Big data-based multi-device state intelligent evaluation diagnosis method and system

Abstract

The application provides a big data-based multi-device state intelligent evaluation and diagnosis method and system, belongs to the field of multi-device state evaluation and diagnosis, and is used for solving the problems of multi-source time-varying interference covering weak fault characteristics, insufficient generalization ability in small sample scenes, and difficulty of fixed models in adapting to changes in the whole life cycle of devices in the related art. The method collects multi-source data and pre-processes, fuses to form a diagnosis data set, and generates a diagnosis result through dynamic interference cancellation, weak feature enhancement, cross-device knowledge transfer and health index driven optimization. The system includes a sensor, an edge computing node and a cloud server. It can improve the diagnosis accuracy, adaptability and early fault identification ability.

Description

Technical Field

[0001] This application relates to the field of multi-device status assessment and diagnosis, and in particular to a method and system for intelligent assessment and diagnosis of multi-device status based on big data. Background Technology

[0002] In industrial production, multi-device collaborative operation is widespread, and the stability of equipment status directly affects production efficiency and safety. Therefore, accurate assessment and diagnosis of multi-device status is of great significance. Current technologies for multi-device status assessment and diagnosis largely rely on the status data of individual devices, employing simple data filtering and fixed models for analysis, resulting in a relatively crude processing method. However, real-world industrial environments are subject to multi-source time-varying interference, such as equipment operating noise, power grid fluctuations, and environmental changes. This interference can easily mask weak fault characteristics of equipment. Furthermore, the generalization ability of equipment features is insufficient in small-sample scenarios, and the status changes throughout the equipment's entire lifecycle are difficult to adapt to fixed models, leading to poor diagnostic accuracy and adaptability, and failing to meet actual production needs. Summary of the Invention

[0003] This application provides a multi-device status intelligent assessment and diagnosis method and system based on big data, which can comprehensively process multi-source data, effectively cope with multi-source interference, improve the ability to extract weak fault features and the ability to generalize in small sample scenarios, adapt to changes in the entire life cycle of equipment, and improve diagnostic accuracy and adaptability.

[0004] Firstly, this application provides a multi-device status intelligent assessment and diagnosis method based on big data. It includes the following steps: real-time collection of status data, interference source data, and control signal data from multiple devices using edge computing nodes, followed by preprocessing; construction of a multi-dimensional data fusion framework to fuse the preprocessed multi-source data into a unified diagnostic dataset; adoption of a multi-source time-varying interference dynamic cancellation system to perform layered processing of interference in the diagnostic data, including pre-reaction prediction compensation, dynamic feature pruning, and device coupling interference cancellation; design of a weak fault feature enhancement mechanism, extracting weak fault features through noise co-occurrence utilization algorithms and adaptive mode decomposition; construction of a federated learning cross-device transfer module to perform knowledge transfer between multiple devices, improving feature generalization ability in small sample scenarios; and establishment of a health index-driven diagnostic optimization mechanism to dynamically adjust model parameters according to device health status and generate diagnostic results.

[0005] By adopting the above technical solutions, it is possible to comprehensively collect and integrate multi-dimensional data, process multi-source time-varying interference in a hierarchical manner, enhance the extraction of weak fault features, realize cross-device knowledge transfer, and dynamically adapt to changes in equipment health status, thereby comprehensively improving the accuracy and adaptability of multi-device status assessment and diagnosis, and effectively meeting the diagnostic needs in complex industrial environments.

[0006] Furthermore, the construction steps of the multi-dimensional data fusion framework include: performing protocol conversion and data cleaning on device status data, interference source data, and control signal data; and adopting a spatiotemporal synchronization mechanism to keep the timestamps of multi-source data consistent.

[0007] By adopting the above technical solutions, the consistency and validity of multi-source data can be ensured, laying a reliable data foundation for subsequent interference processing and feature extraction, and reducing diagnostic errors caused by data inconsistency.

[0008] Furthermore, the processing steps of the multi-source time-varying interference dynamic cancellation system include: generating an anti-phase compensation signal based on a time series prediction model through a pre-action prediction compensation module to cancel the predicted interference; using a dynamic feature trimming module to adaptively trim frequency domain features according to the signal-to-noise ratio and retain the weak feature sensitive region; and using a device coupling interference cancellation module to cancel cross-device interference through dynamic coupling modeling and transmission delay compensation.

[0009] By adopting the above technical solutions, multi-source time-varying interference can be processed from multiple levels, including prediction compensation, frequency domain clipping, and cross-device interference cancellation, effectively reducing the impact of interference on fault characteristics and improving the accuracy of feature extraction.

[0010] Furthermore, the processing steps of the pre-reaction prediction compensation module include: predicting the intensity and trend of the interference based on a time series prediction model; generating a compensation signal with the opposite phase to the interference based on the prediction results to cancel the interference.

