Machine equipment on-line state monitoring and fault diagnosis system

Abstract

The invention relates to the technical field of industrial Internet of Things, in particular to a machine equipment online state monitoring and fault diagnosis system, which comprises the following steps of: acquiring multi-source heterogeneous sensing data through an edge computing node deployed on an equipment body, performing adaptive noise filtering and feature dimension reduction processing on original data, and acquiring multi-source heterogeneous sensing data; outputting a standardized equipment state vector set; inputting the equipment state vector set into a dynamic knowledge graph engine, constructing a fault evolution network comprising space-time correlation characteristics based on an equipment operation entropy change quantification model, and generating a graph node connection relationship with a weight coefficient; and inputting the fault evolution network into a migration reinforcement learning module, and outputting a diagnosis decision set comprising a fault type, a severity degree and an evolution path through knowledge migration of a cross-device fault mode. According to the method, the problems of edge redundancy and single feature expression in traditional rule-based atlas construction are effectively avoided, and the structuring ability and physical traceability of fault recognition are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial Internet of Things, and in particular to an online status monitoring and fault diagnosis system for machine equipment. Background Art

[0002] With the increasing intelligence of industrial equipment, online condition monitoring and fault diagnosis have become key technologies to ensure stable equipment operation and extend its service life. Existing monitoring systems generally rely on a single type of sensor (such as vibration or temperature) to collect data and identify anomalies through rule-based or static threshold methods. However, as equipment operating environments become increasingly complex and failure modes become more diverse and cross-cutting, traditional solutions have significant shortcomings in the following areas:

[0003] Fusion of heterogeneous multi-source data is challenging: Vibration, temperature, and current sensors commonly used in industrial settings have varying sampling frequencies, transmission protocols, and physical dimensions, making unified modeling difficult. Existing methods typically employ simple concatenation or single-channel compression, lacking systematic spatiotemporal alignment and noise decoupling mechanisms. This results in inaccurate state representation and is prone to false positives and omissions.

[0004] Fault relationship modeling lacks structure and explainability: Current mainstream methods mostly perform anomaly detection based on statistical features or deep features, ignoring the spatial topological structure of the equipment itself and the coupled evolution process between multiple physical quantities. This makes it difficult to form a visual description of the fault propagation path and potential causal relationships, limiting the effectiveness of fault tracing and preventive maintenance.

[0005] Weak cross-device knowledge transfer and poor generalization: Differences in industrial equipment models, operating conditions, and load conditions make it difficult to directly transfer trained models to different devices. Traditional rule-based or shallow model-based solutions rely on manual parameter adjustment and lack effective cross-domain alignment mechanisms, making them difficult to adapt to multi-scenario deployment requirements.

[0006] Therefore, there is an urgent need for an intelligent diagnosis system that can integrate multi-source heterogeneous data, has the ability to model structured relationships, and supports cross-device transfer learning to improve the accuracy, stability, and explainability of fault identification in complex scenarios. Summary of the Invention

[0007] The present invention provides a machine equipment online status monitoring and fault diagnosis system.

[0008] A machine equipment online status monitoring and fault diagnosis system includes the following steps:

[0009] S1. The edge computing nodes deployed on the device collect multi-source heterogeneous sensor data, perform adaptive noise filtering and feature dimensionality reduction on the raw data, and output a standardized device state vector set.

[0010] S2. Input the device state vector set into a dynamic knowledge graph engine, construct a fault evolution network including spatiotemporal correlation features based on a variable model of device operation entropy, and generate a graph node connection relationship with a weight coefficient;

[0011] S3. Input the fault evolution network into the transfer reinforcement learning module, and output a diagnostic decision set including fault type, severity and evolution path through knowledge transfer across device failure modes.

[0012] Optionally, the S1 specifically includes:

[0013] S11. Spatiotemporal alignment of multi-source data: A hardware timestamp synchronization module performs millisecond-level time alignment on raw data from vibration, temperature, and current sensors. A sliding window dynamic compensation mechanism eliminates multi-protocol transmission delays, generating heterogeneous data streams with strict temporal alignment.

[0014] S12. Layered Adaptive Filtering:

[0015] An improved wavelet threshold denoising algorithm is used for vibration acceleration signals, and the threshold function is dynamically adjusted based on the frequency band energy distribution.

