Equipment state monitoring and analysis evaluation method and system based on big data

Abstract

The invention relates to the field of equipment state monitoring in industrial Internet of Things, and discloses an equipment state monitoring, analysis and evaluation method and system based on big data, and the method comprises the steps: carrying out the adaptation of a multi-source heterogeneous data protocol, and carrying out the cleaning of a dynamic mask, and generating a standardized data stream; constructing a dynamic hypergraph of an embedded constraint equation based on physical topology; combining incremental tensor decomposition with manifold constraint to update a core tensor; abnormal association is positioned based on singular value distribution and a hyperedge propagation algorithm; cross-equipment model migration is realized through topological optimal transmission and knowledge distillation, and a target equipment evaluation model is generated; the system comprises a data preprocessing module, a hypergraph modeling module, a tensor analysis module, a state evaluation module, a transfer learning module and a dynamic tuning module. According to the method, through multi-source data dynamic cleaning, physical constraint hypergraph modeling, incremental tensor decomposition and manifold constraint, and in combination with an abnormal positioning closed loop and cross-equipment topology migration, equipment state monitoring and rapid model adaptation are realized.

Description

Technical Field

[0001] The present invention relates to the field of equipment status monitoring in the industrial Internet of Things, and specifically to a method and system for equipment status monitoring and analysis evaluation based on big data. Background Art

[0002] In the field of industrial equipment condition monitoring, traditional methods rely on a single type of sensor data or simple fusion techniques for anomaly detection. These methods struggle to effectively handle the significant differences in protocol formats, sampling frequencies, and noise levels among heterogeneous sensor data from multiple sources. Raw data, without deep cleaning and standardization, is prone to introducing noise interference and anomalies, resulting in insufficient reliability for subsequent modeling and analysis. Furthermore, existing models often utilize purely data-driven graph neural networks or statistical methods, lacking explicit embedding of the physical laws of the equipment. This disconnects the state representation from the actual physical mechanisms, resulting in weak interpretability and limited adaptability to operating conditions.

[0003] Furthermore, most methods employ a static modeling framework that is unable to dynamically respond to data distribution drift caused by changes in operating conditions and component degradation during equipment operation, resulting in significant degradation of model performance during long-term monitoring. In cross-device knowledge migration scenarios, existing technologies struggle to quantify the similarity of topological structures between devices and ignore the consistent alignment of physical constraints, resulting in insufficient generalization of the migration model on the target device. More critically, existing systems generally lack closed-loop optimization mechanisms, making it impossible to reversely adjust model parameters based on real-time monitoring results, making it difficult to achieve adaptive dynamic evolution and continuous performance improvement.

[0004] Therefore, the present invention proposes a device status monitoring and analysis evaluation method and system based on big data to address the shortcomings of the existing technology. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a method and system for equipment status monitoring and analysis evaluation based on big data, which solves the problems of difficulty in cleaning multi-source data in industrial equipment monitoring, disconnection between physical laws and data models, insufficient adaptability to dynamic working conditions, and low efficiency of cross-equipment knowledge transfer.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a device status monitoring and analysis and evaluation method based on big data, the method comprising the following steps:

[0007] S1, perform protocol adaptation and dynamic mask cleaning on multi-source heterogeneous sensor data to generate standardized data stream;

[0008] S2. Based on the standardized data flow and the physical topological relationship of the device, a dynamic hypergraph topological structure is constructed, and the device operation physical constraint equation is embedded in the hypergraph convolution operation;

[0009] S3. performing incremental tensor decomposition on the dynamic hypergraph topology, updating the core tensor based on a sliding window, and maintaining physical law consistency through manifold space constraints;

[0010] S4. Calculate a device anomaly score based on the singular value distribution of the core tensor, and locate the associated hyperedge that caused the anomaly based on a hyperedge influence propagation algorithm model;

[0011] S5. Based on the topological properties of the associated hyperedges, similarities between devices are measured through optimal transmission of hypergraph topology, and cross-device analysis model migration is achieved in combination with the knowledge distillation loss function to generate a state assessment model for the target device.

[0012] Preferably, in step S1, the step of performing protocol adaptation and dynamic mask cleaning on multi-source heterogeneous sensor data to generate a standardized data stream includes:

[0013] Mask generation: Generate a dynamic binary mask matrix M based on the sliding median filter and the historical health status of the device mask ∈{0,1} N×D , and mask the abnormal time period or sensor channel, where N represents the number of sensors and D represents the dimension of the original data;

[0014] Noise suppression: the original input data Apply sliding median filtering on each channel to remove impulse noise and generate element-by-element multiplication fusion filtering results;

[0015] Data cleaning: Based on the element-by-element multiplication operation, fusion filtering results, data cleaning model and dynamic binary mask matrix, multi-source heterogeneous sensor data is cleaned to generate cleaned data. The data cleaning model is:

[0016] X clean =MedianFilter(X raw )⊙M mask ;

[0017] in, is the original input data; ⊙ represents the element-by-element multiplication operation, M mask is a dynamic binary mask matrix;

[0018] Using the joint feature matrix F edge Performing time-frequency domain joint feature generation on the cleaned data to obtain the standardized data stream;

[0019] The joint feature matrix F edge By the short-time Fourier transform feature F STF1 and continuous wavelet transform feature F CWT It is obtained by splicing along the feature dimension, and the splicing satisfies the relational expression: F edge=F STFT ⊕F CWT ;

[0020] Among them, ⊕ represents the feature concatenation operation; F STFT and F CWT The calculation is performed by intercepting the signal segments through the sliding window.

