A rotating machinery health situation awareness system based on vibration wireless sensing

Abstract

The present invention discloses a rotating machinery health situation perception system based on vibration wireless sensing, which belongs to the technical field of equipment monitoring and analysis; it is used to solve the technical problems of poor data processing and system robustness of existing solutions; through the reliable conversion from original vibration signal acquisition to high-quality feature input, the system's perception sensitivity to early faults is significantly improved; by constructing a feature dynamic weighted matrix and weighted feature vector, the accuracy of rotating machinery health situation perception can be effectively improved; through a pre-built health perception model, end-to-end intelligent evaluation of the health status of rotating machinery can be achieved: from dynamic weighted feature input to health level output, the health assessment accuracy and fault early warning rate under variable working conditions can be effectively improved; multi-node data privacy protection is achieved through federated learning, and the dynamic trust weight improves the data reliability screening capability, and finally outputs a robustness-optimized health situation perception result.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment monitoring and analysis, and in particular to a rotating machinery health situation awareness system based on vibration wireless sensing. Background Art

[0002] Rotating machinery health awareness is a technology that uses wireless vibration sensors installed on rotating machinery to monitor its operating status in real time. This technology collects and analyzes vibration signals from the machinery during operation to assess the equipment's health and identify potential faults and their severity.

[0003] Existing technical solutions, some of which adjust federated learning weights by introducing data quality scores, fail to incorporate dynamic updates of device health status. Others employ retransmission mechanisms to compensate for packet loss, but fail to consider the impact of latency on model timeliness. However, achieving robust health situational awareness by integrating node health, data quality, and transmission reliability within a federated learning framework remains a key unresolved technical challenge. Summary of the Invention

[0004] The purpose of the present invention is to provide a rotating machinery health situation awareness system based on vibration wireless sensing, which is used to solve the technical problems of poor data processing and system robustness in existing solutions.

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] A rotating machinery health situation awareness system based on vibration wireless sensing, comprising:

[0007] Signal acquisition and processing module: Build a distributed vibration wireless sensor network to collect vibration signals from key components of rotating machinery, perform time synchronization calibration on the collected vibration signals, and perform preprocessing of joint time-frequency domain denoising based on an adaptive noise threshold. The kurtosis value K of the original vibration signal is calculated. The kurtosis value K reflects the impact characteristics. The calculation formula is: Where, is the vibration signal sampling point, μ is the signal mean, σ is the signal standard deviation, and N is the number of sampling points;

[0008] Signal extraction and fusion module: Based on the preprocessed vibration signal, it extracts a multidimensional feature set including rotation frequency harmonic features, impact pulse features, and nonlinear dynamic features, and constructs a feature dynamic weighted matrix and weighted feature vector based on the real-time operating parameters of the machine.

[0009] Health status perception module: Inputs the dynamically weighted multidimensional feature set into the pre-built health perception model and outputs the real-time health status assessment value of the rotating machinery;

[0010] Perception optimization processing output module: Build a multi-node data fusion framework based on federated learning, dynamically adjust the trust weight of each sensor node according to the health status assessment value, and output robustness-optimized health situation perception results.

[0011] Preferably, the time-frequency domain joint denoising preprocessing adopts a method combining improved ensemble empirical mode decomposition and wavelet threshold denoising, and screens effective intrinsic mode components through the kurtosis-correlation coefficient joint criterion.

[0012] Preferably, the dominant noise component is removed, and the filtered intrinsic mode component is denoised using an improved soft threshold function, the function expression is:

[0013] Where, is the wavelet coefficient. Compared with the traditional soft threshold, this function has better continuity and avoids signal distortion.

[0014] The denoised intrinsic mode components are superimposed to obtain the preprocessed vibration signal.

[0015] Preferably, the harmonic characteristics of the rotation frequency include the fundamental frequency amplitude , harmonic energy ratio , phase difference ;

[0016] Shock pulse characteristics, including crest factor , Kurtosis K, Pulse Index , impact duration ;

[0017] Nonlinear dynamic characteristics, including approximate entropy AE, sample entropy SE, and fractal dimension .

[0018] Preferably, according to the correlation between the mechanical operating parameters and the feature sensitivity, the weight of each feature is dynamically adjusted to construct a weighted feature vector ; Where W is the weight matrix and F is the original eigenvector.

