Diesel engine cylinder fault diagnosis method and system based on MMFE-PSO-KELM

Abstract

The invention provides a diesel engine cylinder fault diagnosis method and system based on MMFE-PSO-KELM, and the method comprises the steps: S1, collecting a vibration signal of the surface of a cylinder cover of a diesel engine cylinder and a rotating speed signal of a crankshaft, and carrying out the preprocessing of the vibration signal; s2, an MMFE algorithm is adopted to analyze and process the preprocessed vibration signals, and MMFE feature vectors rich in high-frequency and low-frequency fault mode information are constructed; s3, optimizing a regularization parameter C and a kernel function parameter sigma of the KELM model by using a PSO (Particle Swarm Optimization) algorithm, and searching an optimal parameter combination which maximizes the classification accuracy by taking the classification accuracy as a fitness function; and S4, constructing a KELM diagnosis model based on the optimal parameter combination, and performing fault diagnosis on the diesel engine cylinder by taking the high-frequency and low-frequency fault mode information extracted by the MMFE algorithm as input. According to the method, strong noise interference can be effectively inhibited, fault feature information can be comprehensively reserved, KELM key parameters can be optimized, and the accuracy and reliability of fault diagnosis of the diesel engine cylinder are improved.

Description

Technical Field

[0001] This invention relates to the field of diesel engine fault diagnosis technology, and in particular to a diesel engine cylinder fault diagnosis method and system based on MMFE-PSO-KELM. Background Technology

[0002] Currently, mainstream diesel engine fault diagnosis methods include oil analysis, parameter detection, instantaneous speed analysis, and vibration analysis. Among these, vibration signal-based fault diagnosis methods have become the most widely used approach in engineering practice due to their advantages such as no need to disassemble equipment, convenient testing, low cost, and high diagnostic accuracy.

[0003] Within the framework of vibration analysis, fault feature extraction is a core step in the diagnostic process. Existing feature extraction methods mainly include time-frequency domain statistical feature analysis, time-frequency analysis, and wavelet analysis. Time-frequency domain statistical feature analysis is computationally efficient, has good real-time performance, and clear physical meaning, but it is sensitive to noise and struggles to extract effective features under strong background noise. Furthermore, its ability to distinguish non-stationary signals is limited. While time-frequency analysis can better adapt to the characteristics of non-stationary signals and improve feature separability, it suffers from problems such as the mutual constraint between time and frequency resolution, parameter settings relying on experience, and susceptibility to noise and cross-term interference. Wavelet analysis has the advantage of adaptive time-frequency focusing and is highly effective in extracting features from transient impact faults. However, its performance is highly dependent on the selection of wavelet basis functions and the number of decomposition levels, resulting in insufficient universality. It also has weak ability to extract features from broadband non-transient faults and may introduce frequency aliasing errors, affecting feature accuracy.

[0004] Given the complex operating conditions and harsh environment of diesel engines, their vibration signals are often accompanied by strong random noise interference from exhaust and radiation. Traditional feature extraction methods, as described above, are prone to decreased sensitivity to fault features under these conditions. To enhance noise resistance, complexity metrics such as fuzzy entropy have been introduced into the field of rotating machinery fault diagnosis. However, fuzzy entropy and its derivative methods (such as multi-scale fuzzy entropy) are prone to losing high-frequency detail information of the original signal during coarse-grained processing, making it difficult to completely preserve the signal's pattern characteristics and limiting its diagnostic effectiveness under complex operating conditions.

[0005] In the fault mode identification stage, machine learning algorithms, represented by Kernel Extreme Learning Machine (KELM), are widely used due to their fast training speed, strong nonlinear mapping ability, and good generalization performance. The performance of KELM is highly dependent on the proper selection of kernel function parameters and regularization parameters: inappropriate kernel function parameters can easily lead to overfitting or underfitting of the model; improper selection of regularization parameters may cause the model to oversimplify or overfit to the training data.

[0006] Therefore, there is an urgent need for a diesel engine cylinder fault diagnosis method and system based on MMFE-PSO-KELM, which can effectively suppress strong noise interference, fully retain fault characteristic information, and optimize KELM key parameters, thereby improving the accuracy and reliability of diesel engine cylinder fault diagnosis. Summary of the Invention

[0007] To address the aforementioned deficiencies in the prior art, the present invention aims to provide a diesel engine cylinder fault diagnosis method and system based on MMFE-PSO-KELM, which seeks to solve the technical problems of incomplete extraction of diesel engine cylinder fault features under strong noise and low diagnostic accuracy due to the difficulty in optimizing KELM parameters.

[0008] To achieve the above objectives, in a first aspect, the present invention provides a diesel engine cylinder fault diagnosis method based on MMFE-PSO-KELM, the steps of which include:

[0009] S1. Collect vibration signals from the surface of the diesel engine cylinder head and crankshaft speed signals, and preprocess the vibration signals.

[0010] S2. The preprocessed vibration signal is analyzed and processed using the multi-level and multi-scale fuzzy entropy (MMFE) algorithm to construct MMFE feature vectors rich in high- and low-frequency fault mode information.

[0011] S3. Use the Particle Swarm Optimization (PSO) algorithm to optimize the regularization parameter C and kernel function parameter σ of the Kernel Extreme Learning Machine (KELM) model, and search for the optimal combination of parameters that maximizes the classification accuracy, with the classification accuracy as the fitness function.

[0012] S4. Construct a KELM diagnostic model based on the optimal parameter combination, and use the high and low frequency fault mode information extracted by the MMFE algorithm as input to perform fault diagnosis on the diesel engine cylinder.

[0013] As a further improvement to the above scheme, in step S1, when preprocessing the vibration signal, the pulse counting method is used to obtain the information of each working cycle of the cylinder and preprocess the vibration signal.

[0014] As a further improvement to the above scheme, the specific steps for obtaining cylinder working cycle information using the pulse counting method are as follows:

[0015] Based on the crankshaft signal disk parameters, set the crankshaft rotation angle corresponding to a single pulse;

[0016] Calculate the top dead center reference value of the cylinder based on the number of cylinders, number of strokes, and crankshaft reference position of the diesel engine;

[0017] The pulse count calculates the current crankshaft position and the total angle rotated, and combines this with the top dead center reference position to calculate the start and end times of the current working cycle of the cylinder.

[0018] As a further improvement to the above scheme, the specific steps for preprocessing the vibration signal in step S1 are as follows:

[0019] S11, Signal Denoising: This process reduces the noise of the original vibration signal. A low-pass filter is used to eliminate interference from random noise such as exhaust gas and radiation. A Butterworth low-pass filter is preferred, and its specific form is shown below:

[0020] ;

[0021] In the formula, The angular frequency of the input signal. For the transition frequency, Let be the system order. Let H(jΩ) be the passband ripple coefficient, and H(jΩ) be the system transfer function in the complex frequency domain.

[0022] S12, Signal Smoothing: This smooths the noise-reduced vibration signal. By using a smoothing window to reduce noise and unwanted high-frequency components, the mean smoothing model takes the following form:

[0023] ;

[0024] In the formula, It is the length of the mean smoothing window;

[0025] S13. Construct a single working cycle data sample and smooth the vibration signal. The data is divided into segments based on the start and end times of each working cycle of the cylinder, and each segment of data contains the vibration response of one working cycle of the cylinder.

[0026] After preprocessing, a series of data sample sets containing vibration information of a single working cycle of the cylinder were obtained. .

