input" imposed on the system by the external environment.

[0085] The input transform group action operator is the transform operation. The actual physical twisting and stretching matrix of the underlying coupled feature manifold graph spatial coordinate system; describing abrupt changes in operating conditions. The "actual physical distortion" caused to the acquired underlying data (coupled feature manifold graph) is rigorously derived through the exponential mapping of Lie algebras and the transformation of the underlying coordinate system. It is assumed that the input data is a two-dimensional grid (image or spectrogram). It is a specific 2×2 affine transformation matrix (such as a rotation matrix or scaling matrix). Multiplying this matrix by the coordinates of the underlying data realistically simulates the physical process of the waveform being stretched or distorted.

[0086] The output representation matrix is ​​the transformation operation. The isomorphic linear space transformation matrix mapped to the hidden layer feature dimension of the high-level neural network. It represents the "proportional mathematical inverse transformation" that the high-level features of the neural network need to make to counteract the distortion of the input. It is rigorously calculated using "group representation theory" in mathematics. After determining which Lie group (e.g., SO(2)) and the number of channels in the network are used, It can be precisely written as a series of rotation matrices with specific frequencies (usually in block diagonal matrix form). It is a set of mathematical rules pre-defined by the system, which force high-level features to rotate according to the rules of this matrix.

[0087] Based on this constraint equation, when the input signal undergoes physical drift due to changing operating conditions, the deep hidden features of the operator will be forced to undergo a completely isomorphic mathematical reverse rotation, thereby completely removing operating condition noise.

[0088] After global pooling compression, the output is a one-dimensional fatigue degradation feature vector, which is the absolute health state representation vector obtained by eliminating the spatial dimension of the anti-drift hidden layer high-dimensional tensor generated under equality constraints.

[0089] It's important to understand that, in order to transform "uncontrollable external disturbances" into "computable structural rules," thus preventing traditional models from mistakenly absorbing changes in operating conditions as degenerate features, symmetry constraints are pre-embedded in the model structure. This allows the network to automatically rearrange features through inherent geometric consistency, rather than passively learning compensation relationships based on data-driven approaches when faced with input changes. Essentially, this approach shifts the problem from "improving model robustness" to "eliminating unnecessary degrees of freedom," ensuring the model only expresses features in effective directions relevant to degradation, thereby enhancing feature stability and discriminative power. Simultaneously, by introducing strict mapping consistency in the high-dimensional feature space, representations under different states are comparable, avoiding feature space distortion caused by changes in input scale or morphology. This provides a stable and reliable feature foundation for subsequent state evolution modeling and lifetime prediction.

[0090] S4: Project the feature vector onto the information geometric manifold space, introduce geodesic constraints that satisfy the thermodynamic dissipation law, so that the model state evolves monotonically along the shortest physical path, and output the remaining service life prediction curve through Riemann arc length integral.

[0091] The process of projecting the feature vector onto the information geometric manifold space essentially expands the current degradation state of the device from a deterministic description to a probabilistic description. Specifically, the fatigue degradation feature vector is used as input, and global weight parameters are applied... The deep learning network generates a family of conditional probability distributions representing the uncertainty in the prediction of the remaining lifetime of equipment. This approach moves beyond limiting lifetime prediction to a single numerical value, instead obtaining complete probability distribution information. This provides a foundation for constructing an information geometric manifold, enabling feature representations to reflect both state changes and the confidence structure of the prediction results. Consequently, it enhances the model's ability to express complex degradation processes and its stability.

[0092] in, Global weight parameters represent the set of real numbers representing the weight matrices and bias vectors of all connection layers; The true degradation label of a given device is generated through feedforward network mapping. The conditional probability function output is given.