[0019] Preferably, when constructing the health perception model, it is constructed based on the deep belief network DBN, and a 4-layer deep belief network model is designed, including 3 layers of restricted Boltzmann machines RBM and 1 layer of back propagation output layer.

[0020] Preferably, when fine-tuning the target domain of the model, the parameters of the first two layers of RBM are frozen to retain the general feature extraction capability, and only the parameters of the third layer of RBM and the output layer are fine-tuned. The target labeled data of the target machine is used for supervised learning, and the loss function adopts the mean square error (MSE): ; In the formula, yi is the real health score, annotated by experts; is the model prediction score; M is the number of labeled samples;

[0021] Also, fine-tune the pre-trained parameters, and set the fine-tuned model as the health perception model.

[0022] Preferably, when building a multi-node data fusion framework based on federated learning, it includes node type classification, communication protocol and security mechanism, and architecture parameter setting.

[0023] Compared with the existing solutions, the present invention achieves the following beneficial effects:

[0024] The present invention reliably converts raw vibration signal acquisition into high-quality feature input, providing a robust data foundation for subsequent health status assessment. This significantly improves the system's sensitivity to early faults, especially addressing the non-stationarity and impact of rotating machinery vibration signals and noise interference from wireless sensors. By constructing a dynamic feature weighting matrix and weighted feature vectors, the conversion from raw signals to highly recognizable feature vectors is achieved, preserving fault-sensitive information while also eliminating interference from operating condition fluctuations and noise through dynamic weighting. This provides robust input for subsequent data analysis and effectively improves the accuracy of health status perception of rotating machinery. A pre-built health perception model enables end-to-end intelligent assessment of the health status of rotating machinery: from dynamic weighted feature input to health level output, no human intervention is required. This effectively improves the accuracy of health assessments and the early warning rate of faults under variable operating conditions, significantly reducing the risk of sudden mechanical failures. Multi-node data privacy protection is achieved through federated learning, and dynamic trust weights enhance data reliability screening capabilities. Ultimately, robustly optimized health status perception results are output, effectively improving the accuracy of health assessments and the robustness of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The present invention will be further described below with reference to the accompanying drawings.

[0026] Figure 1 This is a flowchart of the operation of a rotating machinery health situation awareness system based on vibration wireless sensing in the present invention. DETAILED DESCRIPTION

[0027] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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.

[0028] like Figure 1As shown, the present invention is a rotating machinery health situation awareness system based on vibration wireless sensing, comprising:

[0029] Build a distributed vibration wireless sensor network to collect vibration signals from key components of rotating machinery, perform time synchronization calibration on the collected vibration signals, and perform preprocessing for joint time-frequency domain denoising based on an adaptive noise threshold. The specific steps include:

[0030] When deploying sensor nodes, MEMS vibration acceleration sensors are installed on the surfaces of key rotating machinery components, with a sampling frequency of 1kHz-10kHz. Key rotating machinery components, such as bearing seats, gearboxes, and motor end covers, are configured with a number of nodes based on the complexity of the mechanical structure, typically 3-8 nodes. A star-mesh hybrid networking topology is used, and wireless communication is achieved through the ZigBee / Bluetooth low-power protocol.

[0031] In addition, each sensor node integrates a microcontroller, low-noise amplifier, 16-bit ADC, and lithium battery, supporting local edge computing preprocessing such as mean filtering and outlier clipping.

[0032] In the embodiment of the present invention, multi-node deployment can simultaneously collect vibration signals from different parts of the rotating machinery, such as the horizontal / vertical directions of the bearing and the gear meshing point, which can effectively avoid the one-sidedness of single-point monitoring;

[0033] When performing time synchronization calibration, the IEEE 1588 PTP protocol is used to achieve master-slave node time synchronization, and the synchronization accuracy is controlled within ±10μs;

[0034] The master node is deployed on a fixed mechanical base, where vibration interference is minimized. It uses GPS timing or a local crystal oscillator to provide a reference clock and performs a synchronization check every P minutes, where P is a positive integer and can be 5. Based on the periodic vibration characteristics of rotating machinery, the time deviation caused by wireless transmission delay or crystal oscillator drift is corrected by calculating the mutual correlation coefficient of the signals collected by each node. The calculation formula is: Where, is the time deviation from node i; is the master node reference signal, is the signal from the i-th slave node, Search delay range, specifically ±500μs;

[0035] In an embodiment of the present invention, through this time synchronization calibration, it is possible to ensure that multi-node signals are aligned on the time axis, providing a reliable basis for subsequent feature fusion; through real-time calibration of cross-correlation, the crystal oscillator drift and wireless transmission delay in long-term operation can be offset, which can solve the problem of traditional static synchronization failing over time.

