A data-driven based fault intelligent diagnosis system and method for a fuel cell reciprocating supercharging compressor

Abstract

This invention discloses a data-driven intelligent fault diagnosis system and method for fuel cell reciprocating booster compressors, belonging to the field of equipment fault diagnosis technology. It includes acquiring real-time operating status data of mechanical equipment in multiple states collected by sensors; extracting fault features based on multi-scale data, and performing signal denoising and sensitive component screening through optimization algorithms; finally, inputting the real-time operating status data into the model. This invention acquires operating status data of different components of the fuel cell reciprocating booster compressor under different operating conditions, and uses time-domain, frequency-domain, and time-frequency-domain feature extraction methods to mine the correlation between various fault modes, establishing a lightweight MCNN-KELM fusion diagnostic model. This invention effectively suppresses multi-source coupling interference through deep coupling of time-frequency analysis and weighted sample entropy, significantly improving the identification accuracy and robustness of weak faults such as valve plate wear and spring failure under strongly non-stationary and low signal-to-noise ratio conditions, achieving end-to-end intelligent fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of equipment fault diagnosis technology, and in particular to a data-driven intelligent fault diagnosis system and method for fuel cell reciprocating booster compressors. Background Technology

[0002] Fuel cell systems, with energy conversion efficiencies exceeding 60%, are a core component of hydrogen energy utilization. The reciprocating booster compressor, a critical gas supply device in the fuel cell system, directly impacts the system's lifespan and safety. Compressor components such as valves, pistons, and springs are subjected to alternating loads and high-temperature, high-pressure conditions over extended periods, making them prone to failures such as valve wear and spring failure. These failures initially manifest as weak impact components in the vibration signal. Furthermore, the compressor's vibration signal exhibits strong non-stationarity, high impact, and low signal-to-noise ratio characteristics. The coupling and superposition of multiple vibration sources easily masks the fault characteristics, posing significant challenges to accurate diagnosis.

[0003] Currently, fault diagnosis methods for reciprocating compressors are mainly divided into two categories. First, traditional fault diagnosis methods rely on manual experience to extract features in the time domain, frequency domain, or time-frequency domain, and then combine this with a shallow classifier for discrimination. This method suffers from strong subjectivity in feature selection, poor generalization ability, and insufficient depth information mining for strongly non-stationary signals with low signal-to-noise ratios, making it difficult to meet the needs of online monitoring. Second, although data-driven intelligent fault diagnosis methods are gradually being applied in this field, key technical bottlenecks still exist: the number of modes K and the penalty factor α in variational mode decomposition (VMD) are often determined manually through trial and error; improper parameter selection can easily lead to over-decomposition or under-decomposition, causing weak fault features to be lost during the decomposition process; existing features such as sample entropy (SE) are not sensitive to changes in signal amplitude, making it difficult to effectively characterize the severity of impact-type faults; conventional convolutional neural networks use fixed-size convolutional kernels, resulting in a single feature extraction scale, and the deep model inference speed is slow, making it difficult to balance diagnostic accuracy and real-time performance.

[0004] Furthermore, existing diagnostic systems are mostly offline analysis tools or single-function software, with the collection, processing, and identification processes being isolated from each other. Key parameters need to be set manually, lacking adaptive capabilities and a fully automated processing mechanism. They also lack a fully automated processing mechanism and remote operation and maintenance functions, resulting in a low level of intelligence and failing to meet the actual engineering needs of industrial sites for real-time, robust, and predictive maintenance. Summary of the Invention

[0005] The technical problem this invention aims to solve is as follows: Existing methods for fault diagnosis of reciprocating booster compressors in fuel cells suffer from a combination of shortcomings, including VMD parameter dependence on manual trial and error leading to the loss of weak features, insensitivity of traditional sample entropy to changes in the amplitude of impact-type faults, insufficient single-scale CNN feature extraction with poor real-time performance, and a lack of a full-process adaptive intelligent diagnostic system. This invention provides a data-driven intelligent fault diagnosis system and method for reciprocating booster compressors in fuel cells, aiming to achieve adaptive signal decomposition, enhancement of impact-sensitive features, efficient extraction and lightweight classification of multi-scale deep features, and end-to-end automated intelligent diagnosis.

[0006] On the one hand, the technical solution adopted by this invention to solve its technical problem is: a data-driven intelligent fault diagnosis method for a reciprocating booster compressor of a fuel cell, comprising the following steps: S1. Multi-channel vibration signal acquisition: The accelerometer is fixed on the test point of the reciprocating booster compressor, and the vibration signals of the X, Y and Z axes are acquired synchronously through the data acquisition card to realize the synchronous acquisition and transmission of multi-directional vibration signals.

[0007] S2. Signal preprocessing: The vibration signal acquired in real time is preprocessed.