[0027] As a further improvement to the above scheme, the method for constructing the MMFE feature vector in step S2 is as follows:

[0028] S21. Process the preprocessed data sample set Single-cycle vibration response sequence Perform k-level (k is a positive integer) high- and low-frequency decomposition to obtain A length of High and low frequency mode signal sequences , where N is Data length, The sequence number is the index of the modal signal sequence. ;

[0029] S22, Vibration response sequence Each mode signal sequence Multi-scale analysis was performed to obtain a length of A multi-scale time series ,in This is a scale parameter, typically taken as an integer power of 2;

[0030] S23, Vibration response sequence Calculate the signal sequences of each mode separately. Multiscale time series Fuzzy entropy The construction length is The MMFE eigenvectors.

[0031] As a further improvement to the above scheme, the method for generating the high-frequency and low-frequency mode signal sequences in step S21 is as follows:

[0032] Constructing a one-dimensional time series The high- and low-frequency decomposition operators are as follows:

[0033] Low-frequency operators: ;

[0034] High-frequency operators: ;

[0035] in, and They are respectively for The high- and low-frequency mode signal sequences obtained by performing a single-level decomposition have a length of After k-level decomposition, we obtain A length of The high- and low-frequency mode signal sequences, wherein the first... A sequence of modal signals ( The form is as follows:

[0036] ;

[0037] In the formula, It is the operator used in each level of the k-level decomposition. The value is 0 or 1, corresponding to the low-frequency operator Q0 or the high-frequency operator Q1, that is, each layer selects the "low-frequency / high-frequency" operator to split the signal; ,in, Indicates that the calculation is not greater than The smallest integer.

[0038] As a further improvement to the above scheme, the specific steps of the multi-scale analysis in step S22 are as follows:

[0039] For one-dimensional time series Set scale parameters The form of the multi-scale time series is shown below:

[0040] .

[0041] As a further improvement to the above scheme, in step S23, the fuzzy entropy of the time series is calculated. The specific steps are as follows:

[0042] S231. Set the embedding dimension parameter m and the similarity tolerance parameter r;

[0043] S232, For one-dimensional time series N represents the total number of sampling points in the entire signal, which is then transformed into an m-dimensional vector. Its form is as follows:

[0044] ;

[0045] In the formula, ;

[0046] S233, Calculation Chebyshev distance between two adjacent m-dimensional vectors As shown below:

[0047] ;

[0048] S234. Calculate the fuzzy similarity between two vectors. As shown below:

[0049] ;

[0050] In the formula, For fuzzy functions, the similarity tolerance parameter is generally taken as... ,in Let X be the standard deviation. The gradient of the fuzzy function is generally taken as... M is a length correction parameter related to the embedding dimension m;

[0051] S235. Repeat the above steps for dimension m+1 to obtain... Calculate the fuzzy entropy of X as follows:

[0052] .

[0053] As a further improvement to the above scheme, the specific steps for finding the optimal parameter combination of the KELM model in step S3 are as follows:

[0054] S31. Initialize the population: Initialize parameters such as the maximum population size, maximum number of evolutions, local search capability, and global search capability of PSO-KELM, initialize the position and velocity of each particle, and calculate the initial fitness.

[0055] S32. Calculate the fitness value of a particle: In PSO-KELM, the position of each particle represents a KELM regularization parameter. and kernel function parameters The combination of parameters is used to calculate the fitness value of each particle based on the fitness function. Each particle updates its individual optimal position and global optimal position according to the result of the fitness function, and moves in the optimal direction according to the updated result to search for the minimum value of the fitness function and find the optimal combination of parameters. The form of the fitness function that maximizes classification accuracy is shown below:

[0056] ;

[0057] In the formula, It is about optimizing the combination of parameters. Indicates the length of the validation set. It is an indicator function; a correct classification is 1, and a wrong classification is 0. and These are the actual class and the predicted class of the i-th sample, respectively;

[0058] S33. Finding individual and group extreme values: Compare the fitness value of each particle with the individual extreme value. If the fitness value is smaller, update the individual extreme value with the fitness value and compare the updated individual extreme value with the global best fitness value. If the individual extreme value is smaller, use the individual extreme value as the current group extreme value.

[0059] S34. Update particle velocity and particle position, as shown in the following formula:

[0060] ;

[0061] In the formula, The inertial weight of a particle. and It is the learning factor, i.e., the speedup constant. and It is a uniformly random number between 0 and 1, used to limit the position and velocity of particles; , and These are the position, velocity, and historical best position of the i-th particle, respectively. It is the optimal position of the entire particle swarm, i.e., the global optimum.

[0062] S35: Continue iteration or output the optimal value. Determine whether the current particle is within the allowable error range of the fitness function. If it is, terminate the iteration, output the current particle position as the optimal parameter combination, and terminate the current iteration. Otherwise, return to S32 to continue the iteration calculation.

[0063] As a further improvement to the above scheme, the specific steps for constructing the KELM diagnostic model in step S4 are as follows:

[0064] S41. Output-based regularization parameters and kernel function parameters The optimal combination is used to construct the original KELM fault diagnosis model;

[0065] S42. Divide the preprocessed diesel engine cylinder head surface vibration data into training set and test set, and calculate and construct MMFE feature vector;

[0066] S43. The original KELM fault diagnosis model is trained and tested using the training set and the test set until the diagnostic performance of the original KELM fault diagnosis model meets the preset requirements, and then used as the optimized KELM diagnosis model.

[0067] In a second aspect, the present invention also provides a system for diagnosing diesel engine cylinder faults based on the MMFE-PSO-KELM method provided in the first aspect, comprising:

[0068] The data acquisition module is used to collect vibration signals from the surface of the diesel engine cylinder head and crankshaft speed signals.

[0069] The data preprocessing module is used to calculate the working cycle information of the diesel engine cylinder, perform noise reduction and smoothing on the collected vibration signals, and divide the data samples according to the working cycle information.

[0070] The feature extraction module performs multi-level, multi-scale analysis on the preprocessed vibration data samples to construct MMFE feature vectors.

[0071] The diagnostic model optimization module uses the PSO algorithm to search for the optimal combination of regularization parameters and kernel function parameters of the KELM model.

[0072] The fault diagnosis module is used to build an optimized KELM fault diagnosis model, using MMFE feature vectors as input to identify fault modes in diesel engine cylinders.

[0073] Because the present invention adopts the above technical solutions, the beneficial effects of this application are as follows:

[0074] This invention provides a diesel engine cylinder fault diagnosis method based on MMFE-PSO-KELM. First, a multi-level, multi-scale fuzzy entropy (MMFE) algorithm is used to analyze the preprocessed vibration signal. Because the MMFE algorithm combines the advantages of multi-level and multi-scale analysis and introduces fuzzy entropy to measure the complexity and regularity of the signal, it can examine the signal at different levels and scales, effectively distinguishing weak fault features submerged in strong noise, while also taking into account the high-frequency details and low-frequency trends contained in the signal. Therefore, compared with traditional methods, this invention can more comprehensively capture and retain the fault feature information unique to different fault modes of diesel engine cylinders that is easily masked by noise, providing richer and more reliable input features for subsequent accurate diagnosis.

[0075] Furthermore, a particle swarm optimization (PSO) algorithm is introduced to jointly optimize the regularization parameter C and kernel function parameter g of the KELM model. As a heuristic global optimization algorithm, PSO is characterized by its simple principle, fast convergence speed, and ability to escape local optima. By applying the global search capability of the PSO algorithm to the specific technical scenario of KELM model parameter optimization, this invention can automatically and efficiently search for the optimal combination of regularization parameter C and kernel function parameter g that achieves optimal KELM model performance (such as classification accuracy or mean squared error). This avoids the shortcomings of traditional KELM model parameters, which rely on manual experience or time-consuming grid search and are prone to getting trapped in local optima. Therefore, the constructed KELM diagnostic model can better adapt to the characteristics of different fault data, improving the model's generalization ability and the diagnostic accuracy and reliability of diesel engine cylinder faults.