[0036] When performing preprocessing for joint time-frequency domain denoising based on the adaptive noise threshold, the kurtosis value K of the original vibration signal is calculated. The kurtosis value K reflects the impact characteristics. For normal mechanical signals, K≈3, while for fault signals, K>5. The calculation formula is: Where, is the vibration signal sampling point, μ is the signal mean, σ is the signal standard deviation, and N is the number of sampling points;

[0037] Use Donoho threshold formula to calculate the initial noise threshold , the formula is: ;

[0038] In order to avoid the fault impulse signal being over-suppressed, the initial noise threshold is adjusted by the kurtosis K. Make dynamic adjustments: Where α is the empirical correction coefficient, which can be set to 0.1; (K-3) is the kurtosis deviation. For normal signals, K-3≈0, and the threshold remains unchanged; for fault signals, K-3>0, the threshold increases, and the impact component is retained.

[0039] During signal preprocessing, each time a segment of data is processed, such as a 1-second vibration signal, corresponding to N = 1000 sampling points, the kurtosis value K is recalculated and the noise threshold λ is updated to ensure that the noise threshold changes dynamically with the signal characteristics;

[0040] Improved ensemble empirical mode decomposition is used to decompose the signal into 8-12 intrinsic mode components. The added white noise intensity is 0.2 times the signal standard deviation, and the ensemble average is 20 times to suppress modal aliasing.

[0041] The effective eigenmode components are screened by the kurtosis-correlation coefficient joint criterion; specifically, the components with kurtosis K>3 and correlation coefficient r>0.5 with the original signal are retained, and the noise-dominated components are eliminated;

[0042] The improved soft threshold function is used to denoise the filtered intrinsic mode components. The function expression is:

[0043] Where, is the wavelet coefficient. Compared with the traditional soft threshold, this function has better continuity and avoids signal distortion.

[0044] The denoised intrinsic mode components are superimposed to obtain the preprocessed vibration signal.

[0045] It should be noted that the dynamic threshold based on kurtosis can automatically adjust the denoising intensity according to the signal characteristics, solving the contradiction between noise suppression and signal fidelity of the fixed threshold. By improving the method of ensemble empirical mode decomposition combined with wavelet threshold, it can not only decompose the non-stationary impact components in the vibration signal, such as the periodic impact of bearing failure, but also filter out the Gaussian white noise introduced by wireless transmission, which can effectively improve the signal-to-noise ratio.

[0046] In the embodiment of the present invention, by reliably converting the original vibration signal collection into high-quality feature input, a robust data basis is provided for subsequent health status assessment. In particular, the system's sensitivity to early fault perception is significantly improved, especially for the non-stationarity and impact of rotating machinery vibration signals and the noise interference of wireless sensors.

[0047] Based on the preprocessed vibration signal, a multidimensional feature set including rotation frequency harmonic features, impact pulse features, and nonlinear dynamic features is extracted, and a feature dynamic weighted matrix and weighted feature vector are constructed in combination with the real-time operating parameters of the machine; the real-time operating parameters include but are not limited to speed and load;

[0048] Among them, the rotation frequency harmonic characteristics, including the fundamental frequency amplitude , harmonic energy ratio , phase difference ;

[0049] Shock pulse characteristics, including crest factor , Kurtosis K, Pulse Index , impact duration ;

[0050] Nonlinear dynamic characteristics, including approximate entropy AE, sample entropy SE, and fractal dimension ;

[0051] The extraction of multi-dimensional characteristic indicators from vibration signals is an existing conventional technical means, and the specific implementation steps are not detailed here;

[0052] It should be noted that by collecting the harmonic characteristics of the rotation frequency, the impact pulse characteristics, and the nonlinear dynamic characteristics, it is possible to capture the sensitive characteristics of different types of faults, such as the impact of bearing faults, the harmonic characteristics of gear faults, and the periodicity of rotor imbalance, which can effectively solve the problem of missed detection due to a single feature.