[0093] Calculate the log-likelihood function of the family of conditional probability distributions with respect to the global weight parameters. The second-order partial derivatives are calculated, and their expected values ​​are obtained to derive the Fisher information matrix. Specifically, the calculation process is as follows: First, the forward mapping of the deep learning prediction network constructs a family of conditional probability distribution densities representing the uncertainty of lifetime prediction. To transform the probability product into a linear superposition and facilitate optimization, the system takes the logarithm of the distribution family to obtain the log-likelihood function, and then calculates the log-likelihood function with respect to the network's global weight parameters. The second-order partial derivative (i.e., the Hessian matrix) is the parameter space in the context of information geometry. In the face of a single, specific observation data, this second-order partial derivative directly characterizes the parameter space. The "local curvature" at that time (i.e., observational information). However, due to the single industrial data collection... Inevitably, it carries transient physical random noise. Directly utilizing this local curvature would lead to extreme oscillations in the model's feature space and a lack of global representativeness. Therefore, this embodiment performs a crucial "mathematical expectation" mapping step. By calculating the mathematical expectation of the aforementioned second-order partial derivatives over the entire probability distribution of the real data, the model is thoroughly smoothed and eliminated from the underlying mathematical integral, eliminating specific random sampling points. The random variable interference introduced by this. After the aggregation and transformation of this expectation operation, the random matrix that was originally fluctuating due to local noise interference strictly converges to a matrix that only relates to the global parameters. The associated absolute deterministic matrix—that is, the derived Fisher information matrix. In this embodiment, the Fisher information matrix is ​​equivalent to the Riemannian metric tensor, which precisely quantifies the global geometric curvature and prediction confidence of the model in different parameter regions. It completely replaces the flat "identity matrix" in traditional Euclidean space, becoming the Riemannian metric tensor that reshapes the curved topography of the parameter space. This embodiment, through a mechanism of "first calculating the local second derivative, then the global expectation," elevates the flat, blindly fitted Euclidean weight space of the traditional model into an information manifold space with undulating terrain features, thus providing a geometric and topological foundation for subsequently applying a geodesic penalty term based on the physical thermodynamic dissipation law.

[0094] Furthermore, using the Fisher information matrix... The identity matrix in the alternative Euclidean space is used as the Riemannian metric tensor to reconstruct the feature prediction mapping dimension of the deep learning network, generating a non-flat information geometric manifold space, and incorporating the fatigue-degraded feature vectors. As the origin of the dynamic coordinate system, it is mapped into the information geometric manifold space.

[0095] The introduction of geodesic constraints that satisfy the thermodynamic dissipation law, and the output of the remaining service life prediction curve through Riemann arc length integration, includes:

[0096] The fatigue degradation feature vector is used as the origin of the state mapping and projected onto the information geometric manifold space. The implementation process is as follows: First, the output of the preceding isovariant operator is received. One-dimensional tensor This tensor represents the pure degradation increment after removing operational disturbances. Considering the dimensionality inconsistency between the feature representation space and the predicted evolution probability space, a nonlinear projection operator based on a multilayer perceptron (MLP) structure is constructed. The projection operator uses a preset weight matrix to... A linear weighted and nonlinear activation transformation is performed to map the features from a flat Euclidean feature coordinate system to a manifold coordinate representation conforming to the Fisher information metric structure. Subsequently, the mapped high-dimensional features are forcibly assigned to the initial hidden state boundary conditions of the backend neural differential equation prediction module. 0). At the mathematical topological level, this process is equivalent to precisely anchoring a geometric origin representing the current true damage state of the device on a curved probability manifold measured by the Fisher information matrix. Using this origin as the baseline for path planning, subsequent steps can utilize the manifold geodesic algorithm to calculate the optimal degradation trajectory constrained by the law of physical dissipation. Through this projection and anchoring mechanism, this embodiment eliminates the accumulated error caused by initial state drift in traditional prediction algorithms at the digital level, ensuring that the remaining lifetime prediction results always converge to the true physical degradation path.