[0021] Preferably, in step S2, the steps of constructing a dynamic hypergraph topology structure based on the standardized data flow and the physical topology relationship of the device, and embedding the device operation physical constraint equation into the hypergraph convolution operation include:

[0022] The dynamic hypergraph topology is defined as a hypergraph tensor in, is the set of sensor nodes; ε t is a set of hyperedges and each hyperedge connects at least three sensor nodes; is the multimodal interaction weight tensor, K represents the number of sensor types, and M represents the number of working modes;

[0023] The physical constraint equations for the device operation are generated by discretizing the partial differential equations of the physical laws of the device and embedded into the hypergraph convolution operation. The formula for the convolution operation is:

[0024]

[0025] Among them, H (l+1) is the (l+1)th layer vertex feature matrix, represents the vertex feature matrix of the lth layer, N represents the number of sensors, d l Represents the original data dimension in the vertex feature matrix of the lth layer;

[0026] is the learnable parameter matrix; is the physical constraint term; σ is the nonlinear activation function; ×1 and ×2 represent the 1-module product and 2-module product of the tensor, respectively.

[0027] Preferably, the physical constraint Generates conservation or kinetic equations by discretizing the device operation, including heat conduction equations or the finite difference form of the Navier-Stokes equations;

[0028] In the hypergraph convolution operation, the 1-module product × 1 represents the weight tensor and the vertex feature matrix H (l) The row-by-row interaction, 2 modular product × 2 represents the weight tensor With the parameter matrix Θ (l)The column-by-column interaction is mapped to the high-dimensional space through the nonlinear activation function σ.

[0029] Preferably, the step S3 includes:

[0030] Intercepting the time series feature tensor through the sliding window And construct an optimization objective function, where N is the number of sensors, D′ is the feature dimension, t is the sliding window, and Δ is the window length. The optimization objective function is:

[0031]

[0032] The constraints are and in, is the core tensor; U (n) is a factor matrix; is the discretization matrix of the physical equation; R max is the preset rank upper limit;

[0033] The core tensor is updated based on the sliding window, and the parameter update is achieved through the manifold space constraint model and Stiefel manifold projection. The manifold space constraint model is:

[0034]

[0035] Where η is the learning rate; The objective function is The gradient, is the factor matrix of the nth mode at the kth iteration; the QR decomposition operation projects the gradient updated matrix onto the Stiefel manifold projection, which is

[0036] Preferably, the step S4 includes:

[0037] Through Core Tensor The singular value and the abnormality score calculation formula are used to perform distribution calculation to obtain the device abnormality score. The abnormality score calculation formula is:

[0038]

[0039] Among them, S anomaly The value for the device anomaly score, represents the slice of the core tensor at the rth mode; is the physical constraint factor matrix; σ r is the rth singular value; is the mean of the singular values;

[0040] Based on the hyperedge influence propagation algorithm model, the gradient contribution of the hyperedge weight to the anomaly score is calculated, and the associated hyperedge that causes the anomaly is determined in combination with the importance measurement of the sensor node. The hyperedge influence propagation algorithm model is:

[0041]

[0042] Among them, I(e j ) is the associated hyperedge that caused the exception, Represents the hyperedge e j The weight parameter of PageRank(v i ) is the sensor node v i The global topological importance score of I(e j )>θ I The hyperedges of v constitute the associated hyperedge set, i is the hyperedge e j The i-th sensor node contained in belongs to the sensor node set θ I is the dynamically adjusted abnormal threshold; is the partial derivative.

[0043] Preferably, in step S5:

[0044] The optimal transmission of the hypergraph topology is achieved by solving the source device hypergraph Hypergraph with target device The optimal transmission matrix between To achieve this, the transmission distance is defined as:

[0045]

[0046] Where S represents the hypergraph topology of the source device; T represents the target device; Π(μ S ,μ T ) satisfies the marginal distribution μ S ,μ T The set of transmission matrices; and All are hypergraph topology embedding functions;

[0047] The knowledge distillation loss function is implemented by constraining the core tensor distribution consistency between the source device and the target device, and is defined as:

[0048]

[0049] in, is the optimal transmission matrix; Denotes the i-th and j-th core tensor slices of the source and destination devices, respectively; D KL is the Kullback-Leibler divergence.

[0050] Preferably, the total loss function for cross-device analysis model migration is a weighted combination of knowledge distillation loss and optimal transmission loss, and the knowledge distillation loss function is:

[0051]

[0052] in, is the total loss function of transfer learning; is the optimal transmission loss term, and are the node embedding vectors of the source device and the target device respectively; α∈(0,1) is the balancing weight coefficient.

[0053] The present invention also provides a big data-based equipment status monitoring and analysis and evaluation system, which includes the following modules:

[0054] The data preprocessing module performs protocol conversion and dynamic mask cleaning of multi-source heterogeneous sensor data, and outputs a standardized data stream containing joint features in the time and frequency domains;

[0055] The hypergraph modeling module builds a dynamic hypergraph based on the physical topology of devices and standardized data flows, and embeds the physical constraint equations of device operation into the hypergraph convolution;

[0056] Tensor analysis module, which performs incremental tensor decomposition on dynamic hypergraphs and updates core tensors through manifold space constraints;

[0057] The status assessment module calculates the device anomaly score based on the core tensor singular value distribution and locates the associated hyperedges in combination with the hyperedge influence propagation algorithm;

[0058] The transfer learning module realizes cross-device knowledge transfer and generates a target device evaluation model based on the topological properties of the associated hyperedges and the optimal transmission distance;

[0059] The dynamic tuning module adjusts the hypergraph weight parameters in real time based on anomaly scoring and transfer learning results to drive the dynamic evolution of the system.