[0076] Furthermore, the MMFE feature vector, rich in high- and low-frequency fault mode information obtained in step S2 (which possesses strong noise resistance and complete information due to the advantages of the MMFE algorithm), is used as input to the KELM diagnostic model constructed based on the optimal parameter combination obtained in step S4 using PSO optimization. This effective combination of the two core technical features, "MMFE feature extraction" and "PSO-KELM model optimization," ensures, on the one hand, that the features input to the diagnostic model have high quality and high discriminative power, and on the other hand, that the diagnostic model itself has good parameter configuration and learning capabilities. The synergistic effect of these two aspects enables the method of this invention to more accurately identify various fault types in diesel engine cylinders, effectively overcome strong noise interference and parameter optimization challenges, significantly improve the accuracy and reliability of diesel engine cylinder fault diagnosis, and meet the needs of practical engineering applications. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0078] Figure 1 This is a flowchart illustrating a diesel engine cylinder fault diagnosis method based on MMFE-PSO-KELM disclosed in this invention.

[0079] Figure 2 This is a schematic diagram of the process for obtaining cylinder working cycle information using the pulse counting method disclosed in this invention;

[0080] Figure 3 This is a schematic diagram of a preprocessing flow for vibration signals disclosed in this invention;

[0081] Figure 4 This is a schematic diagram of the MMFE feature vector extraction process disclosed in this invention;

[0082] Figure 5 This is a schematic diagram of the process for calculating the fuzzy entropy of a time series disclosed in this invention;

[0083] Figure 6 This is a schematic diagram of the KELM parameter optimization process disclosed in this invention;

[0084] Figure 7 This is a schematic diagram of the process for constructing a KELM diagnostic model disclosed in this invention;

[0085] Figure 8 This is a schematic diagram of the installation of the vibration acceleration sensor disclosed in this invention;

[0086] Figure 9 It is the prediction confusion matrix after inputting different feature vectors into KELM as disclosed in this invention;

[0087] Figure 10 This is a schematic diagram of the diagnostic results of KELM and PSO-KELM disclosed in this invention;

[0088] Figure 11 This is a structural diagram of a diesel engine cylinder fault diagnosis system based on MMFE-PSO-KELM disclosed in this invention.

[0089] The realization of the objective, functional characteristics and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0090] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0091] It should be noted that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0092] Example 1

[0093] See Figures 1-10 The present invention provides a diesel engine cylinder fault diagnosis method based on MMFE-PSO-KELM, the steps of which include:

[0094] S1. Signal Acquisition and Preprocessing:

[0095] A vibration acceleration sensor is installed on the cylinder head surface of the diesel engine, and a speed sensor is installed at the crankshaft end to synchronously collect vibration signals and corresponding crankshaft speed signals during cylinder operation. The collected vibration signals are processed by a detrending term to eliminate baseline drift, and then bandpass filtered to remove high-frequency environmental interference and low-frequency power frequency noise, resulting in a preprocessed vibration signal with improved signal-to-noise ratio. This preprocessing can suppress background noise interference with subsequent feature analysis and prevent noise from masking effective fault information.

[0096] S2. Fault feature extraction based on MMFE algorithm:

[0097] A multi-level, multi-scale fuzzy entropy (MMFE) algorithm is used to process the preprocessed vibration signal. First, multiple analysis levels (such as a low level reflecting impact details and a high level reflecting modulation patterns) and multiple scale factors (covering high-frequency details and low-frequency trends) are set. Fuzzy entropy values ​​are calculated for the signal sequences at each level and scale. These entropy values ​​quantify the complexity and regularity differences of the signal at different levels and scales, thereby constructing MMFE feature vectors containing high- and low-frequency fault mode information. Because it simultaneously captures multi-level detailed features and multi-scale global features of the signal, the MMFE feature vectors can more comprehensively reflect the unique regularities of vibration signals under different fault modes such as cylinder wear and ignition anomalies, solving the problem of missing key information in single-scale or single-level features under strong noise environments.

[0098] S3. KELM parameter optimization based on PSO algorithm:

[0099] Using the regularization parameter C and kernel function parameter σ of the Kernel Extreme Learning Machine (KELM) as variables to be optimized, and the classification accuracy of the KELM model on the training set as the fitness function, the Particle Swarm Optimization (PSO) algorithm is employed for optimization: the position and velocity of the particle swarm are initialized, and the optimal positions of individual particles and the swarm as a whole are iteratively updated until a preset number of iterations or fitness convergence is reached. Through the adaptive search mechanism of PSO, the optimal combination of C and σ for KELM generalization performance can be efficiently located in the parameter space, avoiding the blindness of manual parameter tuning and providing a reliable parameter foundation for subsequent model diagnosis.

[0100] S4 and KELM Diagnostic Model Construction and Fault Identification:

[0101] A KELM diagnostic model is constructed based on the optimal parameter combination found by PSO. The MMFE feature vector extracted from S2 is used as input for supervised training of the model. After training, the MMFE feature vector of the test sample is input, and the model outputs the corresponding cylinder fault type (such as normal, excessive valve clearance, piston ring wear, etc.). Since the input features fully cover high- and low-frequency fault information, and the classification boundary of the model is more closely aligned with the data distribution after optimization, the diagnostic model's recognition accuracy is significantly improved in noisy environments.

[0102] In this embodiment, the preprocessing step improves signal quality through detrending and filtering, laying the foundation for subsequent feature extraction; the MMFE algorithm effectively overcomes the deficiency of incomplete information in a single feature dimension under strong noise by analyzing the extracted feature vectors through multi-level and multi-scale analysis; the KELM parameters optimized by the PSO algorithm enhance the model's generalization ability and avoid the decrease in diagnostic accuracy caused by improper parameters; finally, comprehensive extraction and high-precision diagnosis of diesel engine cylinder fault features under strong noise environment are achieved.

[0103] In a preferred embodiment, this embodiment introduces a pulse counting method in the vibration signal preprocessing stage to obtain the working cycle information of each cylinder of the diesel engine, thereby achieving time-series alignment and segmentation of the vibration signal and improving the accuracy of subsequent feature extraction. The specific steps are as follows:

[0104] S11. Set the crankshaft angle corresponding to a single pulse:

[0105] Based on the number of teeth or the scale parameters of the crankshaft signal disk, the crankshaft angle value corresponding to each pulse is determined. For example, for a signal disk with N evenly distributed teeth, the crankshaft angle corresponding to a single pulse is 360° / N. This setting provides an angular reference for the subsequent accurate calculation of the crankshaft position, avoiding errors in the work cycle division caused by sampling timing deviations.

[0106] S12. Calculate the reference value of the cylinder top dead center:

[0107] Based on the number of cylinders, number of strokes (four-stroke or two-stroke), and crankshaft reference position (such as the top dead center signal trigger position), the crankshaft angle reference value at top dead center for each cylinder is calculated. The crankshaft angle reference value is used to map the crankshaft angle obtained from pulse counting to the working phase of each cylinder, ensuring that the vibration signals of different cylinders can be correctly segmented according to the working cycle.

[0108] S13. Count pulses and determine the start and end times of the work cycle:

[0109] During the data acquisition process, the crankshaft speed signal is counted in real time, and the current crankshaft position and the total angle rotated from the starting point are accumulated. Combined with the top dead center reference value obtained from S12, it is determined whether the cylinder corresponding to the current count has entered the starting phase of a certain working cycle, and the start and end times of the cycle are determined accordingly. Subsequently, the vibration signal is truncated and aligned according to the start and end times of each working cycle to form vibration data segments corresponding to each cycle.