[0011] By adopting the above technical solutions, interference can be predicted in advance and compensation signals can be actively generated, thereby achieving proactive processing of interference and reducing the masking of weak fault characteristics by instantaneous interference.

[0012] Furthermore, the processing steps of the dynamic feature trimming module include: performing frequency domain transformation on the diagnostic data to obtain the signal-to-noise ratio of each frequency band; and trimming the interference-dominant frequency band according to the signal-to-noise ratio threshold, while retaining the frequency band where weak features are located.

[0013] By adopting the above technical solutions, it is possible to accurately identify and cut off the frequency bands that dominate interference, and selectively retain the frequency bands where weak fault characteristics are located, thereby improving the identification of weak characteristics.

[0014] Furthermore, the processing steps of the weak fault feature enhancement mechanism include: dividing noise into equipment-related noise and irrelevant noise through a noise co-occurrence utilization algorithm, and extracting weak features from the equipment-related noise; and using adaptive mode decomposition to dynamically adjust the number of modes according to the signal complexity to decompose the signal and extract weak fault features.

[0015] By adopting the above technical solution, weak feature information in equipment-related noise can be fully utilized, and the extraction effect of weak fault features can be improved by dynamically adjusting the mode decomposition parameters, thereby enhancing the ability to identify early faults.

[0016] Furthermore, the processing steps of the federated learning cross-device transfer module include: calculating the similarity and feature differences between devices to generate transfer weights; and performing knowledge transfer between multiple devices based on the transfer weights and feature distribution alignment algorithm, while preserving individual device differences.

[0017] By adopting the above technical solutions, knowledge can be effectively transferred between multiple devices, improving the generalization ability of device features in small sample scenarios, while preserving the individual characteristics of each device and improving the adaptability of diagnosis.

[0018] Furthermore, the processing steps of the health index-driven diagnostic optimization mechanism include: calculating the health index of the equipment by integrating equipment status data and historical diagnostic results; dividing the health stage of the equipment according to the health index; and adjusting the modal decomposition parameters and classifier weights in a coordinated manner.

[0019] By adopting the above technical solution, the model parameters can be dynamically adjusted based on the health status of the equipment, so that the diagnostic model can adapt to the status changes of the equipment throughout its entire life cycle and maintain the stability of diagnostic accuracy.

[0020] Furthermore, the health stage includes a health stage, a degradation stage, and a near-failure stage, with different mode decomposition parameters and classifier weight adjustment strategies corresponding to different stages.

[0021] By adopting the above technical solutions, differentiated parameter adjustment strategies can be formulated for the characteristics of different health stages of the equipment, which can more accurately adapt to changes in equipment status and further improve the diagnostic effect.

[0022] Secondly, this application provides a multi-device status intelligent assessment and diagnosis system based on big data. It includes sensors, edge computing nodes, and a cloud server for deployment in industrial scenarios where multiple devices operate collaboratively. The system is used to execute the multi-device status intelligent assessment and diagnosis method based on big data as described in any of the first aspects above. The sensors are used to collect status data, interference source data, and control signal data from multiple devices; the edge computing nodes are used for real-time data processing; and the cloud server is used for storing and analyzing big data.

[0023] By adopting the above technical solutions, hardware support can be provided for intelligent assessment and diagnosis of the status of multiple devices, realizing the integration of data collection, processing, storage and analysis, ensuring the effective execution of the method, and thus achieving accurate assessment and diagnosis of the status of multiple devices.

[0024] In summary, this application has at least the following beneficial effects:

[0025] 1. A method and system for intelligent assessment and diagnosis of the status of multiple devices based on big data is provided, which can effectively improve the accuracy and adaptability of diagnosis;

[0026] 2. It can process multi-source time-varying interference in a hierarchical manner, enhancing the ability to extract weak fault features;

[0027] 3. It can achieve cross-device knowledge transfer, adapt to changes throughout the device's lifecycle, and improve generalization ability in small sample scenarios.

[0028] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0029] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0030] Figure 1 The diagram illustrates a schematic of a multi-device status intelligent assessment and diagnosis system based on big data, according to an embodiment of this application.

[0031] Figure 2 A flowchart of a multi-device status intelligent assessment and diagnosis method based on big data is shown in an embodiment of this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0034] This application provides a multi-device status intelligent assessment and diagnosis method and system based on big data, which can accurately assess the status of multiple devices, effectively cope with multi-source time-varying interference, enhance the extraction of weak fault features, adapt to the entire life cycle of devices, improve the generalization ability of small sample scenarios, and significantly improve diagnostic efficiency.

[0035] In a first aspect, embodiments of this application disclose a multi-device status intelligent assessment and diagnosis system based on big data.

[0036] Figure 1 The diagram illustrates a schematic of a multi-device status intelligent assessment and diagnosis system based on big data, according to an embodiment of this application.