[0097] Within the information geometric manifold space, by finding the optimal monotonic path of analytical state evolution where the curve arc length becomes zero under the Fisher-Riemann metric, a theoretically shortest smooth curve connecting the current healthy point and the endpoint of complete physical failure is solved, defined as the ideal physical dissipation geodesic. The specific calculation and implementation process is as follows: First, in the manifold space defined by the Fisher Information Matrix (FIM), a "current healthy point" representing the true degradation state at the current moment is marked. And a pre-defined "complete physical failure endpoint" representing the probability boundary of irreversible damage to the device. Subsequently, abandoning the traditional linear Euclidean distance extrapolation method in Euclidean space, we instead used the Fisher information matrix as the Riemannian metric tensor to construct an integral functional of the trajectory arc length connecting the two points. Based on the physical fact that equipment fatigue wear is an irreversible entropy increase process in thermodynamics, an absolute "optimal monotonicity" mathematical constraint was imposed on this evolution path, that is, the state evolution vector must strictly point to the failure direction of system entropy increase in its time derivative, prohibiting state reversal and self-healing phenomena that violate physical common sense.

[0098] Under the aforementioned monotonic constraints, the system utilizes the variational method to solve for the functional extrema. By constructing and solving the Euler-Lagrange differential equation system, the system seeks to find the first variation that makes the arc length of the trajectory strictly zero. The analytical solution is obtained. This variational extremum condition, in a physical sense, defines a theoretically shortest smooth curve, namely the "ideal physical dissipation geodesic". This geodesic perfectly fits the underlying curvature characteristics of complex non-Euclidean probability manifolds, representing the true degradation trajectory of the device under test under current operating boundary conditions, following the path of minimum energy dissipation resistance, which best conforms to the objective physical evolution law. Thus, it provides a closed-loop analytical basis for extremely accurate calculation of remaining useful life (RUL).

[0099] Simultaneously, the fatigue degradation feature vector is tracked by continuous time step. Network evolution trajectory as depicted The specific implementation is as follows: taking the current moment... The origin of the anchored manifold state Set a fixed continuous time step as the initial boundary condition. (Or an adaptive step-size differential solver can be used). Within each time step, the prediction network calculates the nonlinear degradation increment for the next time step based on its internally learned dynamic weight matrix, and then recursively updates or integrates to obtain the hidden variable state for the next time step. With time step As the network progresses continuously over time, a series of discrete latent variable points, derived from the network's actual outputs, are sequentially calculated and generated. From a topological perspective within the information geometric manifold space, spatially connecting these time-series-arranged discrete latent variable points reveals a high-dimensional continuous evolution curve. This curve is the so-called "actual network extrapolation trajectory." Because deep learning models are essentially black-box systems that rely on pure data fitting, they inevitably accumulate gradient errors or are affected by uncertainties in model parameters during long-term extrapolation. This causes the actual extrapolation trajectory to easily generate non-physical oscillations or divergences, gradually deviating from objective thermodynamic evolution laws.

[0100] Calculate the network inference trajectory With the ideal physical dissipation geodesic Riemannian geometric distance between The "Ideal Physical Dissipation Geodesic" is mathematically and geometrically represented as the absolute shortest smooth curve in a high-dimensional probability manifold space with the Fisher information matrix as the Riemannian metric tensor, strictly connecting the origin of the current healthy state and the endpoint of complete physical failure, making the first variation of the arc length zero. At the physical mechanism level, this geodesic profoundly reflects the "principle of least action" and the "law of irreversible entropy increase" in thermodynamics. It represents the "absolute physical evolution standard path" for the tested mechanical equipment under given boundary conditions, characterized by minimal energy dissipation resistance, naturalness, and conformity to the objective microscopic material fatigue laws.