[0060] Preferably, the dynamic tuning module is specifically configured to perform the following operations:

[0061] Based on the anomaly score and the associated hyperedge identifier, the hyperedge weight tensor is updated through a gradient-driven mechanism;

[0062] Fusion of transfer learning loss and core tensor features to generate topology evolution gradient signals;

[0063] Adaptively adjust the physical constraint discretization parameters of the hypergraph modeling module;

[0064] Hypergraph weights and physical constraints are updated synchronously through cross-module interfaces to form a closed-loop optimization.

[0065] The present invention provides a method and system for equipment status monitoring, analysis and evaluation based on big data. It has the following beneficial effects:

[0066] 1. This invention effectively addresses the issues of heterogeneous data formats and high noise interference from multi-source sensor data through protocol conversion and dynamic mask cleaning technology, generating a standardized data stream containing joint time-frequency domain features. The dynamic masking mechanism, combined with device health status, adaptively shields abnormal data points, ensuring the integrity and reliability of data input to subsequent modules and providing a high-quality data foundation for device status analysis.

[0067] 2. This invention constructs a hypergraph structure based on the physical topological relationships of devices, embedding discretized device operating physics equations into hypergraph convolution operations, achieving dual constraints of data-driven and physical laws. This design overcomes the limitation of traditional graph models that rely solely on data associations, making the state representation more closely aligned with the actual operating characteristics of the device, significantly improving the model's interpretability and generalization capabilities.

[0068] 3. This invention uses a sliding window mechanism to achieve incremental tensor decomposition of dynamic hypergraphs, combined with Stiefel manifold projection to maintain parameter orthogonality and effectively avoid modal aliasing. Physical constraints force the decomposition results to conform to conservation equations, ensuring the physical consistency of the core tensor during temporal evolution and enhancing adaptability to dynamic changes in equipment operating conditions.

[0069] 4. This invention uses the core tensor singular value distribution to quantify anomaly scores and combines gradient contributions with topological importance in a hyperedge influence propagation algorithm to achieve multi-dimensional localization of anomaly sources. A dynamic tuning module adjusts hypergraph weights in real time based on anomaly feedback, forming an "evaluation-localization-optimization" closed loop, improving the system's response speed and accuracy to complex faults.

[0070] 5. This invention uses hypergraph topology optimal transmission and a knowledge distillation loss function to achieve core tensor distribution alignment and node feature mapping between source and target devices. This mechanism overcomes the limitations of single-device data silos and leverages topological similarity to rapidly generate a state assessment model adapted to the target device, significantly reducing the labeled data requirements and training costs for new device deployments. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 is a flow chart of the method of the present invention;

[0072] Figure 2 This is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0073] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0074] See also Figure 1 The present invention provides a method for monitoring and analyzing equipment status based on big data, which includes the following steps:

[0075] S1, perform protocol adaptation and dynamic mask cleaning on multi-source heterogeneous sensor data to generate standardized data stream;

[0076] In this embodiment, the protocol adaptation and dynamic mask cleaning process of multi-source heterogeneous sensor data is achieved by generating a standardized data stream through the following technical solutions:

[0077] For sensor devices with different communication protocols (such as Modbus, CAN, MQTT, etc.), a multi-protocol parsing rule base is established to convert the original data stream into a unified structured format. Through the preset field mapping table, the function code, register address, and data packet identifier in the protocol are converted into standard field names, and the timing data, device ID and timestamp information are extracted. For unstructured data (such as JSON or binary streams), a recursive parsing algorithm is used to extract key parameters to form an initial data matrix with consistent dimensions. Where N represents the number of sensors and D represents the raw data dimension of a single sensor.

[0078] In order to eliminate sensor noise and outlier interference, a sliding median filter is used to filter the original data matrix X raw Perform channel-by-channel filtering. Specifically, the time series data of each sensor Slide the data segment with the window length W, calculate the median of the data in the window, and generate the filtered data matrix X filtered .

[0079] On this basis, the autoencoder model is trained based on the historical health data of the device, and the current data reconstruction error ∈ is calculated in real time. When ∈ exceeds the preset threshold, a dynamic binary mask matrix M is generated. mask ∈{0,1} N×D , masking outliers. The cleaning process is achieved through element-by-element multiplication:

[0080] X clean =MedianFilter(X raw )⊙Mmask ;

[0081] Among them, ⊙ represents the element-by-element multiplication of the matrix, and the final output is the cleaned data matrix X after removing noise and outliers. clean .

[0082] In order to fully characterize the equipment operation status, the cleaned data matrix X clean Perform joint feature extraction in the time-frequency domain. First, use the short-time Fourier transform (STFT) to perform time-frequency analysis on the time series data segment of each sensor, intercept the signal segment through the Hanning window function, calculate its spectral energy distribution, and generate the time-frequency feature matrix. Among them, F1 is the frequency domain feature dimension. At the same time, continuous wavelet transform (CWT) is used to perform multi-scale decomposition on the same signal segment, and the Morlet wavelet basis function is used to generate a scale coefficient matrix with adjustable time-frequency resolution. Among them, F2 is the wavelet scale number. The two types of features are spliced ​​along the feature dimension to construct a joint feature matrix:

[0083] F edge =F STFT ⊕F CWT ;

[0084] Among them, ⊕ represents the column-wise concatenation operation of the matrix, and the final dimension is Standardized data flow.