[0016] Implement Kalman filtering based on sudden noise suppression for temperature gradient data, and constrain the filter parameter update through the heat conduction physical model;

[0017] A notch filter group is used to eliminate power frequency interference from current harmonic components, and the center frequency of the filter is adjusted in real time according to the equipment speed;

[0018] S13. Lightweight Manifold Dimensionality Reduction: The filtered multidimensional data is fed into an improved local linear embedding algorithm. By introducing equipment operating condition labels to constrain the manifold projection direction, the original data is mapped from the high-dimensional space to a 3D feature space.

[0019] S14. Dynamic normalization processing: Based on the extreme value distribution model of the historical operation data of the equipment, the feature vector after dimensionality reduction is adaptively normalized to generate a standardized set of equipment state vectors, where the value range of each equipment state vector element is compressed to the interval [0,1].

[0020] Optionally, the multi-source heterogeneous sensing data in S1 includes vibration acceleration signals, temperature gradient data, and current harmonic components, and each data channel uses an independent adaptive filter to suppress noise.

[0021] Optionally, the S2 specifically includes:

[0022] S21. Graph node generation:

[0023] Each vector in the device state vector set is mapped to an entity node of the knowledge graph, and the attributes of each node include:

[0024] The time-frequency domain feature entropy value comes from the time-frequency domain feature entropy calculation unit;

[0025] The multi-physics field coupling entropy value is output by the multi-physics field coupling entropy calculation unit;

[0026] The entropy change trend prediction value is generated by the entropy change trend prediction unit;

[0027] S22. Spatiotemporal Correlation Modeling:

[0028] Time dimension association: A sliding time window is used to analyze the evolution of node entropy values. A time-series transmission edge is established for nodes that satisfy ΔH ≥ α within N consecutive sampling periods. ΔH represents the entropy change trend value, and α represents the adaptive anomaly detection threshold of the entropy change trend.

[0029] Spatial dimension association: Based on the physical topology of the device, spatial coupling edges are established for nodes that simultaneously meet the spatial correlation conditions;

[0030] S23. Weight coefficient calculation:

[0031] Assign weights to timing conduction edges Reflects the significance of the entropy change trend; where k represents the slope factor corresponding to the device type;

[0032] Assign weights to spatial coupling edges Among them, ρ ij Indicates the physical connection strength coefficient between equipment components, Indicates H in historical data c The maximum value, H c (i) H c (j) represents the coupling entropy of nodes i and j, ρ ij Indicates the strength coefficient of the physical connection between equipment components;

[0033] S24. Dynamic graph optimization:

[0034] Implement the graph dynamic pruning algorithm to adjust the weight coefficient W t (i)+W s Soft deletion is performed on edge connections with (i, j) < θ, retaining potential fault transmission paths; where θ represents the edge weight clipping threshold;

[0035] Generate a fault evolution network with weight coefficients, where nodes represent equipment status modes and edges represent the possibility of fault transmission.

[0036] Optionally, the processing unit of the device running the entropy variable quantization model includes:

[0037] The time-frequency domain characteristic entropy calculation unit calculates the energy entropy of each frequency band after performing wavelet packet decomposition on the vibration signal;

[0038] Multi-physics field coupling entropy calculation unit, which integrates temperature field gradient and current harmonic distortion rate to construct a joint entropy index;

[0039] The entropy change trend prediction unit uses a long short-term memory network to predict the evolution path of the device entropy value.

[0040] Optionally, the dynamic knowledge graph engine in S2 adopts a graph dynamic pruning algorithm, specifically including:

[0041] The node activity evaluation module calculates the importance score of graph nodes based on the Mahalanobis distance of the device state vector;

[0042] The association decay module implements exponential decay on edge connections that have not been activated for N consecutive sampling periods;

[0043] The graph topology optimization module dynamically adjusts the graph hierarchy structure according to the equipment operation stage.

[0044] Optionally, the S3 specifically includes:

[0045] S31. Cross-domain feature alignment:

[0046] Input the graph node connection relationship in the fault evolution network into the source domain knowledge extractor to extract a transferable fault mode feature set, including:

[0047] a) Node entropy change sensitivity vector;

[0048] b) spatial coupling propagation mode matrix;

[0049] The feature distributions of source and target domain devices are aligned through the domain adversarial training module, and the gradient reversal layer is used to enable the feature extractor to generate domain-invariant features.