[0008] S3. SABO-VMD Adaptive Parameter Co-optimization Based on Envelope Entropy Optimization Criterion: The Subtractive Average Optimization (SABO) algorithm is adopted, with the minimum envelope entropy as the fitness function, to collaboratively iteratively search for the optimal combination of the number of modes K and the penalty factor α in variational mode decomposition (VMD).

[0009] S4. Adaptive Decomposition of Non-Stationary Fault Signals: Using the optimal parameter combination obtained above, variational mode decomposition is performed on the non-stationary fault signal (the vibration signal in S1, also known as the original vibration signal) to obtain multiple intrinsic mode function (IMF) components. Variational mode decomposition can effectively separate vibration information in different frequency bands and highlight weak fault characteristics.

[0010] S5. Screening of Fault-Sensitive Components: The IMF components are screened using the kurtosis-correlation coefficient joint criterion: First, redundant IMF components with weak correlation to the original vibration signal are removed by using the correlation coefficient threshold, and IMF components that are highly correlated with the equipment operating status are retained; then, the IMF component with the largest kurtosis value is selected as the fault-sensitive component to complete the initial noise reduction and feature purification.

[0011] An amplitude-weighted sample entropy (WSE) mechanism is introduced to perform amplitude weighting processing on the fault-sensitive components. Feature weights are assigned according to the magnitude of the vibration impact amplitude, achieving a dynamic allocation where "the larger the fault impact amplitude, the higher the feature weight." This differs from the global or fixed weight mode of conventional attention mechanisms / frequency domain weighting, further enhancing the impact characteristics of weak faults unique to reciprocating compressors, such as valve wear and spring failure, while suppressing background noise interference and outputting a weighted feature vector. .

[0012] S6. Time-Frequency Sample Construction: Two-dimensional time-frequency image samples are generated by continuous wavelet transform of the selected fault-sensitive components. All two-dimensional time-frequency image samples are stratified, with the training and test sample sets divided in a 7:3 ratio. This step converts the one-dimensional time-domain vibration signal into a time-frequency domain feature image adapted for multi-scale convolutional learning, highlighting the instantaneous impact characteristics of reciprocating compressor valve failures and providing targeted input for subsequent deep feature learning.

[0013] S7. Deep Feature Adaptive Learning Based on Multi-Scale Convolutional Neural Network (MCNN): The two-dimensional time-frequency image samples are input into the MCNN, and multi-frequency and multi-level fault impact features are extracted in parallel through multi-scale convolutional kernels. This accurately captures the instantaneous impact and time-frequency distribution pattern of valve faults in reciprocating compressors, achieves multi-scale feature fusion, and outputs a deep feature vector. .

[0014] S8. Coupling and fusion of WSE weighted features and MCNN deep features: The WSE weighted feature vector obtained in S6 is combined with the deep features of MCNN. The deep feature vectors of MCNN obtained by S8 Dynamic coupling of feature layers is performed to construct a globally fused feature vector. .

[0015] S9. Fault Intelligent Classification Based on Kernel Extreme Learning Machine (KELM): The fused feature vector A lightweight MCNN-KELM fusion diagnostic model is constructed by inputting a Kernel Extreme Learning Machine (KELM) to minimize structural risk. The training sample set is used for offline iterative training of the MCNN-KELM lightweight fusion diagnostic model, while the test sample set is used independently for generalization performance verification, with no data cross-leakage. The MCNN-KELM lightweight fusion diagnostic model ultimately outputs three fault states: normal, valve plate wear, and spring failure. This model does not simply concatenate deep time-frequency features with weighted features; instead, it achieves dynamic fusion at the feature layer, using a KELM classifier to replace the complex backpropagation process of traditional deep neural networks. It maintains high diagnostic accuracy and fast inference speed even under small sample conditions, making it more suitable for real-time diagnostic needs in industrial settings. This solves the technical problems of insufficient accuracy and poor real-time performance in identifying subtle faults in reciprocating compressors using existing methods.

[0016] Experimental data show that, compared with conventional multi-scale CNN methods, this scheme improves the fault classification accuracy by more than 8% and reduces the false detection rate under noise interference by 12% in the identification of minor faults such as valve wear and spring failure in reciprocating compressors. It can accurately identify three types of fault states: normal, valve wear, and spring failure.

[0017] In some embodiments, in step S1, the acceleration sensor is fixed to key test points such as the reciprocating booster cylinder and valve, and is used to collect vibration signals (stable signals and non-stable fault signals) of the reciprocating booster compressor in real time under normal conditions and various fault conditions, as input data for subsequent fault diagnosis.

[0018] In some embodiments, in step S2, the preprocessing includes denoising and time-frequency analysis of the vibration signal through wavelet transform and FFT transform. The above preprocessing achieves dual purification in the time domain and frequency domain, laying a reliable data quality foundation for high-precision fault diagnosis across the entire link.