[0085] Dynamic binary mask matrix M mask The generation of depends on the real-time assessment of the health status of the equipment. Preferably, the autoencoder model is used to perform unsupervised training on historical normal data to reconstruct the error ∈ = ∥X raw -Decoder(Encoder(X raw ))∥2 is used as a health status indicator. When the real-time data ∈ exceeds the dynamic threshold τ (adaptively adjusted according to the 3σ principle of historical error distribution), the corresponding data point is judged to be abnormal and mask The corresponding positions are set to zero, thereby shielding abnormal interference while retaining valid data.

[0086] The resulting normalized data stream is represented by the feature matrix F edge is a carrier, whose row vectors correspond to each sensor node, and the column vectors contain the joint features of the time and frequency domains. Preferably, to adapt to the subsequent hypergraph modeling requirements, the feature matrix also retains the physical attributes of the sensor nodes (such as location coordinates and type labels) as metadata for constructing the physical topology relationship of the devices.

[0087] S2. Based on the standardized data flow and the physical topological relationship of the device, a dynamic hypergraph topological structure is constructed, and the device operation physical constraint equation is embedded in the hypergraph convolution operation;

[0088] In this embodiment, a dynamic hypergraph topology is constructed based on the relationship between standardized data flows and device physical topology, and a technical solution for implementing hypergraph convolution operations is implemented by embedding device-running physical constraint equations. The specific implementation process is as follows:

[0089] The hypergraph topology is defined as a tensor It contains a set of sensor nodes, a set of hyperedges, and a multimodal interaction weight tensor. It is generated by mapping the physical topology of the device. Each vertex corresponds to a sensor node and carries location coordinates (such as three-dimensional space coordinates) and type attributes (such as temperature, vibration, current, etc.). Hyperedge set ε t The construction is based on the logical and physical coupling relationship of the equipment operation. Each hyperedge connects at least three sensor nodes and is used to characterize the complex interaction between multiple nodes (such as the vibration transmission path of the bearing-gear-shaft in the gearbox). Multimodal interaction weight tensor The generation depends on the joint feature matrix F of the standardized data stream edge , where K represents the number of sensor types and M represents the number of operating modes (such as no-load, full-load, overload, etc.). Preferably, the weight value is calculated by the mutual information of the feature matrix, reflecting the correlation strength of different sensor types under specific operating conditions.

[0090] In order to incorporate the physical laws of device operation into hypergraph convolution operations, conservation equations or dynamic equations need to be discretized. For example, the heat conduction equation Discretize into matrix form by finite difference method Among them, Q is the number of spatial discretization points, and R3 is the modal dimension of the constraint equation. The discretization matrix is ​​used as the physical constraint term Through the learnable projection matrix Embedded into the hypergraph convolutional layer to ensure that the model output conforms to the physical laws of the device.

[0091] Hypergraph convolution operation realizes multimodal interaction and feature propagation through tensor module product. The convolution formula of the lth layer is defined as:

[0092]

[0093] in, represents the vertex feature matrix of the lth layer, N is the total number of sensor nodes, d l is the feature dimension; is a learnable parameter matrix used for feature dimension transformation; σ is a nonlinear activation function (such as ReLU). 1 modular product × 1 represents the weight tensor By row and vertex feature matrix H (l)Interact to capture the correlation pattern between sensor types; the 2-module product × 2 represents the weight tensor column-wise and the parameter matrix Θ (l) Interact to achieve adaptive mapping of feature dimensions. Physical constraints Discrete physical equations are embedded into the convolution output through linear projection to ensure that the model maintains a balance between data-driven and physical laws.

[0094] Weight Tensor Dynamically adjust as the equipment working conditions change. Preferably, the time-frequency feature matrix F based on the standardized data stream edge , calculate the real-time correlation strength between sensor types through mutual information, and update the third dimension (operating mode dimension) parameters of the weight tensor. For a specific operating mode m, its corresponding weight slice Calculated by the following formula:

[0095]

[0096] Here, I(·) represents the conditional mutual information, which is used to quantify the association strength between sensor types k1 and k2 under working condition m.

[0097] Physical constraints The generation of Φ requires selecting the appropriate physical equation according to the type of equipment. Taking rotating mechanical equipment as an example, the finite difference form of the Navier-Stokes equation is used to discretize and generate the constraint matrix Φ PDE , the discretization process includes the following steps:

[0098] Grid the velocity field and pressure field variables in the fluid mechanics equations in the equipment space domain;

[0099] Use central difference approximation to the derivative terms in partial differential equations to construct linear equations;

[0100] Convert the system of equations into matrix form Φ PDE , as a basic template for physical constraints.

[0101] Through the learnable parameter matrix The physical constraint template Φ PDE Dynamically adapt to the feature space of different convolutional layers to achieve deep fusion of physical laws and data features.

[0102] Hypergraph convolution operations break through the limitation of traditional graph neural networks that edges can only connect two nodes through multi-node interactions connected by hyperedges. For example, when describing the coupled vibration of bearings, gears, and shafts, a single hyperedge can simultaneously associate vibration sensors, temperature sensors, and speed sensors. The three-modal interaction modeling method simultaneously models the coupled effects of mechanical vibration, heat conduction, and speed changes. The embedding of physical constraints further constrains the rationality of the convolution output, for example, ensuring that the temperature field prediction results conform to the attenuation characteristics of the heat conduction equation.