[0050] S32. Strengthen strategy optimization:

[0051] Construct a dual-channel strategy network to handle:

[0052] a) Timing conduction channel: predicting the fault development stage based on the timing edge weights of the fault evolution network;

[0053] b) Spatial coupling channel: Calculate the fault diffusion probability based on the spatial edge weight;

[0054] Calculate multi-dimensional reward values ​​through the reward function generator:

[0055]

[0056] Among them, R represents the comprehensive reward value of the strategy network; R accuracy Indicates the diagnostic accuracy, calculated based on maintenance records; R timeliness =exp(-δt) represents the function of warning lead time, where δt is the number of hours of advance; R cost Indicates the maintenance cost savings compared with planned maintenance costs; Represents the reward weighted coefficient, satisfying

[0057] S33. Decision set generation:

[0058] Output a diagnosis decision tuple, wherein the dimensions of the diagnosis decision tuple include: fault type code, severity index, and evolution path topology graph;

[0059] Implement double verification on decision results:

[0060] Test 1: Physical constraint verification: Verify the physical feasibility of the fault type and equipment operating parameters;

[0061] Test 2: Trend consistency check: Compare the similarity of the historical state degradation trajectories of the equipment.

[0062] Optionally, the transfer reinforcement learning module includes:

[0063] Cross-domain feature alignment unit: extracts transferable structured graph features from source domain devices and maps them to the target domain, achieving feature distribution consistency through an adversarial mechanism;

[0064] Strategy Optimization Learning Unit: Learn how to predict the probability of failure and its development process based on the time / space evolution path in the graph structure;

[0065] Diagnostic decision generation and verification unit: Outputs clear diagnostic results and ensures their availability and rationality through dual verification of physical rules and historical trends.

[0066] Optionally, the domain adversarial training module in S31 specifically includes:

[0067] Feature Extractor: Extracts structured features from the input fault evolution graph, taking into account both time series patterns and spatial topology information;

[0068] Gradient reversal layer: keeps the features unchanged during forward propagation and negates the gradient during backpropagation to enhance the feature extractor against the domain classifier;

[0069] Domain classifier: Determines whether the input feature belongs to the source domain or the target domain, and learns the difference in feature distribution;

[0070] Domain adversarial loss function: The optimization goal is to minimize the domain classification error, while enabling the feature extractor to learn the representation of "indistinguishable domains" through GRL.

[0071] Beneficial effects of the present invention:

[0072] This invention deploys edge computing nodes within the device itself to synchronize the collection and processing of multi-source sensor data, including vibration, temperature, and current. It employs a combination of improved wavelet packet threshold filtering, Kalman filtering, and notch filtering to dynamically suppress unstructured noise. It also incorporates lightweight manifold dimensionality reduction and a historical normalization standard model to construct a unified device state vector. This addresses the challenges of multi-source signal fusion and high dimensionality reduction distortion in existing technologies, improving the timeliness and feature sensitivity of the monitoring model.

[0073] This invention introduces a three-entropy fusion mechanism: "time-frequency domain feature entropy," "multi-physics field coupling entropy," and "entropy change trend." This constructs a node graph of device operating status. Dynamic threshold control is used to establish temporal conduction edges and spatial coupling edges. Sigmoid and physical connection strength drive weight generation, forming an interpretable fault evolution network. Dynamic graph pruning further preserves potential fault paths, effectively avoiding the edge redundancy and single feature representation issues inherent in traditional rule-based graph construction, thereby enhancing the structured capabilities and physical traceability of fault identification.