[0019] In some embodiments, step S3 specifically includes the following steps: S31.SABO Initialize Population: Within the preset range of modality number K and penalty factor α, randomly generate N sets of parameter combinations [K, α] to be optimized. Each parameter combination [K, α] corresponds to an individual in the population.

[0020] S32. Calculate the envelope entropy of each agent: For each individual in the population, perform VMD decomposition to obtain multiple IMF components for the parameter combination. Select the first few IMF components to calculate the envelope entropy, and use the minimum envelope entropy value as the fitness value of that individual. The smaller the envelope entropy, the more significant the impact characteristics and the stronger the periodicity in the vibration signal, and the better the corresponding parameter combination.

[0021] S33. Iterative position update: Based on the subtraction average update mechanism of the SABO algorithm, the difference vector between individuals in the population is calculated and averaged to guide individuals to move closer to the current optimal region. At the same time, the search space is extensively explored to update the position of each parameter combination.

[0022] S34. Output the optimal parameter combination after convergence: Determine whether the current iteration count has reached the preset maximum iteration count or whether the fitness value change is less than the set threshold; if the convergence condition is met, output the global optimal parameter combination; otherwise, return to step S32 to continue iterating.

[0023] This solution completely eliminates the reliance on manual experience in parameter determination through SABO-VMD adaptive parameter collaborative optimization, avoiding over- or under-decomposition problems caused by improper parameter settings, and ensuring stable decomposition performance of VMD under different operating conditions. The minimum envelope entropy criterion is highly consistent with the physical characteristics of compressor impact failures, making the decomposition results highlight the impact characteristics of the failure to the greatest extent. It ensures that the impact components generated by early weak faults such as valve plate wear and spring failure are not lost in the decomposition process, guaranteeing the accuracy of the entire chain diagnosis from the source. SABO's global optimization capability avoids getting trapped in local optima, and the decomposition effect is stable and reliable.

[0024] On the other hand, this solution also discloses a data-driven intelligent fault diagnosis system for fuel cell reciprocating booster compressors, including: The data acquisition module includes the accelerometer and the data acquisition card.

[0025] A signal denoising module is used to preprocess the raw vibration signal acquired in real time.

[0026] The fault identification module embeds a lightweight MCNN-KELM fusion diagnostic model to output the fault status.

[0027] The system also includes a remote control module for wireless transmission of vibration signals and diagnostic results from the MCNN-KELM lightweight fusion diagnostic model.

[0028] The beneficial effects of this invention are: the fusion of end-to-end data-driven and deep intelligent algorithms, from signal acquisition, adaptive decomposition, sensitive feature screening, multi-scale deep learning to intelligent classification and recognition, completely eliminates reliance on human experience, significantly improving the diagnostic accuracy, robustness, and real-time performance of fuel cell reciprocating booster compressors under complex operating conditions, low signal-to-noise ratios, and weak fault scenarios, specifically in the following aspects: 1) Data-driven adaptive parameter optimization completely solves the pain points of traditional methods, such as reliance on manual experience for parameters and unstable decomposition results. Existing VMD requires repeated manual trial and error to determine the number of modes K and the penalty factor α, which is prone to over-decomposition / under-decomposition and easily obscures weak fault features. This invention adopts the SABO optimization algorithm + envelope entropy minimization criterion to automatically search for the optimal [K,α] in a data-driven manner without manual intervention. It has stronger adaptive decomposition capabilities for non-stationary, high-impact, and low signal-to-noise ratio vibration signals, and more accurate and repeatable separation of weak fault IMF components, providing a high-quality signal foundation for subsequent feature extraction.

[0029] 2) Data-driven feature enhancement based on weighted sample entropy (WSE) significantly improves the identification and noise resistance of weak faults. Traditional sample entropy is insensitive to the magnitude of fault impact and is prone to failure under low signal-to-noise ratio. This invention introduces an amplitude-weighted distance mechanism, which quantifies the fault contribution of each component through weighted sample entropy, strengthens the feature weight of impact-type faults, suppresses the contribution of noise components, and uses data-driven screening of highly sensitive features to effectively distinguish similar fault modes such as normal, valve plate wear, and spring failure. The recognition rate is significantly improved under low signal-to-noise ratio, solving the problems of high false negative and false positive rates of weak faults in traditional methods.