[0110] By applying the pulse counting method described above, precise matching of vibration signals and operating cycles can be achieved during the preprocessing stage, eliminating timing misalignment caused by speed fluctuations or asynchronous sampling. This allows subsequent MMFE feature extraction to be completed within a unified and stable operating cycle framework. This measure helps maintain the timing consistency of fault modes under strong noise conditions, reduces feature distortion caused by phase misalignment, and thus improves the reliability and stability of the overall diagnosis.

[0111] As a preferred embodiment, to further refine the preprocessing process in step S1 and ensure accurate correspondence with the cylinder's working cycle while maintaining signal quality, this embodiment sequentially performs signal noise reduction, signal smoothing, and single working cycle data sample construction in S1. The specific steps are as follows:

[0112] Signal noise reduction:

[0113] The acquired raw vibration signal A low-pass filter is input to filter out random noise interference such as exhaust pulsation and structural radiation. A Butterworth low-pass filter is preferred because it has the flattest amplitude characteristic within the passband, which can reduce the adverse effects of amplitude-frequency response fluctuations on the signal shape. The system transfer function of this filter is expressed in the complex frequency domain as:

[0114] ;

[0115] In the formula, The angular frequency of the input signal. For the transition frequency, Let be the system order. The passband ripple coefficient is used to effectively reduce high-frequency random noise and retain low-frequency and mid-frequency components related to cylinder mechanical impact, thus providing a cleaner signal basis for subsequent feature extraction.

[0116] Signal smoothing:

[0117] The denoised vibration signal y(t) is passed through a mean-smoothing window to reduce residual noise and unnecessary high-frequency oscillations. The expression for the mean-smoothing model is:

[0118] ;

[0119] In the formula, It is the length of the mean smoothing window; choose it appropriately. This step can suppress local noise while avoiding excessive smoothing that would attenuate the impact characteristics. It further stabilizes the signal waveform, making it easier to accurately capture the morphology of the fault impact in subsequent analysis.

[0120] Construct a single work cycle data sample:

[0121] The smoothed vibration signal is obtained by using the start and end times of each cylinder's working cycle, which are obtained by the preceding pulse counting method. The data is segmented according to the cycle boundary, with each segment representing the vibration response of one complete working cycle of the cylinder, forming a single-cycle data sample set that includes time-series consistency. , where i represents the i-th working cycle.

[0122] Through the above preprocessing, Butterworth low-pass filtering is first used to remove the main random noise, then mean smoothing is used to suppress residual high-frequency disturbances, and finally the signal is segmented according to the working cycle boundary. This results in a final data sample set that not only significantly improves the signal-to-noise ratio but also maintains the temporal integrity of the cylinder operation process. This is particularly important under high-noise conditions: it can effectively avoid the loss of fault features due to noise masking or timing disorder, and provide a reliable input basis for the MMFE algorithm to extract high- and low-frequency fault mode information, thereby improving the stability and accuracy of the overall diagnosis.

[0123] As a preferred embodiment, in order to further refine the construction process of the MMFE feature vector in step S2, this embodiment is implemented in the following steps in sequence to achieve comprehensive extraction of high and low frequency fault mode information in the preprocessed vibration signal, thereby providing more discriminative input features for the subsequent KELM model and solving the problem of incomplete feature extraction under strong noise.

[0124] S21, High and Low Frequency Mode Signal Decomposition:

[0125] The preprocessed single-cycle vibration response sequence (n=1,2,...,N, where N is the sequence length) performs k-level (k is a positive integer) high- and low-frequency decomposition, employing multi-scale decomposition methods (such as wavelet packet decomposition, empirical mode decomposition variants, or other applicable high- and low-frequency separation techniques) to decompose the original signal into... A sequence of modal signals ,in Let be the sequence number of the modal signal sequence, and satisfy . The length of each modal signal sequence is .

[0126] This decomposition process adaptively divides the original vibration signal into multiple levels of high-frequency and low-frequency components according to the frequency range, so that each modal signal sequence It can reflect the local characteristics of signals in different frequency bands, especially highlighting key frequency band information related to cylinder failures such as impact and modulation. Through multi-level decomposition, it can effectively overcome the failure modes that may be missed by single-band analysis, and provide a rich frequency domain foundation for subsequent multi-scale and fuzzy entropy analysis.

[0127] S22. Multi-scale time series construction:

[0128] For each modal signal sequence Further multi-scale analysis was conducted using scale parameters. (Usually, the power of 2 is used, such as 2, 4, 8, etc.) Downsampling or moving average operations are performed on the sequence to generate time series at multiple scales. The time series length at each scale is .

[0129] Multiscale analysis can reveal the variation patterns of signals at different temporal resolutions, helping to capture the manifestations of fault characteristics at different time scales. For example, short-term impulses are more obvious at fine scales, while long-term modulations or periodic anomalies are easier to identify at coarse scales. This step expands the temporal perspective of features and enhances the adaptability of feature vectors to complex fault modes.

[0130] S23. Fuzzy Entropy Calculation and Eigenvector Construction:

[0131] For each modal signal sequence at different scales, time series The fuzzy entropy (FEn) of each mode is calculated. Fuzzy entropy is an indicator that measures the complexity and regularity of a time series, reflecting the structural characteristics of a signal under uncertain conditions, and exhibits good robustness to noisy signals. The calculated fuzzy entropy values ​​constitute the characteristic values ​​under this mode-scale combination.

[0132] All The fuzzy entropy values ​​of several modal signal sequences (covering different frequency ranges) calculated at various scales are arranged in a certain order to finally construct a fuzzy entropy array of length [length missing]. (The MMFE feature vector may include all modal-scale combinations, and the feature dimension may be adjusted accordingly, depending on the specific implementation.) The MMFE feature vector integrates the high and low frequency components of the signal as well as multi-scale information, and can comprehensively characterize the dynamic characteristics of the cylinder vibration signal under different operating conditions. In particular, it can still maintain effective information sensitive to faults in a high-noise environment.

[0133] The above implementation first uses high- and low-frequency decomposition to finely divide the original vibration signal according to frequency bands, effectively extracting key frequency band information related to the fault. Further, through multi-scale analysis, it captures the performance of fault features at different time resolutions, enhancing the robustness and discriminative power of the features. Finally, it quantifies the complexity of each modal-scale sequence using fuzzy entropy, constructing an MMFE feature vector containing rich fault mode information. This feature vector not only overcomes the shortcomings of incomplete extraction of single-scale or frequency band features under strong noise, but also provides a more discriminative input for the subsequent KELM model, helping to improve the accuracy and reliability of diesel engine cylinder fault diagnosis.

[0134] As a preferred embodiment, to further clarify the specific generation method of the high- and low-frequency modal signal sequences in step S21, this embodiment adopts a layer-by-layer decomposition method based on the low-frequency operator Q0 and the high-frequency operator Q1. By performing k-level decomposition on the original vibration signal sequence, the following is obtained: This provides a modal signal sequence covering different frequency components, thus offering a fine-grained frequency domain feature foundation for subsequent multi-scale analysis and fuzzy entropy calculation, effectively solving the problem of incomplete feature extraction of vibration signals under strong noise environments.

[0135] Definitions of high-frequency and low-frequency decomposition operators:

[0136] For a one-dimensional time series The low-frequency operator Q0 and the high-frequency operator Q1 are defined as follows:

[0137] ;

[0138] ;

[0139] in, It is a low-frequency operator that reflects the average trend between adjacent sampling points and is used to extract low-frequency components from the signal. This is a high-frequency operator that reflects the differential change between adjacent sampling points and is used to extract high-frequency components from the signal. After one layer of decomposition, low-frequency mode signal sequences and high-frequency mode signal sequences of length are obtained respectively.