[0101] The "Network Inference Trajectory" is characterized as follows: a set of discrete state points actually calculated by a deep learning time series prediction operator (such as a dynamic recursive graph or neural constant differential equation) at continuous time steps, relying on its internally learned dynamic weight matrix, through layer-by-layer iterative mapping of the latent variable space. In terms of algorithmic attributes, this trajectory represents the "pure data-driven fitting path" of the neural network. Since deep learning models are essentially highly nonlinear mathematical black boxes, when extrapolating long time series without strong external constraints, they are highly susceptible to interference from factors such as model parameter uncertainty and gradient accumulation error. This can cause the inference trajectory to diverge, oscillate, or spatially distort in the manifold space, deviating from the common sense of monotonic degradation.

[0102] In a specific embodiment of the present invention, there is inevitably a significant difference in geometric dimension between the network's extrapolated trajectory and the ideal physical dissipative geodesic. The physical origin of this difference lies in the fact that traditional deep learning temporal networks (such as RNNs / ODEs) are essentially high-degree-of-freedom parameter fitting systems lacking physical constraints. In the process of extrapolating future time series without real labels, they inevitably experience spatial drift due to gradient accumulation errors and data uncertainty, even outputting failure trajectories that violate the thermodynamic law of monotonic entropy increase. In contrast, the ideal physical dissipative geodesic is strictly derived from Fisher's manifold curvature and variational extreme conditions, characterizing the physical law of minimizing energy dissipation that the evolution of devices must follow. The difference between the two represents the deviation between "purely data-driven blindness" and "objective physical inevitability." By constructing the Riemannian difference between the actual extrapolated trajectory and the theoretical physical geodesic as the target loss function, this embodiment essentially forces the prior knowledge boundary of physical mechanics into the gradient optimization path of the network's backpropagation. During training, to minimize this geometric penalty, the weight matrix of the hidden layers of the network is forced to compress its redundant nonlinear degrees of freedom, compelling its dynamic behavior to converge to the laws of thermodynamic evolution. This mechanism not only achieves precise correction of the network's "deviation from physical common sense" in its predicted trajectory, but also endows the black-box model with rigorous physical interpretability and extremely high robustness in long-sequence prediction.

[0103] A nonlinear geodesic penalty is set based on the irreversible physical law of thermodynamic dissipation: based on Riemannian geometric distance. Constructing a geodesic penalty term for backpropagation Defined as:

[0104]

[0105] in, The penalty increase / decrease coefficient is a pre-defined, significantly scaled scalar constant used to exponentially dampen the accuracy of the violation evolution. When the network's inferred trajectory deviates from the ideal physical dissipation path, the Riemannian geometric distance between them expands over time, i.e., the time derivative. When the value is greater than zero, the penalty term is immediately activated, forcing the state evolution trajectory to slide absolutely downwards along the manifold surface.

[0106] The Riemann arc length integral is calculated along the state evolution trajectory constrained by the geodesic penalty term, and the remaining lifetime prediction curve with absolute monotonically decreasing is output.

[0107] Specifically, when processing industrial acquisition data with strong transient noise, traditional deep learning networks often output "abnormal lifespan recovery" (i.e., the equipment exhibits a self-healing illusion) that violates physical common sense due to local overfitting. To eliminate this technical bias, this embodiment constructs a loss function with an "asymmetric extremum penalty" property during the backpropagation stage. By introducing a significantly scaled scalar constant, the system grants the monotonicity constraint a "veto power." Once the network's predicted trajectory is induced by data perturbation and attempts to climb upwards (i.e., an abnormal recovery occurs with the time derivative greater than zero), the geodesic penalty term will be activated exponentially. This overwhelming gradient pull constructs a strict physical boundary, forcibly cutting off the neural network's degrees of freedom in the invalid solution space, forcing its predicted trajectory to slide absolutely and uniquely downwards along the information manifold surface. Furthermore, performing Riemann arc length integration along the state evolution trajectory tightly bound by the above penalty term is to complete a rigorous mapping from "abstract geometric distance" to "concrete physical time." This integration process, within the manifold metric space, precisely converts the accumulated path data representing the equipment's degradation state into the remaining lifespan in the real world. This ensures that the system's final output prediction curve possesses the mathematical property of "absolute monotonically decreasing," fundamentally eliminating decision-making interference and false alarm anxiety caused by frequent jumps in prediction results in industrial settings.