[0074] This invention combines the gradient reversal mechanism with graph features to construct a cross-domain adversarial learning framework, achieving feature alignment between the node entropy change sensitivity and the spatial propagation pattern in the source / target domains, solving the problems of large differences in equipment types and the non-transferability of empirical knowledge. At the policy level, a dual-channel reinforcement learning structure is constructed to evaluate the fault evolution stage and diffusion risk based on the temporal and spatial channels of the fault propagation graph, and policy optimization is achieved through a reward function that integrates accuracy, early warning, and economic cost, ultimately outputting fault diagnosis decisions with multi-dimensional semantics. This method breaks through the traditional policy mechanism based on threshold discrimination or static models, and has stronger generalization capabilities and industrial adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0076] Figure 1 A schematic diagram of system execution steps according to an embodiment of the present invention;

[0077] Figure 2Schematic diagram of a transfer reinforcement learning module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0078] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0079] It should be noted that references in the specification to "one embodiment," "an embodiment," "an exemplary embodiment," "some embodiments," etc. indicate that the described embodiments may include specific features, structures, or characteristics, but not every embodiment necessarily includes such specific features, structures, or characteristics. In addition, when specific features, structures, or characteristics are described in conjunction with an embodiment, it is within the knowledge of persons skilled in the relevant art to implement such features, structures, or characteristics in conjunction with other embodiments (whether or not explicitly described).

[0080] In general, terms can be understood, at least in part, from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending at least in part on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0081] like Figure 1-Figure 2 As shown, a machine equipment online status monitoring and fault diagnosis system includes the following steps:

[0082] S1. The edge computing nodes deployed on the device collect multi-source heterogeneous sensor data, perform adaptive noise filtering and feature dimensionality reduction on the raw data, and output a standardized device state vector set.

[0083] S2. Input the device state vector set into the dynamic knowledge graph engine, build a fault evolution network including spatiotemporal correlation features based on the device operation entropy variable model, and generate a graph node connection relationship with weight coefficients;

[0084] S3. Input the fault evolution network into the transfer reinforcement learning module, and through the knowledge transfer of cross-device failure modes, output a diagnostic decision set including fault type, severity and evolution path.

[0085] S1 specifically includes:

[0086] S11. Spatiotemporal alignment of multi-source data:

[0087] The hardware timestamp synchronization module is used to synchronize the data streams from heterogeneous sensors such as vibration (V(t)), temperature (T(t)), and current (I(t)) at the millisecond level. Assume that the sensor data acquisition timestamp set is Use the sliding window mechanism to calculate the transmission delay δ (x) (x∈{V,T,I}) is dynamically compensated to obtain the aligned time series: in, Indicates the original timestamp of the i-th sampling point; δ (x) (t i ) represents the estimated transmission delay value of the i-th sampling point; Indicates the timestamp after time alignment;

[0088] S12. Layered Adaptive Filtering:

[0089] S121. Vibration signal filtering: The vibration acceleration signal V(t) is denoised using an improved wavelet threshold. The filtered output is:

[0090] represents the wavelet coefficient after denoising, w k represents the k-th layer wavelet decomposition coefficient, λ k represents the adaptive threshold of the kth layer, sign(·) represents the sign function;

[0091] in, The dynamic adjustment factor α is based on the current frequency band energy, N represents the total number of signal sampling points, σ k represents the standard deviation of the noise estimation in the kth layer;

[0092] S122. Temperature data filtering: Apply Kalman filtering based on heat conduction physical constraints to the temperature gradient data T(t):

[0093] represents the filtered temperature estimate (at the current moment), Represents the estimated temperature at the previous moment, T k Indicates the current measured temperature value, K k represents the Kalman gain coefficient;

[0094] The gain is: k represents the thermal diffusion coefficient (constraint filter parameter), P k-1 represents the prior error covariance; R k represents the measurement noise covariance;

[0095] S123. Current signal filtering: to filter the harmonic interference components I in the current signal h(t), using a notch filter bank:

[0096]

[0097] Among them, H i (f) represents the transfer function of the i-th notch filter, s represents the Laplace domain complex variable, ζ represents the damping coefficient, ω i Indicates the center angular frequency of the filter, unit: rad / s, ω i =2πf i (t), f i (t) represents the dynamic center frequency based on the rotational speed of the device.

[0098] S13. Lightweight manifold dimensionality reduction: Let the feature matrix before dimensionality reduction be X, the equipment condition label be y, and introduce the constrained local linear embedding (LLE) optimization objective:

[0099]

[0100] Where Y represents the three-dimensional feature matrix after dimensionality reduction; is an indicative function, a weighted penalty across label points; λ represents the label constraint adjustment coefficient; represents the neighbor set of sample i, is the set of neighbors calculated in X according to the Euclidean distance; w ij represents the weight when reconstructing sample i using its neighbor j in X; Represents the characteristic function, if y i ≠y j 1 if yes, 0 otherwise.