[0140] k-level decomposition process and modal signal sequence construction:

[0141] In k-level decomposition, the original sequence By applying low-frequency or high-frequency operators layer by layer, and selecting Q0 or Q1 at each layer, the signal is progressively decomposed into modes with different frequency components. Specifically, the first... A sequence of modal signals (1≤ ≤ It is generated recursively as follows:

[0142] ;

[0143] In the formula, It is the operator used in each level of the k-level decomposition. The value is 0 or 1, corresponding to the low-frequency operator Q0 or the high-frequency operator Q1, that is, each layer selects the "low-frequency / high-frequency" operator to split the signal; ,in, Indicates that the calculation is not greater than The smallest integer. In this way, each level of decomposition is based on the sequence number. Adaptively select the low-frequency or high-frequency operator to be used in the current layer, so that after k-level decomposition, the final result is obtained. A sequence of modal signals Each sequence has a length of .

[0144] In this embodiment, by defining well-defined low-frequency and high-frequency operators and employing a layer-by-layer controllable decomposition strategy, efficient and fine-grained frequency partitioning of the original vibration signal is achieved. Each layer adaptively selects operators based on the sequence index, ensuring that each modal signal sequence can specifically reflect the local characteristics of different frequency bands in the signal. This decomposition method does not rely on complex transformation basis functions, the calculation process is clear and easy to implement, and it can obtain [the desired results] after k-level decomposition. A modal signal sequence containing rich frequency information provides a comprehensive and fine-grained input for subsequent multi-scale analysis and fuzzy entropy calculation, which helps to more accurately capture high and low frequency patterns related to faults in strong noise environments, thereby improving the comprehensiveness of feature extraction and the reliability of fault diagnosis.

[0145] As a preferred embodiment, the obtained modal signal sequences (i.e., the vibration signal or its decomposed modes within a single working cycle), further multi-scale analysis is performed to obtain signal characteristics at different time resolutions. Specifically, scale parameters are set. (Usually, the value is an integer power of 2, such as 2, 4, 8, etc.), the original sequence is averaged according to scale to construct a multi-scale time series. As shown below:

[0146] ;

[0147] The above operation is equivalent to "coarsening" the original signal in the time dimension, at each scale j. Represents a time window (length is) The average state of the signal within a given time frame reflects the overall trend or periodic changes of the signal over a longer period. This can be achieved by selecting different scale parameters. It can acquire the performance of signals at multiple time resolutions, which is helpful for capturing the response differences of fault characteristics at different time scales. For example, short-term shocks are more significant at smaller scales, while long-term modulation or periodic anomalies are easier to identify at larger scales.

[0148] As a preferred embodiment, to further quantify the complexity and regularity of each multi-scale time series, this embodiment uses fuzzy entropy as a feature index to analyze the time series of each modal signal sequence at different scales. (i.e., the sequence after multi-scale processing) performs fuzzy entropy calculation to construct MMFE feature vectors that can distinguish different fault modes.

[0149] S231, Parameter Settings:

[0150] Set the embedding dimension parameter m (usually 2 or 3) and the similarity tolerance parameter r (usually 0.1 to 0.25 times the signal standard deviation SD) for subsequent calculation of the similarity between vectors in the sequence.

[0151] S232, Sequence Vectorization:

[0152] One-dimensional time series Transform into a series of m-dimensional vectors N represents the total number of sampling points in the entire signal, and its form is as follows:

[0153] ;

[0154] In the formula, The mean of this m-dimensional vector is used for centering to reduce the impact of the DC component of the sequence on distance calculation.

[0155] S233, Chebyshev distance calculation:

[0156] Calculate any two m-dimensional vectors and Chebyshev distance between Its definition is as follows:

[0157] ;

[0158] Chebyshev distance reflects the maximum difference between two vectors in their respective dimensions and is used to measure the degree of difference between local patterns in a sequence.

[0159] S234. Fuzzy similarity calculation:

[0160] Based on Chebyshev distance, calculate the fuzzy similarity between two vectors. Its expression is as follows:

[0161] ;

[0162] in, Here, is the fuzzy function, n is the gradient of the fuzzy function, usually taken as 2; r is the similarity tolerance, generally taken as 0.1 to 0.25 times the signal standard deviation; this step reduces the sensitivity of noise to distance measurement through fuzzy logic processing, making the similarity measurement more robust.

[0163] S235, Embedded Dimension Expansion and Fuzzy Entropy Calculation:

[0164] Repeat the above process for embedding dimension m+1 to obtain Finally, the fuzzy entropy of the time series is calculated:

[0165] .

[0166] Fuzzy entropy reflects the complexity of a sequence under a given embedding dimension and similarity tolerance. A larger value indicates stronger uncertainty or disorder in the sequence, while a smaller value indicates a more regular sequence pattern. By calculating the fuzzy entropy of each modal signal sequence at different scales, subtle differences in vibration signals under normal and fault conditions can be effectively distinguished.

[0167] Each modal signal sequence The fuzzy entropy values ​​calculated at different scales r are processed and finally a length of r is constructed. (Or include all modal-scale combinations) MMFE feature vectors. MMFE feature vectors integrate the high and low frequency components of the signal as well as multi-scale and multi-modal fuzzy entropy information, which can comprehensively reflect the vibration characteristics of diesel engine cylinders under different operating conditions, and are particularly suitable for fault mode identification in high-noise environments.

[0168] As a preferred embodiment, to further clarify the optimization process of the optimal parameter combination of the KELM model in step S3, this embodiment elaborates on how to use the Particle Swarm Optimization (PSO) algorithm to optimize the regularization parameter C and kernel function parameter σ of the Kernel Extreme Learning Machine (KELM), so as to solve the problem of low diagnostic accuracy caused by improper parameter selection, and improve the model's generalization ability and fault identification accuracy in noisy environments. The PSO-KELM parameter optimization process is as follows:

[0169] S31. Initialize the population:

[0170] First, the operating parameters of the PSO algorithm are initialized, including the maximum population size (total number of particles), the maximum number of evolutions (iterations), and parameters related to local and global search capabilities (such as inertia weight ω, learning factors c1 and c2, etc.). During initialization, the position and velocity of each particle are randomly generated, where the position of each particle represents a set of KELM parameter combinations to be optimized, namely the regularization parameter C and the kernel function parameter σ. Subsequently, the fitness value of each particle is calculated based on the initial parameter combinations, serving as the benchmark for subsequent searches.

[0171] S32. Fitness Calculation and Update:

[0172] In PSO-KELM, the fitness function is used to evaluate the classification performance of the KELM model under the current parameter combination. In this embodiment, the classification accuracy on the validation set is used as the fitness index, and its expression is:

[0173] ;

[0174] Where θ=(C,σ) is the parameter combination of the current particle, N val To determine the number of samples in the validation set, This is an indicator function (1 for correct classification, 0 for incorrect classification). and Let be the true class and the model predicted class of the i-th validation sample, respectively.

[0175] Each particle updates its individual optimal position (i.e., the best parameter combination ever reached by the particle) and global optimal position (i.e., the best parameter combination found so far in the entire particle swarm) based on its current fitness value. Subsequently, each particle adjusts its position based on its individual optimal and global optimal positions, moving towards better parameter combinations to gradually approach the optimal solution.

[0176] S33, Individual Extreme Value and Population Extreme Value Update:

[0177] In each iteration, the fitness value of the current particle is compared with its historical best fitness value. If the current fitness is better (i.e., the classification accuracy is higher or the fitness function value is smaller), the individual best position of the particle is updated. Then, the updated individual best position is compared with the current global best position. If it is better, the global best position is updated to ensure that the particle swarm always tracks the current best solution.