[0108] S5: Develop an equipment maintenance plan based on the predicted curve to achieve predictive maintenance of the equipment.

[0109] The remaining useful life prediction curve, which has a monotonically decreasing characteristic, generated by integration, is sent in real time to the distributed programmable logic controller in the industrial field equipment control layer.

[0110] Within the distributed programmable logic controller, a dynamic physical life safety threshold is set that is deeply coupled with the current device's ultimate physical fatigue fracture resistance process. ( It represents the dynamic physical life safety threshold, which is derived from a hard alarm scalar calculated based on the minimum remaining time of mechanical fracture limit allowed by the on-site process.

[0111] When real-time polling monitoring detects that the value of the remaining useful life prediction curve is approaching and falling below the dynamic physical life safety threshold... When this occurs, a maintenance command is triggered. For example, the distributed programmable logic controller directly triggers the highest priority hardware interrupt. This forces the inverter drive mechanism of the control equipment to execute physical load reduction protection procedures, including speed reduction or direct disconnection of the spindle power source.

[0112] This embodiment also provides a device prediction and maintenance system based on multi-source data fusion and deep learning, including: a data acquisition unit, which acquires multi-source asynchronous sensor data during device operation and maps heterogeneous physical signals with different sampling frequencies that are not equally distributed on the time scale into a probability distribution measure within a unified time window.

[0113] The computational unit calculates the Wasserstein distance between various probability distribution measures based on optimal transport theory, obtains the minimum work cost matrix across physical fields as dynamic weights, and generates an aligned coupled feature manifold graph.

[0114] The extraction unit performs group convolution processing on the coupled feature manifold graph by constructing an equivariant neural dynamics operator based on Lie group theory, and extracts fatigue degradation feature vectors that remain symmetric under varying operating conditions.

[0115] The evolution unit projects the feature vector into the information geometric manifold space, introduces geodesic constraints that satisfy the thermodynamic dissipation law, and makes the model state monotonically evolve along the shortest physical path. The remaining service life prediction curve is output through Riemann arc length integral.

[0116] The maintenance unit formulates equipment maintenance plans based on the predicted curves to achieve predictive maintenance of the equipment.

[0117] This embodiment also provides a computer device applicable to the device predictive maintenance method based on multi-source data fusion and deep learning, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the device predictive maintenance method based on multi-source data fusion and deep learning as proposed in the above embodiment.

[0118] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0119] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the device predictive maintenance method based on multi-source data fusion and deep learning as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0120] In summary, this invention achieves high-precision and robust predictive maintenance of complex industrial equipment by: mapping multi-source heterogeneous sensor data from the time domain to the probability measure space, utilizing optimal transmission theory to achieve unified alignment of cross-physical field data, and eliminating phase distortion caused by sampling differences; based on this, introducing an equivariant neural dynamics operator based on Lie group theory to model equipment operating condition changes as continuous symmetric transformations, and using a group convolution mechanism to achieve stable decoupling and invariance extraction of features under complex operating condition disturbances; further constructing an information geometric manifold space, using the Fisher information matrix to characterize the non-Euclidean geometry of the model parameter space, and introducing geodesic constraints that satisfy thermodynamic dissipation laws to force the model state to evolve monotonically along the shortest physical path; finally, obtaining strictly monotonically decreasing remaining service life prediction results through Riemann arc length integral, and combining dynamic safety thresholds to achieve closed-loop linkage between prediction results and equipment control strategies.