[0101] S14. Dynamic standardization processing:

[0102] For each dimensional feature vector f j , normalized according to the historical extreme value distribution model:

[0103]

[0104] Among them, f j represents the j-th dimension of the feature vector after dimensionality reduction, represents the historical minimum value of the j-th dimension, represents the historical maximum value of the j-th dimension, ∈ represents a very small constant to prevent the denominator from being zero, Represents the normalized eigenvalue of the j-th dimension, with the range compressed to [0,1].

[0105] S2 specifically includes:

[0106] S21. Graph node generates a state vector for each device Mapped to an entity node v in the graphi , this node includes three entropy attributes:

[0107] 1) Time-frequency domain characteristic entropy value: The time-frequency domain characteristic entropy calculation unit performs wavelet packet decomposition on the vibration signal v(t) and decomposes it to the Lth layer to obtain 2 L sub-bands, the total energy is E i Represents the energy of the i-th frequency band, L represents the number of wavelet packet decomposition layers; calculate the energy probability distribution: p i Represents the energy proportion (frequency band energy distribution probability); based on Shannon entropy, the time-frequency domain energy entropy H of the signal is defined f for:

[0108] 2) Multi-physics coupling entropy value: The multi-physics coupling entropy calculation unit integrates the temperature gradient field and the current harmonic distortion (THD) to construct a joint distribution entropy model;

[0109] The temperature gradient sequence is represents the temperature gradient at the i-th time slice;

[0110] The current harmonic distortion rate is THD = {THD1, THD2, ..., THD n}, THD j Represents the total harmonic distortion rate of the current in the jth time slice, and constructs the joint distribution probability matrix P i,j ,P i,j Represents the joint probability under the i-th temperature gradient and the j-th THD, and calculates the joint entropy: H c =-∑ i ∑ j P i,j ·log(P i,j ), where H c It represents the multi-physics coupling entropy and measures the synergistic complexity of temperature and current fluctuations;

[0111] 3) Entropy change trend prediction value: The entropy change trend prediction unit constructs an input sequence as the entropy value sequence of the previous T time points Input to the LSTM network for sequence modeling and output the predicted value of the next time step The entropy change trend is: Among them, H t represents the total entropy value of the device at time t, It represents the entropy value at the next moment predicted by the LSTM network, and ΔH represents the trend of entropy change, which is used to evaluate system stability or potential failure tendency;

[0112] The LSTM network uses a standard gating structure, and the hidden layer outputs: h t =LSTM(H t ,h t-1 ,c t-1 ), h t Represents the hidden state vector (LSTM output) of the current time step t, which represents the compressed representation of the current input and historical information; H t Indicates the input feature of the current time step. In this application, it is the device entropy value at the current moment, h t-1 represents the hidden state vector of the previous time step t-1, carrying historical memory information; c t-1 Represents the cell state vector of the previous time step, indicating the long-term memory content; LSTM(·) represents the long short-term memory unit function, which includes operations such as input gate, forget gate, output gate, and candidate state update;

[0113] The final output is passed through the fully connected layer to obtain the predicted value.

[0114] S22. Spatiotemporal Correlation Modeling:

[0115] (1) Time dimension association:

[0116] For nodes in N consecutive time windows, if ΔH(i)≥α is satisfied, a time sequence edge is established for them in time order: E t ={(v i ,v i+1 )∣ΔH(i)≥α}, where E t represents the set of sequential conduction edges, which represents the time evolution path connection between nodes that meet the conditions; v i Represents the i-th device status node, corresponding to time t i The state vector of the entity node in the knowledge graph is represented; ΔH(i) represents the entropy change trend value of the i-th node, which represents the change between the device entropy value at the prediction moment and the current entropy value, and is defined as is the entropy value of the next moment predicted by LSTM, H i represents the actual device entropy value at the current i-th time step, i represents the time index, indicating the position of the current sample in the entropy sequence; α represents the adaptive anomaly detection threshold of the entropy change trend, which is used to determine whether to establish a time series edge connection. The threshold α is adaptively defined as α = μ ΔH +3σ ΔH , μ ΔH represents the mean (expected) value of the historical entropy change trend value ΔH, that is, the average change in past data; σ ΔH The standard deviation of the historical entropy change trend value ΔH indicates the degree of fluctuation or dispersion;