[0178] S34, Particle velocity and position update:

[0179] The updates to particle velocity and position follow the standard PSO algorithm formula:

[0180]

[0181] Where ω is the inertia weight, used to balance local search and global search capabilities; c1 and c2 are learning factors (acceleration constants), which control the tendency of particles to move towards individual optimal and global optimal positions; r1 and r2 are uniform random numbers in the interval [0,1], which increase the randomness of the search. , and These represent the position, velocity, and optimal position of the i-th particle in the j-th dimension (i.e., parameter C or σ), respectively. This is the globally optimal position.

[0182] Through the aforementioned update mechanism, the particle dynamically adjusts its search direction and step size in each iteration, gradually moving closer to the parameter region with better fitness.

[0183] S35. Iteration Termination and Optimal Parameter Output:

[0184] After each iteration, determine whether the current global optimal fitness value has met the preset allowable error range or reached the maximum number of iterations. If the termination condition is met, stop the iteration and output the parameter combination C and σ corresponding to the current global optimal position as the optimal parameters of the KELM model; otherwise, return to step S32 to continue the next round of iteration optimization.

[0185] The PSO algorithm automatically searches for the optimal combination of regularization parameter C and kernel function parameter σ in the KELM model, avoiding the drawbacks of traditional manual parameter tuning methods, which rely on experience, are inefficient, and struggle to adapt to complex data distributions. Using the fitness function (classification accuracy) as an evaluation metric ensures that the selected parameter combination maximizes the model's classification performance on the validation set, thereby improving the KELM model's generalization ability and fault diagnosis accuracy in noisy environments. The particle swarm optimization process, through information sharing and collaboration between individuals and the group, balances global search and local fine-grained search, effectively balancing model complexity and fitting ability. Ultimately, it provides a parameter-optimized and stable KELM model for diesel engine cylinder fault diagnosis, significantly improving the reliability and accuracy of diagnostic results.

[0186] As a preferred embodiment, to further clarify the specific process of constructing and optimizing the KELM diagnostic model based on the optimal parameter combination in step S4, this embodiment elaborates on how to use the regularization parameter C and kernel function parameter σ obtained by the PSO algorithm to construct and train the KELM fault diagnosis model, so as to solve the problems of difficult parameter selection and insufficient diagnostic accuracy of traditional KELM, and achieve accurate identification of diesel engine cylinder faults under strong noise environment. Specifically, the KELM diagnostic model construction and training process is as follows:

[0187] S41. Construct the original KELM fault diagnosis model:

[0188] Based on the optimal parameter combination obtained by the PSO algorithm in step S3, namely the regularization parameter C and the kernel function parameter σ, a primitive Kernel Extreme Learning Machine (KELM) fault diagnosis model is constructed. KELM is a single-hidden-layer feedforward neural network based on kernel function learning. It nonlinearly maps input data to a high-dimensional feature space by introducing a kernel function, and constructs a linear model in this space, thereby simplifying the network training process and improving generalization ability. In this embodiment, the optimal parameters C and σ are directly substituted into the KELM model to determine its regularization strength and kernel function characteristics, providing a basic model framework for subsequent training.

[0189] S42. Data partitioning and MMFE feature vector construction:

[0190] The preprocessed diesel engine cylinder head surface vibration signal data (i.e., the data sample set obtained after signal preprocessing in step S1 and MMFE feature extraction in step S2) is divided into training and testing sets, usually randomly divided in a certain ratio (e.g., 7:3 or 8:2) to ensure the independence and reliability of model training and validation. On the divided training and testing sets, the MMFE feature vector constructed in step S2 (i.e., a feature representation containing high- and low-frequency fault mode information and multi-scale fuzzy entropy information) is used as the input features of the KELM model. This feature vector comprehensively reflects the fault sensitivity information of the cylinder vibration signal at different frequency bands and time scales, providing the model with highly discriminative input.

[0191] S43. Model Training and Performance Optimization:

[0192] The original KELM fault diagnosis model is trained using a pre-defined training set. Through kernel function mapping and regularization optimization, the mapping relationship between the input MMFE feature vector and cylinder fault categories is learned. Subsequently, the trained model is tested using a test set to evaluate its diagnostic performance (e.g., classification accuracy, recall). If the model performance does not meet preset requirements (e.g., accuracy below a threshold or false positive rate too high), the reasons are analyzed, and the data partitioning ratio, feature selection, or parameters are re-optimized. If necessary, step S3 can be returned for further fine-tuning of the parameter combination until the model's diagnostic performance on the test set reaches the expected standard. Finally, the model that meets the preset requirements is used as the optimized KELM fault diagnosis model for actual diesel engine cylinder fault diagnosis tasks.

[0193] This implementation pre-optimizes the regularization parameter C and kernel function parameter σ of the KELM model using the PSO algorithm, ensuring the model has a good generalization foundation from the early training stage. This avoids the shortcomings of traditional manual parameter tuning methods, such as low efficiency, strong subjectivity, and difficulty in adapting to complex data distributions. Based on MMFE feature vectors as input, it fully mines high-frequency, low-frequency, and multi-scale fault mode information in vibration signals, enhancing the model's ability to identify fault features. By strictly dividing the training and test sets and continuously optimizing the model based on preset performance indicators, the final optimized KELM diagnostic model can maintain a high fault identification accuracy even in strong noise environments. This effectively solves the problems of difficult parameter optimization and insufficient diagnostic accuracy in traditional KELM models, significantly improving the reliability and practicality of diesel engine cylinder fault diagnosis.

[0194] To verify the effectiveness of the diesel engine cylinder fault diagnosis method based on MMFE-PSO-KELM proposed in this invention, this embodiment tested and compared the method by building a typical diesel engine cylinder fault simulation experimental platform.

[0195] The experiment used an inline 8-cylinder 4-stroke diesel engine, whose one complete working cycle corresponds to a crankshaft rotation of 720°. For example... Figure 8 (As shown in the diagram, the vibration acceleration sensor (range 100g) is mounted on a bracket near the exhaust valve of each cylinder head to collect vibration signals from the cylinder head surface; the speed sensor is mounted on the side of the flywheel at the output end of the diesel engine to obtain the top dead center signal of the first cylinder for subsequent division of the working cycle.

[0196] The experimental conditions were set as follows: diesel engine speed 750 rpm, load 50%, and vibration signal sampling frequency 12.8 kHz. For cylinder 1, various typical cylinder faults were simulated (see Table 1 for specific fault types and corresponding numbers), and experimental data were continuously collected for 3 minutes under each fault mode.

[0197]

[0198] In the data preprocessing stage, the collected vibration signals are first denoised and smoothed. Then, the start and end times of each cylinder's working cycle are calculated based on the rotational speed signal. Finally, single-working-cycle vibration data samples are extracted from the preprocessed signals, with 1000 samples extracted for each fault mode for subsequent feature extraction, model training, and testing. This experiment aims to simulate typical diesel engine cylinder faults under actual operating conditions. By collecting real vibration signals and processing them systematically, the proposed method's ability to extract cylinder fault features and its fault diagnosis accuracy in high-noise environments are verified, providing a reliable basis for the practical application of the method.

[0199] To verify the effectiveness of the MMFE feature vector proposed in this invention in diesel engine cylinder fault diagnosis, this embodiment designed a set of comparative experiments. By comparing the diagnostic performance of the KELM model under different feature inputs, the advantages of the MMFE feature vector are evaluated.