[0053] When collecting and normalizing the real-time operating parameters of a machine, the operating parameters are obtained in real time through sensor nodes or the machine control system. The real-time operating parameters include speed n, load L, and ambient temperature T. The speed is measured in r / min, with a sampling frequency of 1 Hz. The load is measured in %, such as motor current percentage or torque percentage. The ambient temperature is measured in °C to compensate for the impact of temperature on vibration signals.

[0054] Map the operating parameters to the [0,1] interval to eliminate dimension differences:

[0055] , Where, 、 The minimum and maximum speeds for normal operation of the machine; 、 is the minimum and maximum load;

[0056] According to the correlation between mechanical operating parameters and feature sensitivity, the weight of each feature is dynamically adjusted to construct a weighted feature vector ;W is the weight matrix and F is the original eigenvector; specifically:

[0057] By formula Calculate the initial weight of feature i Where, is the balance coefficient, the value range is [0, 1], and the default value is 0.7; 、 They are sensitivity value and anti-interference value respectively;

[0058] in, Where, is the mean value of feature i under fault condition, 、 is the mean and standard deviation of feature i under normal conditions; The larger it is, the more sensitive feature i is to faults;

[0059] Where, is the Pearson correlation coefficient between feature i and speed, is the correlation coefficient between feature i and load; ∈[0,1], the closer it is to 1, the less the characteristic is affected by operating condition fluctuations;

[0060] Preliminary weights Normalize it to the interval [0,1] to ensure that the weight matrix satisfies , m is the total number of features, which avoids the impact of numerical magnitude differences on subsequent model inputs;

[0061] Among them, the weight coefficient ; is the anti-interference item. The lower the correlation with the speed fluctuation, the stronger the anti-interference ability and the higher the weight.

[0062] Construct the final feature weight matrix W and multiply it with the original feature vector F to obtain the weighted feature vector ; .

[0063] It should be noted that by dynamically adjusting the weights of sensitivity and anti-interference values, fault features, such as high-kurtosis impulses, are highlighted, while features that are greatly affected by operating condition fluctuations, such as low-order harmonic amplitudes, are suppressed. This can effectively improve the model's adaptability to complex operating conditions.

[0064] In an embodiment of the present invention, by constructing a feature dynamic weighting matrix and a weighted feature vector, the conversion from the original signal to a high-recognition feature vector is achieved. This not only retains the fault-sensitive information, but also eliminates the interference of operating condition fluctuations and noise through dynamic weighting, providing robust input for subsequent data analysis, and can effectively improve the accuracy of the health status perception of rotating machinery.

[0065] The dynamically weighted multidimensional feature set is input into the pre-built health perception model to output the real-time health status assessment value of the rotating machinery;

[0066] Among them, when building the health perception model, it is built based on the deep belief network DBN. A 4-layer deep belief network model is designed, including 3 layers of restricted Boltzmann machines (RBMs) and 1 layer of back propagation output layer;

[0067] The input is the weighted feature vector obtained from the previous processing ; The value range of dimension m is 8-10; the output is the health status level, specifically 0-100 points, the higher the score, the better the health;

[0068] The number of nodes at each layer is configured as follows:

[0069] Input layer: number of nodes = feature dimension m, for example 8 nodes;

[0070] RBM1 layer: number of nodes = 2m, for example, 16 nodes, extracting low-order local features;

[0071] RBM2 layer: number of nodes = 1.5m, for example, 12 nodes, fusing middle-level features;

[0072] RBM3 layer: number of nodes = m, for example 8 nodes, abstracting high-order semantic features;

[0073] Output layer: 1 node, used to output health status score;

[0074] The contrastive divergence algorithm is used to train each layer of RBM, such as CD-1. CD-1 is a specific form of contrastive divergence. It refers to the process of performing only one Gibbs sampling step during the sampling process, using the output of the previous layer of RBM as the input of the next layer to minimize the reconstruction error:

[0075] ; Where E is the mathematical expectation; v is the visible layer vector, i.e., the input feature; h is the hidden layer vector, i.e., the extracted feature; pdata is the data distribution; Indicates randomly extracting samples v from the training data; Represents the edge probability of generating the visible layer vector v; represents the logarithm of the marginal probability; Represents the conditional distribution p(h|v) of the hidden layer vector h given v; represents the conditional probability distribution of the visible layer v given the hidden layer h; Represents the logarithm of the reconstruction probability, which is used to quantify the reconstruction effect;

[0076] It should be noted that the reconstruction error It consists of two parts, It is a negative log-likelihood term, which can promote the model to learn the true distribution of data and make the generated visible layer vector v as close to the training data as possible;

[0077] To reconstruct the log-likelihood term, the model can be pushed to accurately restore the input v through the hidden layer h, ensuring that the features extracted by the hidden layer contain sufficient information;

[0078] By minimizing the reconstruction error , which enables RBM to learn low-dimensional, highly discriminative features of input data in an unsupervised manner, such as fault shock patterns in rotating machinery vibration signals, and provide reliable feature input for the health status assessment of subsequent health perception models.