[0030] 3) The MCNN-KELM deep fusion intelligent algorithm enables multi-scale feature adaptive learning and high-precision intelligent recognition. Traditional methods rely on manually designed time-domain / frequency-domain features, resulting in limited feature representation, poor generalization, and difficulty in adapting to the complex vibration coupling characteristics of compressors. This invention is based on data-driven time-frequency sample construction and employs a multi-scale convolutional neural network (MCNN). Through 3×3 / 5×5 / 7×7 parallel convolutional kernels, it automatically extracts multi-band, multi-level, and multi-scale deep fault features without the need for manual feature design. Furthermore, it integrates a kernel extreme learning machine (KELM) to construct a strong generalized classifier by minimizing structural risk. This classifier exhibits strong nonlinear fitting ability, fast training speed, and resistance to overfitting, achieving data-driven end-to-end intelligent diagnosis. Compared to traditional machine learning, it improves diagnostic accuracy by 10%–15%, has stronger generalization, and offers higher real-time performance.

[0031] 4) A data-driven intelligent diagnostic system for the entire process, enabling real-time monitoring and intelligent early warning. Existing systems are mostly offline analysis or single-function monitoring, with data processing and algorithms separated, low intelligence, and poor operation and maintenance efficiency. This invention develops a data-driven intelligent diagnostic host computer based on LabVIEW, integrating real-time data acquisition, SABO-VMD adaptive decomposition, WSE feature weighting, MCNN-KELM intelligent recognition, health indicator assessment, and remote transmission modules. It is data-driven throughout the process, with embedded algorithms and automatic operation, outputting fault status and confidence level in real time without manual intervention. It supports remote operation and maintenance, significantly reducing the operation and maintenance threshold and improving the system's intelligence level and engineering practicality.

[0032] In summary, this invention, through a data-driven intelligent algorithm that deeply integrates SABO-VMD adaptive optimization, WSE feature enhancement, and MCNN-KELM, fundamentally overcomes the problems of traditional methods, such as reliance on manual intervention, poor fault identification, weak generalization, and low intelligence. It is particularly well-suited to the non-stationary, high-impact, and low signal-to-noise ratio operating characteristics of reciprocating booster compressors for fuel cells, achieving significant improvements in fault identification accuracy, robustness, real-time performance, and engineering practicality, thus possessing significant engineering application value. Attached Figure Description

[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0034] Figure 1 This is an overall flowchart of the method of the present invention.

[0035] Figure 2 This is a schematic diagram of the processing flow of the MCNN-KELM lightweight fusion diagnostic model of the present invention.

[0036] Figure 3 This is a flowchart for determining the equipment fault status of the present invention.

[0037] Figure 4 This is a system framework diagram of the present invention. Detailed Implementation

[0038] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0039] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0040] The purpose of this invention is to provide a data-driven intelligent fault diagnosis system and method for reciprocating booster compressors of fuel cells, addressing the faults that occur during the operation of reciprocating booster compressors and the convenience and cost requirements of testing methods. This system and method monitor reciprocating booster compressors, using LabVIEW as the development platform and calling MATLAB Script to analyze the time-frequency domain characteristics of signals, multi-scale fault characteristics, and intelligent classification methods.

[0041] like Figures 1-3 As shown, a data-driven intelligent fault diagnosis method for reciprocating booster compressors in fuel cells includes the following steps: S1. Multi-channel vibration signal acquisition: Multiple triaxial accelerometers are fixed to the cylinders, valves, and other measuring points of the reciprocating booster compressor to collect raw vibration signals (vibration time series) in real time under normal and various fault conditions of the reciprocating booster compressor, which serve as input data for subsequent fault diagnosis algorithms.

[0042] The triaxial accelerometer has three sets, which measure the vibration acceleration of an object / device in the X, Y, and Z axes respectively, and has high sensitivity and low drift characteristics.

[0043] The acquisition card acquires and transmits vibration signals from the monitored reciprocating booster compressor measuring points. The acquisition card adopts a three-channel mode and transmits the received data to the computer through the DAQ assistant, realizing the synchronous acquisition and transmission of multi-directional vibration signals.

[0044] Specifically, the acquisition of raw vibration signals of the reciprocating booster compressor under normal and various fault conditions involves: acquiring the vibration time series of the reciprocating booster compressor under different health conditions (normal and various fault conditions) within a preset operating time; the time series is the sequence of changes over time acquired by a triaxial accelerometer (preferably a piezoelectric accelerometer) within the preset time; the preset operating time is the time interval for acquiring equipment signals within a certain time interval and the time for acquiring data in the next time interval.

[0045] S2. Signal Preprocessing: The raw vibration signal acquired in real time in S1 is preprocessed. Preprocessing includes denoising and time-frequency analysis of the vibration signal through wavelet transform and FFT transform to remove noise and distinguish high- and low-frequency components. Preferably, wavelet packet transform is used to simultaneously decompose the high-frequency detail components and low-frequency approximate components of the signal to obtain a finer time-frequency resolution.