[0117] (2) Spatial dimension association: for node vi ,v j , if the following two conditions are met:

[0118] |H f (i)-H f (j)|≤β,|H c (i)-H c (j)|≤γ, then a spatial coupling edge is established:

[0119] E s ={(v i ,v j )|satisfies the above double threshold conditions},

[0120] Among them, H f (i) H f (j) represents the time-frequency domain feature entropy value of the i-th and j-th nodes, H c (i) H c (j) represents the multi-physics field coupling entropy value of the i-th and j-th nodes, β represents the time-frequency domain entropy similarity threshold, and γ represents the coupling entropy similarity threshold.

[0121] S23. Weight coefficient calculation:

[0122] (1) Time series transmission edge weight W t :

[0123]

[0124] Where k represents the slope factor corresponding to the device type, W t (i) represents the temporal transmission edge weight of node i;

[0125] (2) Spatial coupling edge weight W s :

[0126]

[0127] Among them, ρ ij Indicates the physical connection strength coefficient between equipment components, Indicates H in historical data c The maximum value, H c (i) H c (j) represents the coupling entropy of nodes i and j, W s (i,j) represents the edge (v i ,v j )’s spatial coupling weight;

[0128] S24. Dynamic graph optimization:

[0129] (1) Weighted joint judgment for edge soft deletion:

[0130] For each edge (v i ,v j ), satisfy: W t (i)+W s (i, j) < θ, implement soft deletion marking instead of physical deletion to preserve the historical path tracing capability;

[0131] The edge weight clipping threshold θ is dynamically defined as: θ = median(W all )+λ·IQR(W all ), where λ represents the pruning sensitivity coefficient, W all Represents the set of all edge weights in the current graph, median(W all ) represents the median of edge weights, and IQR represents the interquartile range;

[0132] (2) Description of the graph structure:

[0133] The final output graph structure is:

[0134] Node set V = {v1,v2,…,v n} indicates the standardized device status;

[0135] Edge set E = E t ∪E s Indicates the temporal or spatial propagation of fault evolution;

[0136] The weight function W:E→[0,1] represents the conduction strength of the quantized edge;

[0137] Input graph structure for subsequent reinforcement learning modules.

[0138] S3 specifically includes:

[0139] S31. Cross-domain feature alignment:

[0140] S311. Extract the following two types of features from the fault evolution network:

[0141] 1) Node entropy change sensitivity vector:

[0142]

[0143] Among them S i Represents node v i The entropy change sensitivity of node i; ΔH(i) represents the entropy change trend of node i; W t (i) represents the timing conduction weight of node i; k represents the sigmoid mapping slope factor;

[0144] 2) Spatial coupling propagation mode matrix:

[0145]

[0146] Among them A s represents the spatial adjacency weight matrix; W s (i, j) represents the coupling weight of the spatial edge;

[0147] S312. Alignment method:

[0148] The above features are input into the domain adversarial training module, and the gradient reversal layer (GRL) is introduced to extract domain-invariant features to minimize the difference in feature distribution between the source domain and the target domain:

[0149]

[0150] in, represents the domain classification loss; G(x i ) represents the feature extractor output; D(·) represents the domain classifier; d i ∈{source, target} represents the domain label; CE(·) represents the cross entropy loss;

[0151] S32. Strengthen strategy optimization to build a dual-channel strategy network:

[0152] a) Temporal conduction channel: According to the edge weight W t Predicting the stage of fault evolution;

[0153] b) Spatial coupling channel: According to W s Estimate the probability of fault propagation paths.