[0079] In addition, the pre-training parameters are set as follows: learning rate η = 0.01, batch size batch size = 32, number of iterations 100 rounds, momentum factor α0 = 0.9 to accelerate convergence and suppress oscillation;

[0080] When fine-tuning the target domain of the model, the parameters of the first two RBM layers are frozen to retain the general feature extraction capability. Only the parameters of the third RBM layer and the output layer are fine-tuned. Supervised learning is performed using the target annotated data of the target machine. The target annotated data includes 200 groups of normal samples and 100 groups of faulty samples. The loss function uses the mean square error (MSE): ; In the formula, yi is the real health score, annotated by experts; is the model prediction score; M is the number of labeled samples;

[0081] And, fine-tune the pre-training parameters, set the learning rate η to 0.001 to avoid destroying the pre-training features; set the number of iterations to 50 rounds; set the regularization coefficient λ to ;

[0082] The fine-tuned model is set as the health perception model;

[0083] It is important to note that the three-layer RBM automatically extracts multi-level features from local impact features to global fault patterns through layer-by-layer unsupervised learning, avoiding the subjectivity of manual feature selection. For example, traditional methods rely on expert experience to select features, which may miss key fault information.

[0084] The CD-1 algorithm can be used to more efficiently optimize RBM parameters and effectively reduce the model reconstruction error after pre-training. By freezing the underlying parameters, the general feature extraction capability can be retained, and the high-level parameters can be fine-tuned to adapt to the target equipment, so that the model can work stably on rotating machinery of different brands and models.

[0085] When constructing the threshold interval, in the initial stage of trouble-free operation of the machine, for example, the first three months, 1000 sets of weighted feature vectors under normal conditions are collected. , input the trained health perception model to obtain the normal health score distribution ; Among them, the mean ≈90 points, standard deviation ≈3 points;

[0086] Set the health status threshold interval according to the 3σ principle:

[0087] Health: Score ≥ , for example ≥87 points;

[0088] Sub-health: <Rating< , for example 84-87 points;

[0089] Minor fault: <Rating≤ , for example 81-84 points;

[0090] Serious failure: score ≤ , for example 81 points;

[0091] Automatically input new weighted feature vectors every 5 minutes The health perception model outputs a health score and compares it with the dynamic threshold to generate an assessment result. For example, if the current health status is sub-healthy, it is recommended to shut down the machine for inspection within 1 week.

[0092] It should be noted that learning the threshold based on the machine's own normal state can effectively avoid false positives caused by using a fixed threshold. For example, the normal score of a new device is generally ≥95 points, while an old device may only score 88 points. A fixed threshold will mistakenly identify the old device as sub-healthy.

[0093] In an embodiment of the present invention, a pre-built health perception model can be used to achieve end-to-end intelligent assessment of the health status of rotating machinery: from dynamic weighted feature input to health level output, without the need for human intervention, it can effectively improve the accuracy of health assessment and early fault warning rate under variable working conditions, significantly reducing the risk of sudden mechanical failure.

[0094] A multi-node data fusion framework based on federated learning is constructed to dynamically adjust the trust weight of each sensor node according to the health status assessment value, and output robustness-optimized health situation awareness results.

[0095] When building a multi-node data fusion framework based on federated learning, it includes node type classification, communication protocol and security mechanism, and architecture parameter setting;

[0096] Among them, when dividing the node types, it includes edge sensor nodes and aggregation servers;

[0097] Edge sensor nodes: Sensor terminals deployed locally on machines, such as vibration and temperature sensors, responsible for local data collection, preprocessing, and model training;

[0098] Aggregation server: A centralized node responsible for global model initialization, local model aggregation, and trust weight management, without direct access to raw data.