[0046] S3. SABO-VMD Adaptive Parameter Co-optimization Based on Envelope Entropy Optimization Criterion: The Subtractive Average Optimization (SABO) algorithm is adopted, with the minimum envelope entropy of the preprocessed vibration signal as the fitness function. The optimal number of modes K and the penalty factor α in Variational Mode Decomposition (VMD) are searched iteratively in a cooperative manner. The fitness function is expressed as:

[0047] in, Envelope entropy is used to measure the complexity of the preprocessed vibration signal; the smaller the entropy value, the more prominent the fault characteristics.

[0048] Specifically: S31.SABO Initialize Population: Within the preset range of modality number K and penalty factor α, randomly generate N sets of parameter combinations [K, α] to be optimized. Each parameter combination [K, α] corresponds to an individual in the population.

[0049] S32. Calculate the envelope entropy of each agent: For each individual in the population, perform VMD decomposition on the preprocessed vibration signal to obtain multiple IMF components for the parameter combination. Select the first few IMF components to calculate the envelope entropy, and use the minimum envelope entropy value as the fitness value of that individual. The smaller the envelope entropy, the more significant the impact characteristics and the stronger the periodicity in the vibration signal, and the better the corresponding parameter combination.

[0050] S33. Iterative position update: Based on the subtraction average update mechanism of the SABO algorithm, the difference vector between individuals in the population is calculated and averaged to guide individuals to move closer to the current optimal region. At the same time, the search space is extensively explored to update the position of each parameter combination.

[0051] S34. Output the optimal parameter combination after convergence: Determine whether the current iteration count has reached the preset maximum iteration count or whether the change in fitness value is less than the set threshold; if the convergence condition is met, output the global optimal parameter combination [K, α]; otherwise, return to step S32 to continue iterating.

[0052] S4. Adaptive decomposition of non-stationary fault signals: Using the optimized parameter combination [K, α] obtained above, variational mode decomposition (VMD) is performed on the non-stationary fault signals (preprocessed vibration signals under various fault states) to obtain multiple intrinsic mode function (IMF) components:

[0053]

[0054] in: For the first kEach intrinsic mode function (IMF) component represents the decomposition result of the signal (preprocessed vibration signal) in different frequency bands; For the first k The center angular frequency corresponding to each IMF component, in units of rad / s , representing the dominant frequency of this modal component; K The preset total number of modal decompositions (optimal parameters obtained by SABO optimization) controls the number of IMF components after decomposition. This is the Dirac delta function, used to characterize the instantaneous properties of the impact signal; The imaginary unit is used to construct complex frequency domain analytic signals; For time t Find the partial derivative to characterize the rate of change of the signal in the time domain; The L2 norm (Euclidean norm) is used to measure the sum of squares of signal energy. As a penalty factor The equivalent expression (in the original VMD formula) correspond ), controlling the bandwidth constraint strength of each IMF component; The vibration signal is the preprocessed signal, i.e., the non-stationary fault signal to be decomposed.

[0055] S5. Screening of Fault-Sensitive Components: IMF components are screened using a joint kurtosis-correlation coefficient criterion.

[0056] in, R is the kurtosis, and R is the correlation coefficient. These are weighting coefficients used to balance the contributions of kurtosis and correlation coefficients, filtering out high-impact, high-correlation components, i.e., fault-sensitive components. Weighting coefficients The range of values ​​is This invention, combining the fault characteristics of reciprocating compressors, determines the optimal value through grid search optimization. It can effectively enhance the characteristics of fault impact and suppress noise interference.

[0057] Constructing a WSE (Warp-Survey) model to enhance the contribution of fault impact characteristics and suppress noise; assuming the vibration time series (preprocessed vibration signal) is... With embedding dimension m, construct the vector:

[0058] in, Let be the i-th m-dimensional reconstruction vector.

[0059] Define the tolerance threshold r as:

[0060] Where r is the distance threshold; This is a proportionality coefficient, commonly ranging from 0.1 to 0.25; The standard deviation of the original vibration time series. The one-dimensional vibration time series is reconstructed as follows: m The phase space vector is used to prepare for subsequent distance calculations.

[0061] Define two reconstructed vectors and Weighted distance between for:

[0062] in, , The first i , No. j indivual m Reconstructed vectors of dimension i,j=1,2,…,N m+1, This represents the maximum absolute distance between vectors. The average absolute magnitude of the i-th vector is calculated by multiplying the maximum absolute distance of the traditional sample entropy by an magnitude weighting factor. This makes the distance metric sensitive to changes in the vibration signal amplitude. The amplitude weighting factor is essentially the average absolute amplitude of the i-th reconstructed vector—the larger the amplitude, the larger the weighted distance, and the more difficult it is for the two vectors to be judged as a "match," thus increasing the WSE value and quantifying the degree of fault impact contribution of the vibration signal segment.