[0200] Experiment based on Figure 9 The prediction confusion matrix shown in the figure illustrates the classification results for each fault mode after different feature vectors are input into the KELM model. For comparison, five control groups were set up, each using traditional feature extraction methods to construct feature vectors and inputting them into the KELM model for fault diagnosis, as detailed below:

[0201] like Figure 9 As shown in (a), T-KELM extracts 10 time-domain statistical features (such as mean, variance, peak value, etc.) and inputs them into the KELM model;

[0202] like Figure 9 As shown in (b), F-KELM extracts 10 frequency domain statistical features (such as spectral energy, centroid frequency, etc.) and inputs them into the KELM model;

[0203] like Figure 9 As shown in (c), TF-KELM extracts 10 time-domain and 10 frequency-domain statistical features simultaneously and inputs them into the KELM model;

[0204] like Figure 9 As shown in (d), C-KELM: The vibration signal is decomposed into three layers using db2 wavelet packets, and eight frequency band energy features are extracted and input into the KELM model;

[0205] like Figure 9 As shown in (e), TC-KELM extracts 10 time-domain statistical features and 8 wavelet packet energy features, and inputs them into the KELM model.

[0206] The experimental data samples were divided into training and test sets in a 7:3 ratio. The test set contained 300 samples for five typical fault modes. Experimental results show that the traditional feature extraction methods described above have certain limitations in diagnostic accuracy: the test accuracies of the T-KELM, F-KELM, TF-KELM, C-KELM, and TC-KELM models were 72%, 73%, 82%, 77.7%, and 85.3%, respectively.

[0207] In comparison, such as Figure 9 As shown in (f), the MMFE feature vector extracted by this invention (constructed by the multi-level multi-scale fuzzy entropy method, which integrates the high and low frequencies and multi-scale information of the vibration signal) achieves a fault diagnosis accuracy of 92.3% on the test set after being input into the same KELM model.

[0208] The comparative experiments above clearly show that, compared with traditional time-domain, frequency-domain, or wavelet energy features, the MMFE feature vector constructed in this invention can more comprehensively reflect the vibration signal characteristics under diesel engine cylinder fault conditions, and exhibits higher classification accuracy in the KELM model, verifying its effectiveness and superiority in diesel engine cylinder fault diagnosis.

[0209] To verify the impact of parameter optimization on the diagnostic performance of the KELM model, this embodiment compares and analyzes the fault diagnosis effects of the standard KELM model and the PSO-KELM model with optimized parameters using the PSO algorithm. The relevant results are as follows: Figure 10 As shown.

[0210] Figure 10 The diagnostic results of two models under the same test conditions are presented. The standard KELM model uses default or empirically set regularization parameter C and kernel function parameter σ, while the PSO-KELM model automatically optimizes these two key parameters using a particle swarm optimization algorithm to improve the model's classification performance on the test set. Figure 10 In the text, (a) represents the prediction accuracy of both methods. Figure 10 In the diagram, (b) represents the predicted loss for both. Figure 10 In this context, (c) represents the prediction confusion matrix using the MMFE-PSO-KELM method provided by this invention.

[0211] The experiment also used 300 test samples from five typical cylinder failure modes. The results show that the KELM model without parameter optimization is greatly affected by parameter settings and has relatively low diagnostic accuracy; while the PSO-KELM model optimized by the PSO algorithm can effectively select the regularization parameter C and kernel function parameter σ that are adapted to the current data distribution, thereby significantly improving the model's generalization ability, reducing the risk of overfitting or underfitting, and minimizing the loss of diagnostic performance.

[0212] Test results show that the PSO-KELM model achieves an overall diagnostic accuracy of 98.7% across five fault modes, a significant improvement over the standard KELM model.

[0213] Optimizing the key parameters of the KELM model using the PSO algorithm can effectively improve the accuracy and stability of the model in diesel engine cylinder fault diagnosis tasks, verifying the positive role of parameter optimization in improving diagnostic accuracy, and is an important component of the method of this invention.

[0214] Example 2

[0215] See Figure 11This invention also provides a system for diesel engine cylinder fault diagnosis based on MMFE-PSO-KELM, as provided in Embodiment 1, comprising the following six functional modules: a data acquisition module, a data preprocessing module, a feature extraction module, a diagnostic model optimization module, and a fault diagnosis module. These modules work collaboratively to achieve a complete diagnostic process from vibration signal acquisition to fault mode identification. Specifically,

[0216] The data acquisition module is used to collect vibration signals from the cylinder head surface of the diesel engine and crankshaft rotation speed signals. By placing vibration acceleration sensors on the cylinder head surface and installing speed sensors at the crankshaft end, the vibration response and crankshaft rotation state during cylinder operation are acquired synchronously. This module provides the raw data foundation for subsequent signal processing and fault analysis, ensuring the synchronization and authenticity of the signal sources.

[0217] The data preprocessing module mainly includes three sub-functions: calculation of work cycle information, noise reduction and smoothing of vibration signals, and data sample division based on work cycle.

[0218] Working cycle information calculation: The pulse counting method is adopted. The crankshaft angle corresponding to a single pulse is set according to the crankshaft signal disk parameters. Combined with the number of cylinders, number of strokes and crankshaft reference position of the diesel engine, the top dead center reference value of the cylinder is calculated. Furthermore, the current crankshaft position and the total angle rotated are determined by pulse counting, thereby calculating the start and end times of the current working cycle of the cylinder.

[0219] Signal denoising and smoothing: The acquired vibration signal is first filtered out by a low-pass filter (such as a Butterworth low-pass filter) to remove random noise interference such as exhaust and radiation; then a mean smoothing window is used to smooth the denoised signal to reduce residual high-frequency noise and stabilize the signal waveform.

[0220] Data sample segmentation: Based on the start and end times of the work cycle calculated above, the preprocessed vibration signal is segmented into segments according to each work cycle to form a data sample set containing the vibration response of a single work cycle. This ensures that subsequent feature extraction is performed within a unified work cycle, thereby improving the consistency and representativeness of the features.

[0221] This module provides high-quality, time-consistent input data for feature extraction by accurately acquiring working cycle information and performing noise reduction, smoothing, and segmentation processing on the original signal, which helps to accurately capture fault-related vibration features in noisy environments.

[0222] The feature extraction module performs multi-level, multi-scale analysis on the preprocessed vibration data samples to construct MMFE feature vectors, specifically including:

[0223] High- and low-frequency mode decomposition: A multi-level high- and low-frequency decomposition operator (such as low-frequency operator and high-frequency operator) is used to perform k-level decomposition on a single working cycle vibration signal to obtain 2 k There are high- and low-frequency mode signal sequences, each of which reflects the local characteristics of the signal in a specific frequency band.

[0224] Multi-scale analysis: Each modal signal sequence is divided into multiple scales (the scale parameter is usually taken as an integer power of 2) to generate time series at multiple scales in order to capture the performance of fault characteristics at different time resolutions.

[0225] Fuzzy entropy calculation: For each modality-scale time series, calculate the fuzzy entropy (FEn), quantify its complexity and regularity, and finally combine the fuzzy entropy values ​​of all modality-scale combinations to achieve a length of 2. k The MMFE eigenvectors.

[0226] This module comprehensively extracts high- and low-frequency and multi-scale fault mode information contained in vibration signals through multi-level and multi-scale signal decomposition and fuzzy entropy quantization, providing highly discriminative input features for subsequent diagnostic models and effectively solving the problem of incomplete feature extraction under strong noise.

[0227] The diagnostic model optimization module uses the Particle Swarm Optimization (PSO) algorithm to search for the regularization parameter C and kernel function parameters of the KELM model. The optimal combination, specifically the process includes:

[0228] Population initialization and fitness calculation: Initialize the PSO population, where each particle represents a set of parameter combinations (C, The fitness of each particle is evaluated based on the classification accuracy (fitness function) of the validation set, and the individual optimal and global optimal positions are updated.