[0154] The total reward function is defined as:

[0155] Among them, R represents the comprehensive reward value of the strategy network; R accuracy Indicates the diagnostic accuracy, calculated based on maintenance records; R timeliness =exp(-δt) represents the function of warning lead time, where δt is the number of hours of advance; R cost Indicates the maintenance cost saving rate, compared with the planned maintenance cost, Represents the reward weighted coefficient, satisfying

[0156] S33. Decision set generation:

[0157] S331. Final output diagnosis decision triplet: D i =(Code i ,S i ,P i );

[0158] Among them, Code i Indicates the fault type code, S i ∈[0,1] represents the severity index, and the Q value is normalized output; Pi ={p1,p2,…,p k} represents the set of evolution paths, and k fault propagation paths with the highest weights are selected;

[0159] S332. Double verification mechanism:

[0160] 1) Physical constraint verification: Check whether the diagnostic results conflict with the actual operating parameters of the equipment;

[0161] Trend consistency check: Compare similarity with historical state trajectories to ensure that the evolution pattern is reasonable.

[0162] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0163] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A machine equipment online status monitoring and fault diagnosis system, characterized in that: This includes executing the following: S1. The edge computing nodes deployed on the device collect multi-source heterogeneous sensor data, perform adaptive noise filtering and feature dimensionality reduction on the raw data, and output a standardized device state vector set. S2. Input the device state vector set into a dynamic knowledge graph engine, construct a fault evolution network including spatiotemporal correlation features based on a variable model of device operation entropy, and generate a graph node connection relationship with a weight coefficient; S3. Input the fault evolution network into the transfer reinforcement learning module, and output a diagnostic decision set including fault type, severity and evolution path through knowledge transfer across device failure modes.

2. The online status monitoring and fault diagnosis system for machine equipment according to claim 1, characterized in that: Said S1 specifically includes: S11. Spatiotemporal alignment of multi-source data: A hardware timestamp synchronization module performs millisecond-level time alignment on raw data from vibration, temperature, and current sensors. A sliding window dynamic compensation mechanism eliminates multi-protocol transmission delays, generating heterogeneous data streams with strict temporal alignment. S12. Layered Adaptive Filtering: An improved wavelet threshold denoising algorithm is used for vibration acceleration signals, and the threshold function is dynamically adjusted based on the frequency band energy distribution. Implement Kalman filtering based on sudden noise suppression for temperature gradient data, and constrain the filter parameter update through the heat conduction physical model; A notch filter group is used to eliminate power frequency interference from current harmonic components, and the center frequency of the filter is adjusted in real time according to the equipment speed; S13. Lightweight Manifold Dimensionality Reduction: The filtered multidimensional data is fed into an improved local linear embedding algorithm. By introducing equipment operating condition labels to constrain the manifold projection direction, the original data is mapped from the high-dimensional space to a 3D feature space. S14. Dynamic normalization processing: Based on the extreme value distribution model of the historical operation data of the equipment, the feature vector after dimensionality reduction is adaptively normalized to generate a standardized set of equipment state vectors, where the value range of each equipment state vector element is compressed to the interval [0,1].

3. The online status monitoring and fault diagnosis system for machine equipment according to claim 2, characterized in that: The multi-source heterogeneous sensor data in S1 includes vibration acceleration signals, temperature gradient data, and current harmonic components. Each data channel uses an independent adaptive filter to suppress noise.

4. The online status monitoring and fault diagnosis system for machine equipment according to claim 1, characterized in that: The S2 specifically includes: S21. Graph node generation: Each vector in the device state vector set is mapped to an entity node of the knowledge graph, and the attributes of each node include: The time-frequency domain feature entropy value comes from the time-frequency domain feature entropy calculation unit; The multi-physics field coupling entropy value is output by the multi-physics field coupling entropy calculation unit; The entropy change trend prediction value is generated by the entropy change trend prediction unit; S22. Spatiotemporal Correlation Modeling: Time dimension association: A sliding time window is used to analyze the evolution of node entropy values. A time-series transmission edge is established for nodes that satisfy ΔH ≥ α within N consecutive sampling periods. ΔH represents the entropy change trend value, and α represents the adaptive anomaly detection threshold of the entropy change trend. Spatial dimension association: Based on the physical topology of the device, spatial coupling edges are established for nodes that simultaneously meet the spatial correlation conditions; S23. Weight coefficient calculation: Assign weights to timing conduction edges Reflects the significance of the entropy change trend; where k represents the slope factor corresponding to the device type; Assign weights to spatial coupling edges Among them, ρ ij Indicates the physical connection strength coefficient between equipment components, Indicates H in historical data c The maximum value, H c (i) H c (j) represents the coupling entropy of nodes i and j, ρ ij Indicates the strength coefficient of the physical connection between equipment components; S24. Dynamic graph optimization: Implement the graph dynamic pruning algorithm to adjust the weight coefficient W t (i)+W s Soft deletion is performed on edge connections where (i, j) < θ, retaining potential fault transmission paths; where θ represents the edge weight pruning threshold; Generate a fault evolution network with weight coefficients, where nodes represent equipment status modes and edges represent the possibility of fault transmission.