[0103] S3. performing incremental tensor decomposition on the dynamic hypergraph topology, updating the core tensor based on a sliding window, and maintaining physical law consistency through manifold space constraints;

[0104] In this embodiment, a technical solution for performing incremental tensor decomposition on a dynamic hypergraph topology and maintaining consistency of physical laws through manifold space constraints is implemented as follows:

[0105] Temporal feature tensor generated based on dynamic hypergraph topology (N is the number of sensors, D ′ is the feature dimension, T is the time series length), and a sliding window mechanism is used to intercept local time series fragments. Where Δ is the window length. By constructing an optimization objective function that includes reconstruction error and incremental smoothing terms, the core tensor is realized. With the factor matrix U (n) The objective function is defined as:

[0106]

[0107] in, is the core tensor; U (n) is a factor matrix; is the discretization matrix of the physical equation; R max is the preset rank upper limit;

[0108] The first term is the reconstruction error term, which ensures the fitting accuracy of the decomposition result to the window data; the second term is the incremental smoothing term, which controls the smoothness of the adjacent window factor matrix through the regularization factor λ to avoid parameter mutation. The constraints include:

[0109] Rank constraint: The modal rank of the core tensor satisfies To control model complexity;

[0110] Physical consistency constraint: The product of the core tensor and the factor matrix must satisfy the discretized physical equations in, is the discretization matrix of the partial differential equation.

[0111] Iteratively update the core tensor using the alternating direction method of multipliers (ADMM) With the factor matrix U (n) Specifically:

[0112] Core Tensor Update: Fixed Factor Matrix U(n) , solved by the least squares method Satisfy rank constraints

[0113] Factor Matrix Update: Fixed For each factor matrix U (n) Perform gradient descent optimization and adjust the parameter update direction in combination with incremental smoothing terms.

[0114] Factor matrix U (n) The update must satisfy the Stiefel manifold constraint U (n)T U (n) =I to ensure the orthogonality of the decomposition results and the consistency of physical laws. This is achieved through the projected gradient descent method:

[0115]

[0116] Where η is the learning rate; The objective function is The gradient, is the factor matrix of the nth mode (tensor decomposition dimension) at the kth iteration; the QR decomposition operation projects the gradient updated matrix onto the Stiefel manifold Preferably, the projection process is implemented by Gram-Schmidt orthogonalization or Householder transformation to ensure the orthogonality of the matrix column vectors.

[0117] Discretization matrix of physical equations Φ PDE The generation of is based on the conservation equations (such as conservation of mass and energy) or dynamic equations (such as the Navier-Stokes equations) of the equipment operation. Taking the fluid mechanics equation as an example, the discretization process includes:

[0118] Grid the continuity equation in the device space domain to generate a set of spatial discrete points;

[0119] Finite difference approximation is used for partial derivatives to construct a system of linear equations;

[0120] Convert the system of equations into matrix form Φ PDE , as a template for physical constraints.

[0121] By constraints Force the core tensor to be consistent with the discretized equation in the third mode (physical constraint mode), so as to ensure that the decomposition result conforms to the physical laws of the device.

[0122] To adapt to changes in equipment operating conditions, the window length Δ and the regularization factor λ are dynamically adjusted according to the data distribution. Preferably, the adaptive parameters are calculated based on the feature similarity ρ of the data in the sliding window:

[0123] λ=λ base ·exp(-γρ);

[0124] Among them, ρ th is the similarity threshold; γ is the attenuation coefficient. When the data distribution changes drastically (ρ<ρ th ), shorten the window length to quickly respond to mutations; otherwise, extend the window to enhance stability.

[0125] S4. Calculate a device anomaly score based on the singular value distribution of the core tensor, and locate the associated hyperedge that caused the anomaly based on a hyperedge influence propagation algorithm model;

[0126] In this embodiment, a technical solution is provided for calculating the device anomaly score based on the singular value distribution of the core tensor and locating the associated hyperedges that cause the anomaly through the hyperedge influence propagation algorithm. The specific implementation process is as follows:

[0127] Core tensor singular value analysis:

[0128] The core tensor obtained by incremental tensor decomposition Perform mode slicing and singular value decomposition. Slicing of the third mode (physical constraint mode) Calculate its physical constraint factor matrix The 3-module product of:

[0129]

[0130] in, represents the slice of the core tensor at the kth physical constraint mode; U (3) (k,:) is the kth row vector of the factor matrix. Perform singular value decomposition (SVD) on the product result to obtain the singular value set Characterize the energy distribution characteristics of each mode.

[0131] Equipment abnormality score S anomaly Through the statistical characteristics of the singular value distribution and the energy weighted calculation of the physical constraint mode, the formula is defined as:

[0132]

[0133] in, is the mean of the singular values; ∥·∥2 is the spectral norm of the matrix. This formula implements anomaly detection through the following mechanism:

[0134] Energy weighting term Reflects the energy intensity under physical constraint mode. Under abnormal conditions, the physical laws are destroyed and the energy distribution deviates from the normal mode.

[0135] Exponential decay term exp(-σ r / σavg ): Suppresses the contribution of high singular values, amplifies the abnormal sensitivity of low singular values, and enhances the ability to capture subtle abnormal features.

[0136] To locate the associated hyperedges that cause anomalies, the gradient contribution of the hyperedge weight to the anomaly score is calculated based on the back-propagation mechanism and fused with the node topology importance measure:

[0137] Gradient contribution calculation: Calculate hyperedge weights through automatic differentiation techniques To S anomaly The partial derivative of Reflects the sensitivity of hyperedge weight changes to anomaly scores;

[0138] Node importance measurement: Use PageRank algorithm to calculate node v i Global topological importance of PageRank (v i ), which measures the connection centrality and information dissemination ability of nodes in the hypergraph;

[0139] Hyperedge influence integration: For each hyperedge e j , aggregates the influence of its connected nodes, the formula is:

[0140]

[0141] in, Represents the hyperedge e j The weight parameter of PageRank(v i ) is the node v i The global topological importance score of i is the hyperedge e j The i-th sensor node contained in belongs to the sensor node set θ I is the dynamically adjusted abnormal threshold; is the partial derivative, which is used to describe the rate of change of a variable in a multivariate function;

[0142] The formulation combines local gradient contributions with global topological importance to identify hyperedges that are both sensitive to anomalies and in critical topological positions.