[0121] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A device predictive maintenance method based on multi-source data fusion and deep learning, characterized in that: This includes acquiring multi-source asynchronous sensor data during equipment operation and mapping heterogeneous physical signals with different sampling frequencies that result in non-equidistant distributions on the time scale into a probability distribution measure within a unified time window; Based on the optimal transport theory, the Wasserstein distance between various probability distribution measures is calculated, and the minimum work cost matrix across physical fields is obtained as a dynamic weight to generate an aligned coupled feature manifold graph. By constructing an equivariant neural dynamics operator based on Lie group theory, the coupled feature manifold graph is subjected to group convolution processing to extract fatigue degradation feature vectors that remain symmetric under varying operating conditions. The feature vector is projected onto the information geometric manifold space, and geodesic constraints that satisfy the thermodynamic dissipation law are introduced to make the model state monotonically evolve along the shortest physical path. The remaining lifetime prediction curve is output through Riemann arc length integral. Based on the predicted curve, a maintenance plan is developed to achieve predictive maintenance of the equipment.

2. The equipment prediction and maintenance method based on multi-source data fusion and deep learning as described in claim 1, characterized in that: The heterogeneous physical signals include high-frequency vibration signals and low-frequency temperature signals from the synchronous acquisition device; The probability distribution measure includes the source probability distribution measure converted from the high-frequency vibration signal. The conversion of the low-frequency temperature signal is a measure of the target probability distribution. ; The source probability distribution measure is a probability density function generated by estimating the kernel density of the high-frequency vibration signal amplitude within a sliding time window; the target probability distribution measure is a probability density function generated by estimating the kernel density of the low-frequency temperature signal within a time window. In the topological space formed by the sample points, a geometric cost function for measuring the physical transfer distance is constructed by mapping the discrete sample points of the source measure and the discrete sample points of the target measure through the square of the Euclidean distance. In satisfying and In the solution space formed by all joint probability distributions that serve as constraints, the specific mass transfer scheme matrix is ​​solved using the optimal transfer integral equation. The mass transfer scheme matrix obtained by the solution is used as a dynamic alignment weight matrix, and a dot product operation is performed with a multidimensional tensor containing the features of the original high-frequency vibration signal to generate the coupled feature manifold graph that completely eliminates the time phase deviation.

3. The equipment prediction and maintenance method based on multi-source data fusion and deep learning as described in claim 2, characterized in that: The construction of the isovariant neurodynamic operator based on Lie group theory includes, For the spatial motion degrees of freedom of the core rotation axis system of the device under test, a mathematical Lie group is preset in the deep learning framework to describe continuous symmetric physical transformations; wherein, the mathematical Lie group is a set of group elements in group theory that describes the continuous rotation and translational rigid body motion of an object and has a topological geometric manifold structure. Based on the exponential mapping of Lie algebras, the convolution operation defined in Euclidean space is reconstructed into a group convolution kernel operator defined on the Lie group manifold, so as to realize the equivariant mapping of features under group action. The network feature extraction layer is constructed using the group convolution kernel operator, which extends parameter sharing from Euclidean space to group manifold space.

4. The equipment prediction and maintenance method based on multi-source data fusion and deep learning as described in claim 3, characterized in that: The extraction of fatigue degradation feature vectors that remain symmetrical under varying operating conditions includes, The coupled feature manifold graph Input to contain group convolution kernel operator In deep learning networks; Transformation operation for any physical disturbance representing a sudden change in operating conditions Force the group convolution kernel During network forward propagation, the equivalence space equation is satisfied: in, The mathematical Lie group is described In this context, a single group element represents the specific environmental rotational speed or load disturbance transformation; The input transform group action operator is the transform operation. The actual physical twisting and stretching matrix of the spatial coordinate system of the underlying coupled feature manifold graph; The output representation matrix is ​​the transformation operation. The isomorphic linear space transformation matrix mapped onto the feature dimension of the hidden layer of a high-level neural network; After global pooling compression, a one-dimensional fatigue degradation feature vector is output.