[0099] Regarding communication protocols and security mechanisms: Federated learning communication protocols, such as FedML, are used. Edge nodes and aggregation servers transmit model parameters through encrypted channels, such as TLS 1.3.

[0100] Gaussian noise is added before uploading local model parameters, where the noise intensity σ=0.01 to prevent the aggregation server from inferring the original data;

[0101] When setting the architecture parameters, the number of nodes K ranges from 5 to 10 and is adjusted according to the scale of the machine. For example, a large unit can deploy 10 nodes.

[0102] The communication cycle T = 5 minutes, which is synchronized with the health assessment cycle;

[0103] Differential privacy parameters =1.0, privacy budget. The smaller the value, the stronger the privacy protection. The default value is industrial-grade security.

[0104] It should be noted that by building a multi-node data fusion framework based on federated learning, the data privacy protection and fusion accuracy problems of multiple sensor nodes can be solved. Federated learning can be used to achieve data static model dynamics, and the robustness of health situation awareness can be improved by combining dynamic trust weights.

[0105] When dynamically adjusting the trust weight of each sensor node according to the health status evaluation value, the trust weight of each node i' is calculated using the health status evaluation value, signal-to-noise ratio and historical contribution , the calculation formula is:

[0106] Where, is the health status evaluation value of node i'; is the set of values ​​evaluated for all health states; is the signal-to-noise ratio of node i'; is the set of all signal-to-noise ratios; The historical contribution of node i', which can be the improvement in the accuracy of the past model; is the set of all historical contributions; a, b, and c are all weight coefficients greater than 0 and less than 1, and a+b+c=1, with a=0.5, b=0.3, and c=0.2 by default; j is the sum index variable, indicating traversing all nodes from the 1st node to the Kth node, where K is the total number of nodes;

[0107] When the health evaluation value of node i´ When the score is less than 85, that is, the sub-health threshold, the weight a is automatically reduced to 0.3 to suppress interference from abnormal nodes;

[0108] When the node i´ has three consecutive cycles of data quality When <20dB, the corresponding trust weight will be temporarily Set to 0 to exclude invalid data.

[0109] When performing robustness optimization, the aggregation server aggregates the local model based on the trust weight and the amount of node data. , the formula is: Where, is the updated global model parameter; The amount of node data, such as the number of data collected by sensors deployed on the node and the number of training samples;

[0110] Calculate the parameter differences between each local model and the global model If the difference exceeds 3 times the standard deviation of the mean, the node weight Temporarily set to 0 to resist attacks from malicious nodes;

[0111] The aggregation server will global model parameters Send to each node;

[0112] Each node fine-tunes the local model based on the received data. For example, the learning rate η is set to 0.001, and iterates for 10 rounds to generate a new local model.

[0113] Repeat the above steps until the global model converges, which is determined by the accuracy fluctuation being <1% for three consecutive iterations;

[0114] The health status evaluation value of each node Weighted fusion to obtain the final health status assessment value : ;

[0115] And the output robustness indicator: the variance of the fusion result , the smaller the variance, the more reliable the result.

[0116] In an embodiment of the present invention, multi-node data privacy protection is achieved through federated learning, dynamic trust weights are used to improve data reliability screening capabilities, and ultimately robustness-optimized health situation awareness results are output, which can effectively improve the accuracy of health assessments and the robustness of the system.

[0117] In the several embodiments provided by the present invention, it should be understood that the disclosed system can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative. For example, the division of modules is only a logical function division, and other division methods may be used in actual implementation.

[0118] Modules described as separate components may or may not be physically separate, and components shown as modules may or may not be physical modules, and may be located in one place or distributed across multiple network modules. Some or all of these modules may be selected to achieve the objectives of this embodiment based on actual needs.

[0119] In addition, the functional modules in various embodiments of the present invention may be integrated into a single processing module, each module may exist physically separately, or two or more modules may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or hardware plus software functional modules.