[0143] Dynamically adjust the threshold θ based on the distribution characteristics of historical anomaly scores I Preferably, an adaptive quantile method is used:

[0144] θ I =μ I +k·σ I ;

[0145] Among them, μ I is the historical influence mean; σI is the standard deviation; k is the sensitivity coefficient. j )>θ I The hyperedges of the connected hyperedge set ε fault ,outputs the set of connected sensor nodes and physical coupling paths,,providing root cause information for subsequent maintenance decisions.

[0146] S5. Based on the topological properties of the associated hyperedges, measure the similarity between devices through optimal transmission of the hypergraph topology, and implement cross-device analysis model migration in combination with the knowledge distillation loss function to generate a state assessment model for the target device;

[0147] In this embodiment, a technical solution for implementing cross-device analysis model migration based on the topological properties of associated hyperedges is used to generate a state assessment model for the target device through the joint optimization of the hypergraph topology optimal transmission and the knowledge distillation loss function. The specific implementation process is as follows:

[0148] Define the source device hypergraph Hypergraph with target device Its topological units are sub-hypergraph sets and Embedding the network through hypergraph Map the sub-hypergraph into a low-dimensional feature vector and calculate the topologically optimal transmission distance:

[0149]

[0150] Where S represents the hypergraph topology of the source device (Source-Device), including its sensor node set, hyperedge connection relationship and physical constraints; T represents the hypergraph topology of the target device (Target-Device), which must have similar physical topology characteristics to the source device; Π(μ S ,μ T ) satisfies the marginal distribution μ S (Source device sub-hypergraph distribution) and μ T The transmission matrix set of (target device sub-hypergraph distribution). The optimal transmission matrix Γ * The Sinkhorn algorithm is used to iteratively solve the problem, balancing the transmission cost and the entropy regularization term to ensure that overfitting is avoided while preserving the topological similarity.

[0151] In order to transfer the knowledge of the source device to the target device, the global consistency of the core tensor distribution of the two devices is constrained. The knowledge distillation loss function is defined as:

[0152]

[0153] in, is the optimal transmission matrix; Denotes the i-th and j-th core tensor slices of the source and destination devices, respectively; D KL is the Kullback-Leibler divergence, which is used to measure the similarity of the distribution of two core tensors. Force the target device to approximate the physically constrained modal characteristics of the source device in feature space.

[0154] To enhance the consistency of node-level features across devices, define the optimal transmission loss term Constraining local topological similarity via weighted node embedding distance:

[0155]

[0156] in, and are the node embedding vectors of the source device and the target device, respectively, obtained from the last layer output of the hypergraph convolutional network. The total loss function is a weighted combination of the knowledge distillation loss and the optimal transmission loss:

[0157]

[0158] in, is the total loss function for transfer learning, which is used for cross-device knowledge transfer tasks by jointly optimizing knowledge distillation and topology alignment objectives; α∈(0,1) is an adjustable balance weight coefficient used to control the optimization ratio of global distribution alignment and local feature alignment.

[0159] See also Figure 2 The present invention also provides a big data-based equipment status monitoring and analysis and evaluation system, which includes the following modules:

[0160] The data preprocessing module is used to perform protocol conversion and dynamic mask cleaning on multi-source heterogeneous sensor data to generate a standardized data stream containing joint features in the time and frequency domains;

[0161] This module is responsible for protocol adaptation and cleaning of multi-source heterogeneous sensor data to generate standardized data streams. Specifically, for sensor devices with different communication protocols (such as Modbus, CAN bus, MQTT, etc.), the raw data is converted into a unified structured format through a predefined protocol parsing rule base. The data cleaning process uses a sliding median filter to suppress impulse noise, and combines it with a dynamic binary mask matrix generated based on the health status of the device to mask abnormal data points. The cleaned data is extracted through joint feature extraction in the time and frequency domains, fusing the frequency domain energy features of the short-time Fourier transform and the time-frequency multi-scale features of the continuous wavelet transform to generate a high-dimensional feature matrix as input for subsequent hypergraph modeling.

[0162] The hypergraph modeling module is used to build a dynamic hypergraph structure based on the physical topology of devices and standardized data flows, and embed physical constraints of device operation in hypergraph convolution operations;

[0163] This module constructs a dynamic hypergraph topology based on device physical topology and standardized data flows. Hypergraph vertices represent sensor nodes, and hyperedges connect multiple sensors that are physically coupled or logically associated. Hypergraph weight tensors are dynamically generated through a multimodal interaction mechanism, reflecting the strength of association between different sensor types under specific operating conditions. Hypergraph convolution operations incorporate physical laws into the feature propagation process as constraints by embedding discretized device physics equations, ensuring that model outputs conform to actual physical laws.

[0164] Tensor analysis module, used to perform incremental tensor decomposition of dynamic hypergraphs and update core tensors through manifold space constraints;

[0165] This module performs incremental tensor decomposition on dynamic hypergraph topologies, capturing segments of time series data using a sliding window and updating the core tensor using manifold spatial constraints. The core tensor decomposition enforces the constraints of the discretized matrix of the physical equations and maintains the orthogonality of the factor matrix through Stiefel manifold projection to avoid modal aliasing. The decomposition results update the core tensor and factor matrix in real time, providing a low-dimensional, high-fidelity state representation for subsequent anomaly detection.