[0063] For all ,like If the two vectors are matched, then they are considered a match. Let B and A be the total number of matching vector pairs in the m-dimensional and m+1-dimensional reconstruction spaces, respectively. Then the weighted sample entropy (WSE) is defined as:

[0064] In actual calculations, to avoid the formula becoming invalid due to B=0, the following modified formula is used:

[0065] in, , The number of matches; This is a zero-prevention correction term for extremely small positive numbers. The smaller the WSE value, the more regular and periodic the vibration signal; the larger the WSE value, the more complex and random the vibration signal.

[0066] This step involves introducing WSE (Weighted Feature Extraction) to perform amplitude-weighted feature enhancement processing on the screened fault-sensitive components, and then calculating the WSE-weighted feature vector. The specific calculation formula is as follows:

[0067] in: For the first The amplitude-weighted sample entropy (WSE value) of each fault-sensitive component. For the first The magnitude weights of each IMF component satisfy the following conditions: The result is obtained by normalizing the correlation coefficient between the IMF component and the original vibration signal:

[0068] in, For the first The Pearson correlation coefficient between each IMF component and the original vibration signal; The final output WSE weighted feature vector is obtained by quantifying the contribution of each vibration signal to the fault impact through weighted distance calculation, which accurately characterizes the amplitude difference and feature weight distribution of the fault impact of the reciprocating compressor.

[0069] This step involves introducing WSE (Weighted Evaluator) to enhance the amplitude-weighted features of the screened fault-sensitive components. The weighted distance calculation quantifies the fault impact contribution of each vibration signal, resulting in the calculated WSE-weighted feature vector. ,Should The output calculated by the weighted sample entropy algorithm mentioned above can accurately characterize the amplitude difference and feature weight distribution of the failure impact of the reciprocating compressor.

[0070] S6. Time-frequency sample construction: Continuous wavelet transform is applied to the selected fault-sensitive components to generate two-dimensional time-frequency image samples, which are then used as input to a multi-scale convolutional neural network (MCNN). All two-dimensional time-frequency image samples are divided into training and testing sets.

[0071] S7. Deep Feature Adaptive Learning Based on Multi-Scale Convolutional Neural Networks (MCNN): Two-dimensional time-frequency image samples are input into the MCNN, and multi-frequency, multi-level fault impact features are extracted in parallel using 3×3, 5×5, and 7×7 multi-scale convolutional kernels. Let the input two-dimensional time-frequency image sample be... The convolution operation of MCNN can be represented as:

[0072] in, For convolution kernel weights, For bias terms, This is an intermediate convolutional feature. After the fault impact feature is mapped through the fully connected layers of the MCNN, a deep feature vector is obtained. This refers to the deep features of MCNN that participate in feature fusion as described below.

[0073] S8. Coupling and fusion of WSE weighted features and MCNN deep features: WSE weighted eigenvectors Deep feature vectors extracted by MCNN Dynamic coupling and fusion of feature layers are performed to construct a global fused feature vector. This allows the contribution of fault impact amplitude to be embedded into the fusion feature in the form of weights.

[0074] S9. Intelligent Fault Classification and Health Assessment Based on Kernel Extreme Learning Machine (KELM): fuse feature vectors Input the Kernel Extreme Learning Machine (KELM) and construct a lightweight MCNN-KELM fusion diagnostic model, outputting three types of fault states: normal, valve plate wear, and spring failure.

[0075] The KELM decision model can be represented as:

[0076] in, The predicted fault state (i.e., fault type); For kernel functions; It is the identity matrix; The regularization coefficient is used. For the kernel matrix; The true labels for the training sample set; The fused feature vector is obtained by coupling the WSE weighted feature vector with the MCNN deep feature vector. This is the transpose matrix corresponding to the fused feature vector.

[0077] By inputting the deep features of MCNN into a kernel extreme learning machine (KELM), a lightweight MCNN-KELM fusion diagnostic model is constructed.

[0078] in, The kernel matrix is ​​C, and the regularization coefficient is C, representing the normal, valve plate wear, and spring failure fault states.

[0079] The MCNN-KELM lightweight fusion diagnostic model ultimately outputs three types of fault states: compressor normal, valve plate wear, and spring failure.

[0080] Finally, the fault status is determined by WSE (Weighted Sample Entropy) threshold classification, and the threshold is set in conjunction with the LabVIEW online monitoring program logic: When WSE < 0.6761, it is judged to be in a healthy state, corresponding to the normal operating condition of the compressor; When 0.6761<WSE<0.8231, the corresponding compressor valve plate wear fault is determined; When WSE>0.8231, the corresponding compressor spring failure fault is determined; Based on the above threshold rules, the fault state of the compressor is output.