[0120] It is obvious to a person skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, but that the present invention can be implemented in other specific forms without departing from the essential characteristics of the present invention.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A rotating machinery health situation awareness system based on vibration wireless sensing, characterized in that: include: Signal acquisition and processing module: Build a distributed vibration wireless sensor network to collect vibration signals from key components of rotating machinery, perform time synchronization calibration on the collected vibration signals, and pre-process the time-frequency domain joint denoising based on the adaptive noise threshold. When pre-processing the time-frequency domain joint denoising based on the adaptive noise threshold, the Donoho threshold formula is used to calculate the initial noise threshold. , the initial noise threshold is adjusted by the kurtosis value K Make dynamic adjustments: Where α is the empirical correction coefficient; (K-3) is the kurtosis deviation. For normal signals, K-3≈0, and the initial noise threshold remains unchanged. For fault signals, K-3>0, the initial noise threshold increases, and the impact component is retained. The kurtosis value K of the original vibration signal is calculated. The kurtosis value K reflects the impact characteristics. The calculation formula is: Where, is the vibration signal sampling point, μ is the signal mean, σ is the signal standard deviation, and N is the number of sampling points; Signal extraction and fusion module: Based on the preprocessed vibration signal, it extracts a multidimensional feature set including rotation frequency harmonic features, impact pulse features, and nonlinear dynamic features, and constructs a feature dynamic weighted matrix and weighted feature vector based on the real-time operating parameters of the machine. Health status perception module: Inputs the dynamically weighted multidimensional feature set into the pre-built health perception model and outputs the real-time health status assessment value of the rotating machinery; Perception optimization processing output module: Build a multi-node data fusion framework based on federated learning, and dynamically adjust the trust weight of each sensor node according to the health status assessment value, including using the health status assessment value, signal-to-noise ratio and historical contribution to calculate the trust weight of each sensor node i´ , the calculation formula is: Where, is the health status evaluation value of sensor node i´; is the set of values ​​evaluated for all health states; is the signal-to-noise ratio of the sensor node i´; is the set of all signal-to-noise ratios; is the historical contribution of sensor node i´; is the set of all historical contributions; a, b, and c are all weight coefficients greater than 0 and less than 1, and a + b + c = 1; j is the summation index variable, indicating traversal of all nodes from the 1st node to the Kth node, where K is the total number of nodes; the output is the robustness-optimized health situation awareness result.

2. A rotating machinery health situation awareness system based on vibration wireless sensing according to claim 1, characterized in that: The time-frequency domain joint denoising preprocessing adopts the method of combining improved ensemble empirical mode decomposition with wavelet threshold denoising, and the effective intrinsic mode components are screened by the kurtosis-correlation coefficient joint criterion.

3. The rotating machinery health situation awareness system based on vibration wireless sensing according to claim 2 is characterized in that: Eliminate the dominant noise component and use the improved soft threshold function to denoise the filtered intrinsic mode component. The function expression is: Where, is the wavelet coefficient; The denoised intrinsic mode components are superimposed to obtain the preprocessed vibration signal.

4. The rotating machinery health situation awareness system based on vibration wireless sensing according to claim 3 is characterized in that: Rotation frequency harmonic characteristics, including fundamental frequency amplitude , harmonic energy ratio , phase difference ; Shock pulse characteristics, including crest factor , Kurtosis K, Pulse Index , impact duration ; Nonlinear dynamic characteristics, including approximate entropy AE, sample entropy SE, and fractal dimension .

5. The rotating machinery health situation awareness system based on vibration wireless sensing according to claim 4 is characterized in that: According to the correlation between mechanical operating parameters and feature sensitivity, the weight of each feature is dynamically adjusted to construct a weighted feature vector ; Where W is the weight matrix and F is the original eigenvector.

6. The rotating machinery health situation awareness system based on vibration wireless sensing according to claim 5 is characterized in that: When building the health perception model, it is built based on the deep belief network DBN. A 4-layer deep belief network model is designed, which includes 3 layers of restricted Boltzmann machines (RBMs) and 1 layer of back propagation output layer.

7. The rotating machinery health situation awareness system based on vibration wireless sensing according to claim 6 is characterized in that: When fine-tuning the target domain of the model, the parameters of the first two layers of RBM are frozen to retain the general feature extraction capability. Only the parameters of the third layer of RBM and the output layer are fine-tuned. The target labeled data of the target machine is used for supervised learning, and the loss function is the mean square error (MSE): ; In the formula, yi is the real health score, annotated by experts; is the model prediction score; M is the number of labeled samples; Also, fine-tune the pre-trained parameters, and set the fine-tuned model as the health perception model.

8. The rotating machinery health situation awareness system based on vibration wireless sensing according to claim 7 is characterized in that: When building a multi-node data fusion framework based on federated learning, it includes node type classification, communication protocol and security mechanism, and architecture parameter setting.

Images