[0166] A state assessment module, which is used to calculate the device anomaly score based on the singular value distribution of the core tensor and locate the associated hyperedges that cause the anomaly;

[0167] This module calculates device anomaly scores based on the core tensor's singular value distribution. It quantifies the probability of device state anomalies by analyzing the energy attenuation characteristics of the singular values ​​and the degree of deviation from the physical constraint mode. Combined with the hyperedge influence propagation algorithm, it calculates the gradient contribution of the anomaly score to the hyperedge weight and incorporates the node's global topological importance (such as PageRank score) to locate the associated hyperedge set that caused the anomaly and output the sensor node and coupling path that are the root cause of the anomaly.

[0168] The transfer learning module is used to realize cross-device knowledge transfer based on the topological attributes of the associated hyperedges and generate a state assessment model for the target device;

[0169] This module generates an optimal transfer matrix based on the topological properties of associated hyperedges and the similarity of inter-device sub-hypergraphs using the hypergraph topology optimal transfer metric. This module incorporates a knowledge distillation loss function to constrain the consistency of the core tensor distributions of the source and target devices and weightedly align node-level embedding features. By jointly optimizing the knowledge distillation loss and the optimal transfer loss, the target device model is driven to inherit the knowledge of the source device while adapting to its unique operational characteristics, generating a customized state assessment model.

[0170] Dynamic tuning module, used to adjust hypergraph weight parameters in real time based on anomaly scores and transfer learning results, driving the dynamic evolution of the system;

[0171] This module receives the anomaly scores and associated hyperedge identifiers from the state assessment module in real time, models the dynamic evolution of the hypergraph topology through neural differential equations, and updates the hyperedge weight tensor in response to changes in the device state. At the same time, it integrates the model migration loss of the transfer learning module and the core tensor of the tensor analysis module to generate a topology evolution gradient signal to guide the iterative optimization of the hypergraph weight parameters. According to changes in the operating conditions of the equipment, the discretization parameters of the physical constraint items in the hypergraph modeling module are adaptively adjusted to ensure the real-time and accuracy of the embedding of physical laws. Through the cross-module parameter synchronization interface, the updated hypergraph weights and physical constraint matrices are synchronized to the hypergraph modeling module and the tensor analysis module, forming a closed-loop feedback loop of "state assessment-parameter optimization-model update" to drive the continuous dynamic evolution of the system.

[0172] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for equipment status monitoring and analysis and evaluation based on big data, characterized in that: The method comprises the following steps: S1, perform protocol adaptation and dynamic mask cleaning on multi-source heterogeneous sensor data to generate standardized data stream; S2. Based on the standardized data flow and the physical topological relationship of the device, a dynamic hypergraph topological structure is constructed, and the device operation physical constraint equation is embedded in the hypergraph convolution operation; S3. performing incremental tensor decomposition on the dynamic hypergraph topology, updating the core tensor based on a sliding window, and maintaining physical law consistency through manifold space constraints; S4. Calculate a device anomaly score based on the singular value distribution of the core tensor, and locate the associated hyperedge that caused the anomaly based on a hyperedge influence propagation algorithm model; S5. Based on the topological properties of the associated hyperedges, similarities between devices are measured through optimal transmission of hypergraph topology, and cross-device analysis model migration is achieved in combination with the knowledge distillation loss function to generate a state assessment model for the target device.

2. The device status monitoring and analysis and evaluation method based on big data according to claim 1 is characterized in that: In step S1, the step of performing protocol adaptation and dynamic mask cleaning on multi-source heterogeneous sensor data to generate a standardized data stream includes: Mask generation: Generate a dynamic binary mask matrix M based on the sliding median filter and the historical health status of the device mask ∈{0,1} N×D , and mask the abnormal time period or sensor channel, where N represents the number of sensors and D represents the dimension of the original data; Noise suppression: the original input data Apply sliding median filtering on each channel to remove impulse noise and generate element-by-element multiplication fusion filtering results; Data cleaning: Based on the element-by-element multiplication operation, fusion filtering results, data cleaning model and dynamic binary mask matrix, multi-source heterogeneous sensor data is cleaned to generate cleaned data. The data cleaning model is: X clean =MedianFilter(X (aw )⊙M mask ; in, is the original input data; ⊙ represents the element-by-element multiplication operation, M mask is a dynamic binary mask matrix; Using the joint feature matrix F edge Performing time-frequency domain joint feature generation on the cleaned data to obtain the standardized data stream; The joint feature matrix F edge By the short-time Fourier transform feature F STFT and continuous wavelet transform feature F CWT It is obtained by splicing along the feature dimension, and the splicing satisfies the relational expression: F edge =F STFT ⊕F CWT ; Among them, ⊕ represents the feature concatenation operation; F STFT and F CWT The calculation is performed by intercepting the signal segments through the sliding window.

3. The device status monitoring and analysis and evaluation method based on big data according to claim 1 is characterized in that: In step S2, the steps of constructing a dynamic hypergraph topology structure based on the standardized data flow and the physical topology relationship of the device, and embedding the device operation physical constraint equation into the hypergraph convolution operation include: The dynamic hypergraph topology is defined as a hypergraph tensor in, is the set of sensor nodes; ε t is a set of hyperedges and each hyperedge connects at least three sensor nodes; is the multimodal interaction weight tensor, K represents the number of sensor types, and M represents the number of working modes; The physical constraint equations for the device operation are generated by discretizing the partial differential equations of the physical laws of the device and embedded into the hypergraph convolution operation. The formula for the convolution operation is: Among them, H (l+1) is the (l+1)th layer vertex feature matrix, represents the vertex feature matrix of the lth layer, N represents the number of sensors, d l Represents the original data dimension in the vertex feature matrix of the lth layer; is the learnable parameter matrix; is the physical constraint term; σ is the nonlinear activation function; ×1 and ×2 represent the 1-module product and 2-module product of the tensor, respectively.