5. The online status monitoring and fault diagnosis system for machine equipment according to claim 1 is characterized in that The processing unit of the device running the entropy variable model includes: The time-frequency domain characteristic entropy calculation unit calculates the energy entropy of each frequency band after performing wavelet packet decomposition on the vibration signal; Multi-physics field coupling entropy calculation unit, which integrates temperature field gradient and current harmonic distortion rate to construct a joint entropy index; The entropy change trend prediction unit uses a long short-term memory network to predict the evolution path of the device entropy value.

6. The online status monitoring and fault diagnosis system for machine equipment according to claim 1, characterized in that: The dynamic knowledge graph engine in S2 adopts a graph dynamic pruning algorithm, which specifically includes: The node activity evaluation module calculates the importance score of graph nodes based on the Mahalanobis distance of the device state vector; The association decay module implements exponential decay on edge connections that have not been activated for N consecutive sampling periods; The graph topology optimization module dynamically adjusts the graph hierarchy structure according to the equipment operation stage.

7. The online status monitoring and fault diagnosis system for machine equipment according to claim 1, characterized in that: The S3 specifically includes: S31. Cross-domain feature alignment: Input the graph node connection relationship in the fault evolution network into the source domain knowledge extractor to extract a transferable fault mode feature set, including: a) Node entropy change sensitivity vector; b) spatial coupling propagation mode matrix; The feature distributions of source and target domain devices are aligned through the domain adversarial training module, and the gradient reversal layer is used to enable the feature extractor to generate domain-invariant features. S32. Strengthen strategy optimization: Construct a dual-channel strategy network to handle: a) Timing conduction channel: predicting the fault development stage based on the timing edge weights of the fault evolution network; b) Spatial coupling channel: Calculate the fault diffusion probability based on the spatial edge weight; Calculate multi-dimensional reward values ​​through the reward function generator: Among them, R represents the comprehensive reward value of the strategy network; R accuracy Indicates the diagnostic accuracy, calculated based on maintenance records; R timeliness =exp(-δt) represents the function of warning lead time, where δt is the number of hours of advance; R cost Indicates the maintenance cost savings compared with planned maintenance costs; Represents the reward weighted coefficient, satisfying S33. Decision set generation: Output a diagnosis decision tuple, wherein the dimensions of the diagnosis decision tuple include: fault type code, severity index, and evolution path topology graph; Implement double verification on decision results: Test 1: Physical constraint verification: Verify the physical feasibility of the fault type and equipment operating parameters; Test 2: Trend consistency check: Compare the similarity of the historical state degradation trajectories of the equipment.

8. The online status monitoring and fault diagnosis system for machine equipment according to claim 1, characterized in that: The transfer reinforcement learning module includes: Cross-domain feature alignment unit: extracts transferable structured graph features from source domain devices and maps them to the target domain, achieving feature distribution consistency through an adversarial mechanism; Strategy Optimization Learning Unit: Learn how to predict the probability of failure and its development process based on the time / space evolution path in the graph structure; Diagnostic decision generation and verification unit: Outputs clear diagnostic results and ensures their availability and rationality through dual verification of physical rules and historical trends.

9. The online status monitoring and fault diagnosis system for machine equipment according to claim 7, characterized in that: The domain adversarial training module in S31 specifically includes: Feature Extractor: Extracts structured features from the input fault evolution graph, taking into account both time series patterns and spatial topology information; Gradient reversal layer: keeps the features unchanged during forward propagation and negates the gradient during backpropagation to enhance the feature extractor against the domain classifier; Domain classifier: Determines whether the input feature belongs to the source domain or the target domain, and learns the difference in feature distribution; Domain adversarial loss function: The optimization goal is to minimize the domain classification error, while enabling the feature extractor to learn the representation of "indistinguishable domains" through GRL.

Images