[0037] Reference Figure 1 The system comprises a hardware support architecture for intelligent assessment and diagnosis of multiple device states. This architecture consists of a sensor network, edge computing nodes, a cloud server, communication links, and heterogeneous device interface adaptation modules, all working together to construct a complete operating environment. The sensor network is distributed across industrial scenarios where multiple devices operate collaboratively. Vibration, temperature, current, and environmental sensors are deployed according to device type (e.g., rotating machinery, electrical equipment, fluid equipment) to collect device state data, interference source data, and control signal data, respectively. The acquisition frequency and accuracy are adapted to the characteristics and requirements of different devices. Edge computing nodes are deployed near the device cluster and connected to the sensor network via an industrial bus. They receive and preprocess the data collected by the sensors, including data cleaning, spatiotemporal synchronization, and preliminary interference suppression. They also undertake edge-side computing tasks for real-time diagnostic analysis, meeting low-latency processing requirements. The cloud server is deployed in the data center and connected to the edge computing nodes via an industrial Ethernet or 5G communication link. It receives preprocessed data and intermediate diagnostic results uploaded by the edge nodes, providing big data storage, model training, and deep analysis capabilities to support global diagnostic optimization and knowledge accumulation. The communication links include an industrial bus between sensors and edge nodes, and a wide area network between edge nodes and the cloud, ensuring the real-time and reliable transmission of data and adapting to different data volume requirements. A heterogeneous device interface adapter module is integrated into the edge computing node, compatible with the communication protocols and data formats of different devices, enabling unified access and data interaction for multiple types of devices. The various components of the above hardware support system work collaboratively through pre-defined connections: the sensor network provides raw data input, the edge computing nodes perform real-time processing, the cloud server performs in-depth analysis and optimization, the communication links ensure data flow, and the heterogeneous device interface adapter module resolves differences in device access. Together, they provide a stable and efficient operating environment for big data-based multi-device status intelligent assessment and diagnosis methods, ensuring the smooth execution of the diagnostic process.

[0038] Secondly, embodiments of this application disclose a multi-device status intelligent assessment and diagnosis method based on big data.

[0039] Figure 2 A flowchart of a multi-device status intelligent assessment and diagnosis method based on big data is shown in an embodiment of this application.

[0040] Reference Figure 2 The method specifically includes the following steps:

[0041] S1: Real-time acquisition of status data, interference source data, and control signal data from multiple devices using edge computing nodes, followed by preprocessing.

[0042] In this step, edge computing nodes are used to collect real-time status data, interference source data, and control signal data from multiple devices. The collected data is then preprocessed to lay the foundation for subsequent data fusion. Specifically, status data includes equipment vibration acceleration (sampling frequency range 1kHz-10kHz, with higher values ​​taken depending on equipment type, such as rotating machinery and lower values ​​taken for static equipment), temperature (sampling frequency 1Hz-10Hz), and current (sampling frequency 5kHz-20Hz), with collection points distributed at key parts of the equipment (such as bearing housings and motor terminals). Interference source data covers grid voltage (50Hz fundamental and harmonics, sampling frequency 2kHz), ambient temperature and humidity (sampling frequency 10Hz), and airflow disturbance (high-frequency acquisition is triggered when wind speed ≥3m / s, increasing the sampling frequency to 100Hz). Control signals include PLC commands (acquisition triggered by rising / falling edge, timestamp accuracy 1μs). All sensors are timestamped using a distributed clock synchronization protocol (IEEE1588PTP) to ensure that the time error between data from different sources is ≤100μs.

[0043] The preprocessing stage first performs detrending processing. For signals with linear trends (such as temperature drift), the trend term is fitted using the least squares method with a sliding window (window size N = 100 sampling points): Let the original signal be x(t), and the fitted trend term be... Where b is the slope and c is the intercept, by minimizing... Solve for the parameters; the detrended data is: Next, outlier handling is performed, identifying outliers based on the 3σ criterion: the mean μ and standard deviation σ of the data sequence are calculated. If a data point x(t) satisfies |x(t)-μ|>3σ, it is determined to be an outlier, and linear interpolation between two adjacent normal data points is used for replacement. For asynchronously acquired data (such as time misalignment caused by differences in sampling frequencies of different sensors), cubic spline interpolation is used to achieve time alignment: Let device i acquire a discrete time sequence. (n is the total number of sampling nodes, and n≥4), the corresponding sampling value is The goal is to align the data from all devices to a unified target time t0 (t0 satisfies t). k ≤t0≤t k+1 (where k = 0, 1, ..., n-2).