4. The device status monitoring and analysis and evaluation method based on big data according to claim 3 is characterized in that: The physical constraints Generates conservation or kinetic equations by discretizing the device operation, including heat conduction equations or the finite difference form of the Navier-Stokes equations; In the hypergraph convolution operation, the 1-module product × 1 represents the weight tensor and the vertex feature matrix H (l) The row-by-row interaction, 2 modular product × 2 represents the weight tensor With the parameter matrix Θ (l) The column-by-column interaction is mapped to the high-dimensional space through the nonlinear activation function σ.

5. The device status monitoring and analysis and evaluation method based on big data according to claim 1 is characterized in that: The steps of S3 include: Intercepting the time series feature tensor through the sliding window And construct an optimization objective function, where N is the number of sensors, D′ is the feature dimension, t is the sliding window, and Δ is the window length. The optimization objective function is: The constraints are and in, is the core tensor; U (n) is a factor matrix; is the discretization matrix of the physical equation; R max is the preset rank upper limit; The core tensor is updated based on the sliding window, and the parameter update is achieved through the manifold space constraint model and Stiefel manifold projection. The manifold space constraint model is: Where η is the learning rate; The objective function is The gradient, is the factor matrix of the nth mode at the kth iteration; the QR decomposition operation projects the gradient updated matrix onto the Stiefel manifold projection, which is 6. The device status monitoring and analysis and evaluation method based on big data according to claim 1 is characterized in that: The steps of S4 include: Through Core Tensor The singular value and the abnormality score calculation formula are used to perform distribution calculation to obtain the device abnormality score. The abnormality score calculation formula is: Among them, S anomaly The value for the device anomaly score, represents the slice of the core tensor at the rth mode; is the physical constraint factor matrix; σ r is the rth singular value; is the mean of the singular values; Based on the hyperedge influence propagation algorithm model, the gradient contribution of the hyperedge weight to the anomaly score is calculated, and the associated hyperedge that causes the anomaly is determined in combination with the importance measurement of the sensor node. The hyperedge influence propagation algorithm model is: Among them, I(e j ) is the associated hyperedge that caused the exception, Represents the hyperedge e j The weight parameter of PageRank(v i ) is the sensor node v i The global topological importance score of I(e j )>θ I The hyperedges of v constitute the associated hyperedge set, i is the hyperedge e j The i-th sensor node contained in belongs to the sensor node set θ I is the dynamically adjusted abnormal threshold; is the partial derivative.

7. The device status monitoring and analysis and evaluation method based on big data according to claim 1 is characterized in that: In step S5: The optimal transmission of the hypergraph topology is achieved by solving the source device hypergraph Hypergraph with target device The optimal transmission matrix between To achieve this, the transmission distance is defined as: Where S represents the hypergraph topology of the source device; T represents the target device; Π(μ S ,μ T ) satisfies the marginal distribution μ S ,μ T The set of transmission matrices; and All are hypergraph topology embedding functions; The knowledge distillation loss function is implemented by constraining the core tensor distribution consistency between the source device and the target device, and is defined as: in, is the optimal transmission matrix; Denotes the i-th and j-th core tensor slices of the source and destination devices, respectively; D KL is the Kullback-Leibler divergence.

8. The method for equipment status monitoring and analysis and evaluation based on big data according to claim 1 is characterized in that: The total loss function for cross-device analysis model migration is a weighted combination of knowledge distillation loss and optimal transfer loss. The knowledge distillation loss function is: in, is the total loss function of transfer learning; is the optimal transmission loss term, and are the node embedding vectors of the source device and the target device respectively; α∈(0,1) is the balancing weight coefficient.

9. A big data-based equipment status monitoring and analysis and evaluation system, applied to the method according to any one of claims 1 to 8, characterized in that: The system includes the following modules: The data preprocessing module performs protocol conversion and dynamic mask cleaning of multi-source heterogeneous sensor data, and outputs a standardized data stream containing joint features in the time and frequency domains; The hypergraph modeling module builds a dynamic hypergraph based on the physical topology of devices and standardized data flows, and embeds the physical constraint equations of device operation into the hypergraph convolution; Tensor analysis module, which performs incremental tensor decomposition on dynamic hypergraphs and updates core tensors through manifold space constraints; The status assessment module calculates the device anomaly score based on the core tensor singular value distribution and locates the associated hyperedges using the hyperedge influence propagation algorithm; The transfer learning module realizes cross-device knowledge transfer and generates a target device evaluation model based on the topological properties of the associated hyperedges and the optimal transmission distance; The dynamic tuning module adjusts the hypergraph weight parameters in real time based on anomaly scoring and transfer learning results to drive the dynamic evolution of the system.

10. The equipment status monitoring and analysis and evaluation system based on big data according to claim 9 is characterized in that: The dynamic tuning module is specifically used to perform the following operations: Based on the anomaly score and the associated hyperedge identifier, the hyperedge weight tensor is updated through a gradient-driven mechanism; Fusion of transfer learning loss and core tensor features to generate topology evolution gradient signals; Adaptively adjust the physical constraint discretization parameters of the hypergraph modeling module; Hypergraph weights and physical constraints are updated synchronously through cross-module interfaces to form a closed-loop optimization.

Images