[0081] The following is a calculation example: Embodiment 1: Working condition of valve plate wear fault The vibration data of compressor valve plate wear is collected, and the SABO algorithm is used for optimization with the minimum envelope entropy to obtain the optimal VMD parameters: the number of modes K=6, and the penalty factor α=2400. After the vibration signal is decomposed by VMD, the kurtosis-correlation coefficient criterion is adopted to screen out the fault-sensitive IMF component, which is substituted into the weighted sample entropy formula for calculation, and WSE=0.7021 is obtained. This value falls within the wear fault threshold interval, which can effectively identify the slight wear fault of the compressor valve plate.

[0082] Embodiment 2: Working condition of spring failure fault The vibration data of compressor spring failure is collected, and the optimal VMD parameters are obtained through SABO optimization: the number of modes K=4, and the penalty factor α=1200. After completing the vibration signal decomposition and the screening of fault-sensitive IMF components, WSE=1.2654 is obtained by calculation. This value is significantly higher than that under the normal working condition and the valve plate wear working condition, with high fault feature discrimination, and the spring failure fault can be accurately determined.

[0083] As Figure 4 shown, a data-driven intelligent fault diagnosis system for fuel cell reciprocating booster compressor takes LabVIEW as the upper computer (computer) development platform, integrates the functions of data acquisition, signal processing, fault identification and remote transmission, and forms a complete set of online monitoring and intelligent diagnosis platform.

[0084] Specifically, the system includes a main program login module, a data acquisition module, a signal analysis module, a fault identification module and a remote control module which are connected to each other. The main program login module is used for user login of the system, including a password comparison program, a password modification program and an interface exit program; the password comparison program compares the text input by the user with the password book to judge whether the account and password are correct; the password modification program requires the original password to be input first, and a new password can be set only after the original password is verified to be correct; the interface exit program supports one-click shutdown of the entire system.

[0085] The data acquisition module includes the piezoelectric accelerometer (1A314E type accelerometer) and the data acquisition card (IOtech640U data acquisition card) mentioned in the above method. The data acquisition card adopts a three-channel mode to synchronously acquire vibration signals in the X-axis, Y-axis and Z-axis directions, transmits them to the computer through DAQ Assistant, and stores them in TDMS format, named according to the sampling time.

[0086] The signal analysis module calls MATLAB programs through the MATLAB Script node embedded in LabVIEW to perform the following operations on the raw vibration signals acquired in real time by the data acquisition module: it performs time-frequency analysis and denoising of the signal through wavelet packet transform, removing noise and preserving the details of sudden impacts; and it obtains the global frequency domain distribution characteristics of the signal through Fourier transform (FFT).

[0087] The fault identification module incorporates a lightweight fusion diagnostic model of MCNN-KELM. It sequentially performs SABO-VMD adaptive parameter co-optimization based on the envelope entropy optimality criterion, adaptive decomposition of non-stationary fault signals, screening of fault-sensitive components, weighted sample entropy (WSE) feature enhancement, time-frequency sample construction, deep feature adaptive learning of multi-scale convolutional neural network (MCNN), dynamic coupling of feature layers, and intelligent classification of kernel extreme learning machine (KELM). It outputs three types of fault states: normal, valve plate wear, and spring failure.

[0088] The remote control module connects the data acquisition module, computer, and terminal wirelessly (e.g., using a mature Bluetooth communication module to build a wireless data transmission link, with the HC-06 model being the preferred choice) to realize the transmission of vibration signals and diagnostic results, facilitating flexible deployment of field equipment and offline testing.

[0089] This invention presents a complete data-driven intelligent fault diagnosis system for reciprocating booster compressors in fuel cells, built upon the LabVIEW development platform. The system can be quickly ported to a host computer via software packaging and is ready to use after installation. The system simplifies operation and reduces complexity by establishing a graphical user interface on a display screen. It also integrates a Bluetooth communication module for remote data transmission and wireless monitoring, supports multi-device interconnection, and enhances system deployment flexibility and field applicability.

[0090] The specific requirements for construction shall be determined according to actual construction needs, and no specific restrictions shall be imposed in this application.