[0044] Since the core of cubic spline interpolation is to ensure the smoothness of the interpolation curve through piecewise cubic polynomials, it works within the interval [t]. k ,t k+1 Within this range, the data aligned to the target time t0 needs to be represented as a piecewise cubic polynomial:

[0045] x i (t0)=a k (t0-t k ) 3 +b k (t0-t k ) 2 +c k (t0-t k )+d k

[0046] Where a k b k c k d k Let be the coefficient of the cubic polynomial in the k-th segment. This coefficient needs to be solved by solving multiple sets of constraints to ensure that the interpolation curve satisfies the characteristic of second-order continuous differentiability, thereby preserving the smoothness of the original signal.

[0047] The specific constraints include four parts: first, the function value continuity constraint, at sampling node t... k At points (k = 1, 2, ..., n-1), the function values ​​of adjacent piecewise polynomials are equal, that is:

[0048] a k-1 (t k -t k-1 ) 3 +b k-1 (t k -t k-1 ) 2 +c k-1 (t k -t k-1 )+d k-1

[0049] =a k (t k -t k ) 3 +b k (t k -t k ) 2 +c k (tk -t k )+d k

[0050] The simplification reflects the core logic of continuous function values ​​at nodes, meaning that the function values ​​of adjacent segments are consistent at nodes; secondly, it addresses the first-order derivative continuity constraint, which involves finding the first derivative x′ of the interpolation formula. i (t)=3a k (tt k ) 2 +2b k (tt k )+c k Substitute the sampling node t k (k = 1, 2, ..., n-1), the first derivatives of adjacent piecewise polynomials must be equal:

[0051] 3a k-1 (t k -t k-1 ) 2 +2b k-1 (t k -t k-1 )+c k-1 =c k

[0052] Third, the second derivative continuity constraint is applied to the interpolation formula to obtain the second derivative x″. i (t)=6a k (tt k )+2b k Substitute the sampling node t k (k = 1, 2, ..., n-1), the second derivatives of adjacent piecewise polynomials must be equal:

[0053] 6a k-1 (t k -t k-1 )+2b k-1 =2b k

[0054] Fourthly, natural boundary conditions are defined. To avoid unnecessary fluctuations in the interpolation curve at both ends, the beginning and end intervals are set (k=0 corresponds to [t0,t1], k=n-2 corresponds to [t...]). n-2 ,t n-1 The second derivative of ]) is 0, meaning the left endpoint satisfies 6a0(t1-t0)+2b0=0, and the right endpoint satisfies

[0055] 6a n-2 (t n-1 -t n-2 )+2b n-2 =0.

[0056] Through the above 2n constraint equations (including n - 1 groups of continuous function values, n - 1 groups of continuous first-order derivatives, n - 1 groups of continuous second-order derivatives, and 2 boundary conditions, covering the complete solution logic), 4(n - 1) piecewise coefficients can be solved (4 coefficients for each piece, a total of n - 1 pieces), finally achieving the time alignment of asynchronous data of multiple devices, and ensuring that the aligned data retains the smooth characteristics of the original signal, providing high-quality input with consistent timing for subsequent multi-dimensional data fusion.

[0057] S2: Construct a multi-dimensional data fusion framework to fuse the preprocessed multi-source data to form a unified diagnostic dataset.

[0058] The method of this step specifically includes: performing protocol conversion and data cleaning on device status data, interference source data, and control signal data; adopting a spatio-temporal synchronization mechanism to make the timestamps of multi-source data consistent, and through these operations, effectively fuse the multi-source data to form a unified diagnostic dataset.

[0059] In the protocol conversion stage, for the heterogeneous communication protocols of different devices (such as Modbus, Profinet, OPCUA), the data is uniformly mapped to a standardized format through the protocol conversion module of the edge computing node. For numerical data (such as vibration acceleration, temperature), normalization processing is adopted to eliminate the dimension difference. The formula is: where x is the original data, x min and x max are the historical minimum and maximum values of this type of data (derived from the historical dataset during normal device operation), respectively. The range of the normalized data is mapped to the interval [0,1], which is convenient for cross-type data fusion. For discrete data such as control signals (such as the switch instructions of PLC), they are converted into numerical vectors through one-hot encoding. For example, the "start" instruction is encoded as [1,0], and the "stop" instruction is encoded as [0,1].

[0060] In the data cleaning link, redundancy and noise are further eliminated on the basis of preprocessing. For continuously missing data (missing duration ≤ 5 sampling periods), linear interpolation is used to fill it: Let the valid data before and after the missing data point be x(t1) and x(t2), then the calculation formula for the missing point x(t) (t1 < t < t2) is: x(t) = where t1 and t2 are the timestamps of the valid data, and t is the timestamp of the missing point. For high-frequency noise (such as the spike signal caused by electromagnetic interference), moving average filtering is adopted: where M is the window size (taking 5 - 10 according to the noise frequency characteristics, derived from the spectrum analysis result), and Δt is the sampling interval.