5. The equipment prediction and maintenance method based on multi-source data fusion and deep learning as described in claim 4, characterized in that: Projecting the feature vector onto the information geometric manifold space includes taking the fatigue degradation feature vector as input and utilizing global weight parameters. The deep learning network generates a family of conditional probability distributions representing the uncertainty in the prediction of the remaining lifetime of equipment. ; in, Global weight parameters represent the set of real numbers representing the weight matrices and bias vectors of all connection layers; The true degradation label of a given device is generated through feedforward network mapping. The conditional probability function output is given below; Calculate the log-likelihood function of the family of conditional probability distributions with respect to the global weight parameters. The second-order partial derivatives are calculated, and their expected values ​​are obtained to derive the Fisher information matrix. ; With the Fisher information matrix The identity matrix in the alternative Euclidean space is used as the Riemannian metric tensor to reconstruct the feature prediction mapping dimension of the deep learning network, generating a non-flat information geometric manifold space, and incorporating the fatigue-degraded feature vectors. As the origin of the dynamic coordinate system, it is mapped into the information geometric manifold space.

6. The equipment predictive maintenance method based on multi-source data fusion and deep learning as described in claim 5, characterized in that: The introduction of geodesic constraints that satisfy the thermodynamic dissipation law, and the output of the remaining lifetime prediction curve through Riemann arc length integration, include... The fatigue degradation feature vector is used as the origin of the state mapping and projected onto the information geometric manifold space; Within the information geometric manifold space, by finding the optimal monotonic path of analytical state evolution where the curve arc length becomes zero under the Fisher-Riemann metric, a theoretically shortest smooth curve connecting the current healthy point and the endpoint of complete physical failure is solved, defined as the ideal physical dissipation geodesic. ; Simultaneously, the fatigue degradation feature vector is tracked by continuous time step. Network evolution trajectory as depicted ; Calculate the network inference trajectory With the ideal physical dissipation geodesic Riemannian geometric distance between ; A nonlinear geodesic penalty is set based on the irreversible physical law of thermodynamic dissipation: based on Riemannian geometric distance. Constructing a geodesic penalty term for backpropagation Defined as: in, The penalty increase / decrease coefficient is a pre-set, large-scale scaling scalar constant used to exponentially dampen the amplification of the accuracy of the violation evolution. The Riemann arc length integral is calculated along the state evolution trajectory constrained by the geodesic penalty term, and the remaining lifetime prediction curve with absolute monotonically decreasing is output.

7. The equipment prediction and maintenance method based on multi-source data fusion and deep learning as described in claim 6, characterized in that: The predictive maintenance includes setting a dynamic physical life safety threshold that is deeply coupled with the current equipment's ultimate physical fatigue fracture resistance process. When the real-time polling program detects that the value of the remaining useful life prediction curve reaches the dynamic physical life safety threshold, a maintenance command is triggered.

8. A device predictive maintenance system based on multi-source data fusion and deep learning, based on the device predictive maintenance method based on multi-source data fusion and deep learning as described in any one of claims 1 to 7, characterized in that: This includes a data acquisition unit that acquires multi-source asynchronous sensor data during equipment operation and maps heterogeneous physical signals with different sampling frequencies that result in non-equidistant distribution on a time scale into a probability distribution measure within a unified time window. The computational unit calculates the Wasserstein distance between various probability distribution measures based on optimal transport theory, obtains the minimum work cost matrix across physical fields as dynamic weights, and generates an aligned coupled feature manifold graph. The extraction unit extracts fatigue degradation feature vectors that remain symmetric under varying operating conditions by constructing an equivariant neural dynamics operator based on Lie group theory and performing group convolution on the coupled feature manifold. The evolution unit projects the feature vectors into the information geometric manifold space, introduces geodesic constraints that satisfy the thermodynamic dissipation law, and makes the model state monotonically evolve along the shortest physical path. The remaining lifetime prediction curve is output through Riemann arc length integral. The maintenance unit formulates equipment maintenance plans based on the predicted curves to achieve predictive maintenance of the equipment.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the device prediction and maintenance method based on multi-source data fusion and deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the device prediction and maintenance method based on multi-source data fusion and deep learning as described in any one of claims 1 to 7.

Images