[0091] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A data-driven intelligent fault diagnosis method for reciprocating booster compressors in fuel cells, characterized in that, Includes the following steps: S1. Multi-channel vibration signal acquisition: Fix the accelerometer to the test point on the reciprocating booster compressor to acquire vibration signals; S2. Signal preprocessing: The vibration signal is preprocessed; S3. SABO-VMD Adaptive Parameter Co-optimization Based on Envelope Entropy Optimization Criterion: The subtractive average optimization algorithm is adopted, with the minimum value of the preprocessed vibration signal envelope entropy as the fitness function, and the optimal number of modes K and penalty factor α of variational mode decomposition are searched in a co-iterative manner to form the optimal parameter combination. S4. Adaptive decomposition of non-stationary fault signals: The optimal parameter combination is used to perform variational mode decomposition on the non-stationary fault signals in the preprocessed vibration signals to obtain multiple intrinsic mode function (IMF) components. S5. Screening of Fault-Sensitive Components: The IMF components are screened using the kurtosis-correlation coefficient joint criterion to obtain fault-sensitive components; An amplitude-weighted sample entropy (WSE) mechanism is introduced to calculate the WSE of the fault-sensitive component and output a weighted feature vector. ; S6. Time-frequency sample construction: Two-dimensional time-frequency image samples obtained based on the fault-sensitive components; S7. Deep Feature Adaptive Learning Based on Multi-Scale Convolutional Neural Network (MCNN): Input the two-dimensional time-frequency image samples into the MCNN, and output deep feature vectors. ; S8. Coupling and fusion of WSE weighted features and MCNN deep features: combining weighted feature vectors With deep feature vectors Dynamic coupling of feature layers is performed to construct a globally fused feature vector. ; S9. Fault Intelligent Classification Based on Kernel Extreme Learning Machine (KELM): The fused feature vector Input the Kernel Extreme Learning Machine (KELM) to construct a lightweight fusion diagnostic model of MCNN-KELM, and output the current fault status of the device; the fault status includes three categories: normal, valve plate wear, and spring failure.

2. The data-driven intelligent fault diagnosis method for reciprocating booster compressors of fuel cells according to claim 1, characterized in that, In step S1, the vibration signal is a vibration signal along the X, Y, and Z axes.

3. The data-driven intelligent fault diagnosis method for reciprocating booster compressors of fuel cells according to claim 1, characterized in that, In step S2, the preprocessing includes wavelet transform and FFT transform, used to denoise and perform time-frequency analysis on the vibration signal.

4. The data-driven intelligent fault diagnosis method for reciprocating booster compressors of fuel cells according to claim 1, characterized in that, In step S3, the SABO-VMD adaptive parameter co-optimization specifically includes the following steps: S31.SABO Initialize Population: Within the preset range of modality number K and penalty factor α, randomly generate N sets of parameter combinations [K, α] to be optimized, each set of parameter combinations [K, α] corresponding to an individual in the population; S32. Calculate the envelope entropy of each agent: For the parameter combination corresponding to each individual in the population, perform VMD decomposition to obtain multiple IMF components, select the first few IMF components to calculate the envelope entropy, and use the minimum envelope entropy value as the fitness value of the individual. S33. Iterative position update: Based on the subtraction average update mechanism of the SABO algorithm, the position of each parameter combination is updated; S34. Output the optimal parameter combination after convergence.

5. The data-driven intelligent fault diagnosis method for reciprocating booster compressors of fuel cells according to claim 1, characterized in that, In step S5, the fault-sensitive component screening specifically involves: first, eliminating redundant IMF components that are weakly correlated with the vibration signal obtained in S1 using a correlation coefficient threshold; then, selecting the IMF component with the largest kurtosis value as the fault-sensitive component.

6. The data-driven intelligent fault diagnosis method for reciprocating booster compressors of fuel cells according to claim 1, characterized in that, In step S6, the generated two-dimensional time-frequency image samples are divided into training sample set and test sample set according to a fixed ratio, which are used for offline training and online diagnostic effect verification of the MCNN-KELM lightweight fusion diagnostic model, respectively.

7. The data-driven intelligent fault diagnosis method for reciprocating booster compressors of fuel cells according to claim 1, characterized in that, In step S7, the deep feature adaptive learning based on multi-scale convolutional neural network (MCNN) specifically involves: inputting the two-dimensional time-frequency image samples into the MCNN, extracting multi-frequency and multi-level fault impact features in parallel through multi-scale convolutional kernels, achieving multi-scale feature fusion, and outputting the deep feature vector. .

8. The data-driven intelligent fault diagnosis method for reciprocating booster compressors of fuel cells according to claim 1, characterized in that, In step S8, the coupling and fusion of the WSE weighted features and the MCNN deep features specifically involves: combining the WSE weighted feature vector obtained in step S6... The MCNN deep feature vector obtained in step S8 Dynamic coupling of feature layers is performed to construct a globally fused feature vector. This allows the contribution of fault impact amplitude to be embedded into the fusion feature in the form of weights.

9. A data-driven intelligent fault diagnosis system for a fuel cell reciprocating booster compressor according to any one of claims 1-8, characterized in that, include: The data acquisition module includes the accelerometer and the data acquisition card; A signal denoising module is used to preprocess the vibration signal acquired by the data acquisition module; The fault identification module embeds a lightweight MCNN-KELM fusion diagnostic model to output the fault status. A remote control module is used for transmitting vibration signals and the output results of the MCNN-KELM lightweight fusion diagnostic model.

Images