[0061] The spatiotemporal synchronization mechanism achieves timestamp consistency across multiple data sources through dynamic time calibration. Based on the IEEE 1588PTP protocol, edge computing nodes periodically (10ms period) receive master clock synchronization messages and calculate the deviation Δt = t between the local clock and the master clock. local -t master , where t local t represents the local time of the edge node. master The master clock time (sourced from a cloud-based time server) is used. The timestamps of each sensor's data are corrected based on the deviation value: t corr =t raw -Δt, where t raw For the sensor's original timestamp, t corr The corrected timestamps ensure that the time error of data from different sources is controlled within 100μs. For spatially distributed equipment data (such as equipment in different areas of a production line), the equipment's physical location is encoded (e.g., coordinates (x,y)) and associated with a unified spatial coordinate system to facilitate subsequent coupling interference analysis.

[0062] The data fusion phase employs a weighted fusion algorithm to integrate multi-dimensional data, with weights dynamically allocated based on data reliability. Where F(t) represents the fused diagnostic data, N represents the number of data types (e.g., status data, interference source data, control signals), and x i (t) represents the standardized value of the i-th type of data at time t. For dynamic weights (SNR) i (t) represents the signal-to-noise ratio of the i-th type of data at time t, derived from real-time signal quality monitoring. Through the above process, a time-aligned, format-consistent, and noise-suppressed diagnostic dataset is finally formed, providing high-quality input for subsequent interference processing and feature extraction.

[0063] S3: A multi-source time-varying interference dynamic cancellation system is adopted to perform hierarchical processing of interference in diagnostic data, including pre-action prediction compensation, dynamic feature trimming and equipment coupling interference cancellation.

[0064] The specific steps of this method include: generating an inverse compensation signal based on a time-series prediction model using a pre-reaction prediction compensation module to cancel the predicted interference; adaptively cropping frequency domain features according to the signal-to-noise ratio (SNR) using a dynamic feature trimming module to retain the sensitive region of weak features; and canceling cross-device interference through a device coupling interference cancellation module using dynamic coupling modeling and transmission delay compensation. Specifically, the pre-reaction prediction compensation module's processing steps include predicting the intensity and trend of interference based on a time-series prediction model, generating a compensation signal with the opposite phase to the interference based on the prediction results, and canceling the interference; the dynamic feature trimming module's processing steps include frequency domain transformation of the diagnostic data, obtaining the SNR of each frequency band, and cropping the dominant interference frequency band according to the SNR threshold, retaining the frequency band containing weak features.

[0065] In the pre-action prediction and compensation, the time series prediction model adopts a bidirectional LSTM network. The input is the interference sequence I(t-3Δt), I(t-2Δt), and I(t-Δt) of the past three sampling periods (Δt is the sampling interval, which is 50ms and is determined based on the statistical analysis of the interference change rate in the industrial scenario). The output is the predicted interference value for the next 50ms. The loss function for model training considers both the intensity of the disturbance and the trend of change: Where γ = 0.5 is the trend penalty coefficient (obtained through cross-validation optimization). This is the time derivative of the interference intensity, reflecting the rate of change of the interference. For electromagnetic interference (such as power grid harmonics), the formula for generating the compensation signal is: Where k is the impedance matching coefficient, which is solved using the least squares method. x(t) is the original signal; to address mechanical shock interference, LMS adaptive filtering is used to generate compensated vibration V. 补 (t), the filter coefficient update formula is w(n+1)=w(n)+μ·e(n)·x(n), where μ=0.01 is the step size factor (to ensure convergence stability), e(n=x(n)-V 补 (n) represents the error signal.

[0066] In dynamic feature trimming, the frequency domain transformation uses Short-Time Fourier Transform (STFT), with a window size of 256 sampling points (the corresponding time length is determined based on the sampling frequency to ensure coverage of more than two fault characteristic cycles), and an overlap rate of 50%. After calculating the power spectral density P(f) for each frequency band, the interference-dominant frequency band Finterference is defined as: Finterference = {f | P(f) > 0.7P}. max}(P max The maximum power spectral density is 0.7, which is an empirical threshold determined based on the statistical ratio of interference and weak characteristic energy. The signal-to-noise ratio (SNR)(f) is calculated using the following formula: Where S(f) is the signal power of frequency band f (the average power during normal equipment operation), and N(f) is the noise power of the current frequency band. The clipping rule is: if And SNR(f) < γ th (γ th =3dB (the lowest signal-to-noise ratio at which weak features can be identified), then the frequency band is clipped, and the clipped frequency domain signal is X. 剪 If (f) = 0, otherwise retain the original value X(f), and finally reconstruct the time domain signal through inverse STFT.

[0067] In the device coupling interference cancellation process, dynamic coupling modeling employs a graph neural network (GNN), with nodes representing devices and edge weights w. i,j Correlation between physical connections and control signals: w i,j =0.6·exp(-D i,j / D max )+0.4·|Corr(PLC i PLC j )|, where D i,j Let D be the physical distance between devices i and j (in meters). max PLC is the maximum distance between devices in the scene. i Let be the control command sequence for device i, and Corr be the Pearson correlation coefficient. Coupling strength c. i,j (t) is obtained by mapping from the Sigmoid function: Ensure the value is within the interval [0, 1]. Propagation delay τ i,j Calculated using Dynamic Time Warping (DTW): Where d DTW For the dynamic time-warped distance. The final coupling cancellation formula is: N(j) is the set of devices directly associated with device j (determined based on physical connection relationships).

[0068] S4: Design a weak fault feature enhancement mechanism, which extracts weak fault features through noise co-occurrence utilization algorithm and adaptive mode decomposition.

[0069] The specific methods in this step include: using a noise co-occurrence utilization algorithm to divide noise into equipment-related noise and irrelevant noise, and extracting weak features from equipment-related noise; and using adaptive mode decomposition to dynamically adjust the number of modes according to signal complexity to decompose the signal and extract weak fault features.

[0070] In the noise co-occurrence utilization algorithm, noise separation is first performed on the signal after interference cancellation, and wavelet packet decomposition is used to obtain n noise sub-bands n j(j = 1, 2, ..., n). Noisy subbands are classified using an improved random forest classifier. Input features include the kurtosis K of the subband. j Energy entropy H j The cosine similarity SIM with the fault feature library j ,in (F 故障 This is a historical fault feature template, sourced from an equipment fault case library (where m is the feature dimension). The classifier outputs labels l. j ∈{0,1} (0 represents irrelevant noise, 1 represents equipment-related noise), the classification loss function is (CE is the cross-entropy loss, Ent is the noise entropy, enhancing the ability to distinguish complex noise). For device-related noise, it is incorporated into feature extraction through dynamic weighting: F 噪声 =Σ j ω j ·n j where the weight ω j =exp(γ·SIM) j ) / Σ k exp(γ·SIM k (γ = 0.8 is the amplification factor, which increases the contribution of high-similarity noise).

[0071] Adaptive mode decomposition employs variational mode decomposition (VMD), and the number of modes K is dynamically determined based on the signal complexity index I: I = 0.6·K u +0.4·(1-S / S max ), where K u S is the signal kurtosis (measuring impulse characteristics), and S is the sample entropy (measuring signal regularity). max The maximum sample entropy value is derived from normal signal statistics. When I < 1.2 (stable signal), K = 3; when 1.2 ≤ I ≤ 2.0, K = 5; when I > 2.0 (complex signal), K = 7. The objective function of VMD is... Where u k (t) represents the k-th modal component, ω k Let δ(t) be the center frequency, and δ(t) be the Dirac function, solved using the Alternating Direction Multiplier (ADMM) algorithm. After decomposition, the energy characteristics E of each mode are calculated. k =∫|u k (t)| 2 dt, and filter the signal-to-noise ratio gain. modality and The energy values ​​under fault and normal conditions are respectively taken as the energy threshold for weak feature identification (6dB is the threshold for weak feature identification), and are finally fused into a weak fault feature vector F. 弱 =Ek1 E k2 ,…,E km ].

[0072] S5: Construct a federated learning cross-device transfer module to perform knowledge transfer between multiple devices and improve feature generalization ability in small sample scenarios.

[0073] The specific methods in this step include: calculating the similarity and feature differences between devices to generate transfer weights; and performing knowledge transfer between multiple devices based on the transfer weights and feature distribution alignment algorithm, while preserving individual device differences.

[0074] Device similarity calculation is based on the inherent parameters and operating characteristics of the devices, constructing a device parameter vector θ. i = [Model, Rated Power, Service Life, Historical Failure Rate] (numerical processing, such as model coded by category, power normalized to [0,1]), using cosine similarity to measure the similarity between device i and device j: The numerator is the vector dot product, and the denominator is the product of moduli, with the result ranging from [0,1] (1 indicates complete similarity). Feature differences are measured using the maximum mean difference (MMD). Let F be the feature set of device i. i ={f i1 ,f i2 ,…,f ini}(n i (where F is the sample size), the feature set of device j is F. j ,but Where φ(·) is the feature mapping function (using the RBF kernel K(x,y)=exp(-γ‖xy‖) 2 ), where γ = 0.1 is the kernel parameter, set manually, and the stability of the feature mapping is ensured through experimental verification. For the regenerated kernel Hilbert space, the smaller the MMD value, the closer the feature distribution is.

[0075] Transfer weights are generated by combining similarity and feature differences: ω j←i =SIM(θ) i ,θ j )·exp(-MMD(F i ,F j )), where ω j←i Let represent the transfer weight from device i to device j, with a value range of [0,1] (the larger the weight, the stronger the transfer contribution). Feature distribution alignment is achieved through adversarial learning. The generator G is trained to map the features of device i to the feature space of device j, and the discriminator D distinguishes between the mapped features and the original features of device j. The loss function is... Generate aligned transfer features

[0076] The knowledge transfer process preserves individual device differences, and the characteristics of device j after transfer are a weighted fusion of the original features and the transferred features: Where (1-ω j←i Weights are preserved for the original features to ensure that the unique operational characteristics of device j are not masked. For scenarios with multiple reference devices, the transfer weights are normalized: ω′ j←i =ω j←i / Σ k ω j←k The final fusion feature is This mechanism can maintain feature generalization ability even in small sample scenarios (such as device j sample size < 50).

[0077] S6: Establish a health index-driven diagnostic optimization mechanism to dynamically adjust model parameters based on the equipment's health status and generate diagnostic results.

[0078] The specific methods in this step include: calculating the equipment's health index by integrating equipment status data and historical diagnostic results; classifying the equipment's health stages based on the health index; and adjusting the mode decomposition parameters and classifier weights accordingly. The health stages include a healthy stage, a degradation stage, and a near-failure stage. Different stages correspond to different mode decomposition parameter and classifier weight adjustment strategies, and these adjustments generate the final diagnostic result.

[0079] The Health Index (HI) is calculated based on multi-feature fusion, integrating key features from equipment condition data (such as RMS vibration, temperature deviation, and current harmonic distortion rate) and the confidence level of historical diagnostic results. The calculation formula is: Where F represents the number of features (5-8 depending on the equipment type, such as vibration and temperature for rotating equipment), w f The feature weights (determined through random forest feature importance evaluation, satisfying ∑w) f =1), f(t) is the characteristic value at the current time, f 新 f represents the characteristic mean of the new equipment (derived from the equipment's factory calibration data). 旧 This is the fault threshold (a characteristic critical value derived from historical fault cases). The value of HI ranges from [0,1], and the closer the value is to 1, the healthier the equipment is.

[0080] Health stages are defined based on the HI value: healthy stage (HI ≥ 0.8), degradation stage (0.3 ≤ HI < 0.8), and near-failure stage (HI < 0.3) (thresholds are determined through statistical analysis of equipment lifecycle data to ensure each stage covers typical failure evolution cycles). Modal decomposition parameter adjustments are made to the penalty factor α for VMD. VMD Its dynamic adjustment formula is:

[0081]

[0082] Where α0 = 2000 is the initial penalty factor (optimized based on the decomposition effect of normal equipment signals), which enhances the decomposition intensity in the degradation and near-fault stages and improves the weak feature separation capability.

[0083] Classifier weight adjustment for the fusion weights α of SVM and Random Forest k (k represents the fault type), the healthy phase uses α. k =0.4 (emphasizing the noise resistance of random forests), degradation stage α k =0.6 (balancing the advantages of both), α during the near-failure stage k =0.8 (emphasizing the sensitivity of SVM to weak features). The final diagnostic result is output through the fusion classifier: P 融合 =α k ·P SVM +(1-α k )·P RF , where P SVM and P RF The failure confidence scores for the two classifiers are given, and the class with the highest confidence score is taken as the diagnostic result, along with a health index and remaining lifespan prediction (based on the HI decay rate). (Calculation). By dynamically linking health status with model parameters, the stability of diagnostic accuracy is ensured throughout the entire lifecycle of the equipment.

[0084] By acquiring multi-dimensional data (covering equipment status, interference sources, and control signals) and processing it in a spatiotemporally synchronized manner, the integrity and temporal consistency of multi-source data can be ensured, laying a high-quality data foundation for subsequent analysis and avoiding feature misjudgment due to missing or misaligned data. On this basis, the multi-source time-varying interference dynamic cancellation system can effectively weaken the masking of fault features by various interferences by predicting and compensating for interference in advance, dynamically clipping the dominant interference frequency band, and canceling cross-device coupling interference, making weak fault features stand out from complex noise. With the help of the weak feature enhancement mechanism of noise co-occurrence utilization and adaptive mode decomposition, weak fault information can be further extracted and enhanced from equipment-related noise, improving the identifiability of early fault features. Federated learning cross-device transfer achieves knowledge sharing through equipment similarity and feature distribution alignment, which can make up for the lack of data in small sample scenarios and enhance the feature generalization ability of niche devices. The health index-driven dynamic optimization mechanism adjusts model parameters by adapting to the health status of the equipment throughout its entire life cycle, ensuring the diagnostic stability of different aging stages. In summary, the synergistic effect of end-to-end technical means from data acquisition to dynamic optimization can ultimately achieve a comprehensive improvement in the accuracy, adaptability, and early fault identification capabilities of multi-device status assessment and diagnosis.

[0085] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0086] In summary, this application has at least the following beneficial effects:

[0087] 1. It can effectively counteract multi-source time-varying interference. Through pre-prediction compensation, dynamic frequency band clipping and cross-device coupling cancellation, it significantly improves the stability of weak fault feature extraction and solves the problem of interference masking early fault signals in traditional methods.

[0088] 2. By leveraging federated learning for cross-device knowledge transfer, feature generalization ability can be improved in scenarios with small sample sizes or a limited number of devices through feature distribution alignment and dynamic weight allocation, thus compensating for insufficient diagnostic accuracy caused by data scarcity.

[0089] 3. The dynamic optimization mechanism based on health index can adapt to the entire life cycle status of equipment. By adjusting the modal decomposition parameters and classifier weights in stages, it ensures the consistency of equipment diagnosis from the healthy stage to the near-failure stage and enhances the adaptability of the method to the equipment aging process.

[0090] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A multi-device status intelligent assessment and diagnosis method based on big data, characterized in that, Includes the following steps: The system utilizes edge computing nodes to collect real-time status data, interference source data, and control signal data from multiple devices, and performs preprocessing on these data. A multi-dimensional data fusion framework is constructed to fuse preprocessed multi-source data into a unified diagnostic dataset. A multi-source time-varying interference dynamic cancellation system is adopted to perform hierarchical processing on the interference in the diagnostic data, including pre-action prediction compensation, dynamic feature trimming and equipment coupling interference cancellation; A weak fault feature enhancement mechanism is designed, which extracts weak fault features through a noise co-occurrence utilization algorithm and adaptive mode decomposition. A federated learning cross-device transfer module is built to perform knowledge transfer between multiple devices, thereby improving the feature generalization ability in scenarios with few samples. Establish a health index-driven diagnostic optimization mechanism to dynamically adjust model parameters based on equipment health status and generate diagnostic results; The processing steps of the multi-source time-varying interference dynamic cancellation system include: The pre-action prediction compensation module generates an anti-phase compensation signal based on the time series prediction model to cancel out the prediction interference. The dynamic feature trimming module is used to adaptively trim frequency domain features based on the signal-to-noise ratio, retaining the sensitive region of weak features; By using a device coupling interference cancellation module, cross-device interference can be canceled through dynamic coupling modeling and transmission delay compensation; The processing steps of the health index-driven diagnostic optimization mechanism include: Calculate the equipment's health index by combining equipment status data and historical diagnostic results; The health stages of the equipment are determined based on the health index, and the modal decomposition parameters and classifier weights are adjusted accordingly. The processing steps of the weak fault feature enhancement mechanism include: The noise co-occurrence utilization algorithm divides noise into equipment-related noise and irrelevant noise, and extracts weak features from equipment-related noise. Adaptive mode decomposition is employed, which dynamically adjusts the number of modes based on signal complexity to decompose the signal and extract weak fault features.

2. The method according to claim 1, characterized in that, The construction steps of the multi-dimensional data fusion framework include: Perform protocol conversion and data cleaning on equipment status data, interference source data, and control signal data; A time-space synchronization mechanism is adopted to keep the timestamps of multi-source data consistent.

3. The method according to claim 1, characterized in that, The processing steps of the pre-reaction prediction and compensation module include: The intensity and trend of interference are predicted based on time series prediction models; Based on the prediction results, a compensation signal with the opposite phase to the interference is generated to cancel the interference.

4. The method according to claim 1, characterized in that, The processing steps of the dynamic feature cropping module include: The diagnostic data is frequency domain transformed to obtain the signal-to-noise ratio of each frequency band; Based on the signal-to-noise ratio threshold, the frequency bands that dominate interference are cropped, while the frequency bands containing weak features are retained.

5. The method according to claim 1, characterized in that, The processing steps of the federated learning cross-device migration module include: Calculate the similarity and feature differences between devices to generate transfer weights; Based on the migration weight and feature distribution alignment algorithm, knowledge transfer is performed between multiple devices while preserving individual device differences.

6. The method according to claim 1, characterized in that, The health phase includes a healthy phase, a deterioration phase, and a near-failure phase, with different modality decomposition parameters and classifier weight adjustment strategies corresponding to different phases.

7. A multi-device status intelligent assessment and diagnosis system based on big data, characterized in that, The system includes sensors, edge computing nodes, and cloud servers for deployment in industrial scenarios where multiple devices operate collaboratively. The system is used to execute the intelligent assessment and diagnosis method for the status of multiple devices based on big data, as described in any one of claims 1 to 6. The sensors are used to collect status data, interference source data, and control signal data of multiple devices, the edge computing nodes are used to perform real-time data processing, and the cloud server is used to store and analyze big data.

Images