[0229] Parameter search and iterative optimization: By iteratively updating the velocity and position of particles, the system continuously searches towards the optimal parameter region until it finds the parameter combination that maximizes the fitness function (classification accuracy), i.e., the optimal regularization parameters and kernel function parameters of the KELM model.

[0230] This module automatically optimizes the model through intelligent optimization algorithms, avoiding the subjectivity and inefficiency of manual parameter tuning. This ensures that the KELM model has good generalization ability under complex data distributions, laying the foundation for high-precision fault diagnosis.

[0231] The fault diagnosis module, based on the aforementioned optimized KELM parameter combination, constructs an optimized KELM fault diagnosis model and uses the MMFE feature vector as input to perform fault mode identification for diesel engine cylinders. Specifically, it includes:

[0232] Model building and training: Optimal parameters C and Substitute the KELM model into the training set and train the model to learn the mapping relationship between MMFE feature vectors and fault categories.

[0233] Model testing and validation: Evaluate the model's diagnostic performance (such as classification accuracy) on the test set. If it meets the preset requirements, the model is used as the final optimized KELM fault diagnosis model for actual fault identification; otherwise, adjust the parameters or features until the performance meets the standards.

[0234] This module employs an optimized KELM model and combines it with MMFE feature vectors that comprehensively reflect fault characteristics to achieve accurate identification of diesel engine cylinder fault modes, significantly improving the accuracy and reliability of diagnosis and solving the problem of low diagnostic accuracy under strong noise conditions in traditional methods.

[0235] This system, through the organic coordination of its various functional modules, achieves full automation and intelligence across the entire process from vibration signal acquisition, preprocessing, and feature extraction to model optimization and fault diagnosis. The data preprocessing module effectively improves signal quality, the feature extraction module comprehensively captures multi-dimensional fault-related information, the diagnostic model optimization module ensures optimal configuration of KELM model parameters, and finally, the fault diagnosis module achieves high-precision fault mode recognition. The system's overall design is highly targeted, effectively solving the technical problems of incomplete feature extraction for diesel engine cylinder faults under strong noise and the difficulty in optimizing KELM parameters, which leads to low diagnostic accuracy. It possesses high practical value and reliability.

[0236] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct or indirect applications in other related technical fields, are within the patent protection scope of the present invention.

Claims

1. A diesel engine cylinder fault diagnosis method based on MMFE-PSO-KELM, characterized in that, The steps and methods include: S1. Collect vibration signals from the surface of the diesel engine cylinder head and crankshaft speed signals, and preprocess the vibration signals. S2. The MMFE algorithm is used to analyze and process the preprocessed vibration signal to construct MMFE feature vectors rich in high and low frequency fault mode information. S3. Optimize the regularization parameter C and kernel function parameter σ of the KELM model using the PSO algorithm, and search for the optimal combination of parameters that maximizes the classification accuracy, with the classification accuracy as the fitness function. S4. Construct a KELM diagnostic model based on the optimal parameter combination, and use the high and low frequency fault mode information extracted by the MMFE algorithm as input to perform fault diagnosis on the diesel engine cylinder.

2. The diesel engine cylinder fault diagnosis method based on MMFE-PSO-KELM according to claim 1, characterized in that, In step S1, when preprocessing the vibration signal, the pulse counting method is used to obtain the information of each working cycle of the cylinder. The specific steps of obtaining the information of each working cycle of the cylinder using the pulse counting method are as follows: Based on the crankshaft signal disk parameters, set the crankshaft rotation angle corresponding to a single pulse; Calculate the top dead center reference value of the cylinder based on the number of cylinders, number of strokes, and crankshaft reference position of the diesel engine; The pulse count calculates the current crankshaft position and the total angle rotated, and combines this with the top dead center reference position to calculate the start and end times of the current working cycle of the cylinder.

3. A diesel engine cylinder fault diagnosis method based on MMFE-PSO-KELM according to claim 1 or 2, characterized in that, The specific steps for preprocessing the vibration signal in step S1 are as follows: S11. Use a low-pass filter to eliminate random noise in the vibration signal; S12. Use a smoothing window to reduce high-frequency noise components in the denoised signal; S13. Based on the crankshaft signal disk parameters, set the crankshaft angle corresponding to a single pulse. Combine the number of cylinders, number of strokes, and crankshaft reference position of the diesel engine to calculate the cylinder top dead center reference value. Calculate the total angle corresponding to the current crankshaft position using the pulse counting method, and then determine the start and end times of the cylinder's working cycle. Segment the smoothed vibration signal according to the start and end times of the working cycle to obtain a single working cycle data sample set. .

4. The diesel engine cylinder fault diagnosis method based on MMFE-PSO-KELM according to claim 3, characterized in that, In step S2, the steps for constructing the MMFE feature vector are as follows: S21. The preprocessed data sample set Single-cycle vibration response sequence Perform k-level high- and low-frequency decomposition to obtain A length of High and low frequency mode signal sequences , where N is Data length, The sequence number is the index of the modal signal sequence. ; S22, Vibration response sequence Each mode signal sequence Multi-scale analysis was performed to obtain a length of A multi-scale time series ,in For scale parameters; S23, Vibration response sequence Calculate the signal sequences of each mode separately. Multiscale time series Fuzzy entropy The construction length is The MMFE eigenvectors.

5. A diesel engine cylinder fault diagnosis method based on MMFE-PSO-KELM according to claim 4, characterized in that, In step S21, the method for generating the high-frequency and low-frequency mode signal sequences is as follows: Constructing a one-dimensional time series The high- and low-frequency decomposition operators are as follows: Low-frequency operators: ; High-frequency operators: ; in, and They are respectively for The high- and low-frequency mode signal sequences obtained by performing a single-level decomposition have a length of ; After k-level decomposition, we obtain A length of The high- and low-frequency mode signal sequences, wherein the first... A sequence of modal signals The signal X is obtained by sequentially selecting low-frequency or high-frequency operators to perform k-level decomposition. The selection of each layer of operators is determined by parameters. It is confirmed that, among them, Indicates that the calculation is not greater than The smallest integer.

6. The diesel engine cylinder fault diagnosis method based on MMFE-PSO-KELM according to claim 4, characterized in that, In step S23, the fuzzy entropy of the time series is calculated. The specific steps are as follows: S231. Set the embedding dimension parameter m and the similarity tolerance parameter r; S232, For one-dimensional time series N represents the total number of sampling points in the entire signal, which is then transformed into an m-dimensional vector. ; S233, Calculation Chebyshev distance between two adjacent m-dimensional vectors ; S234. Calculate the fuzzy similarity between two adjacent m-dimensional vectors. ; S235. Further calculate the m+1 dimensional fuzzy similarity. Finally, the fuzzy entropy of X is calculated using the following formula, as shown below: 。 7. A diesel engine cylinder fault diagnosis method based on MMFE-PSO-KELM according to claim 1 or 2, characterized in that, In step S3, the specific steps for finding the optimal parameter combination of the KELM model are as follows: S31. Initialize the population: Initialize the parameters of PSO-KELM, initialize the position and velocity of each particle, and calculate the initial fitness. S32. Iteratively update the individual optimal position and global optimal position of the particles, updating the velocity and position according to the following formula, the specific formula of which is shown below. ； In the formula, The inertial weight of a particle. and It is the learning factor, i.e., the speedup constant. and It is a uniformly random number between 0 and 1, used to limit the position and velocity of particles; , and These are the position, velocity, and historical best position of the i-th particle, respectively. It is the optimal position of the entire particle swarm, i.e., the global optimum. S33: When the fitness reaches the allowable error range, output the optimal parameter combination (C,σ).

8. A diesel engine cylinder fault diagnosis method based on MMFE-PSO-KELM according to claim 7, characterized in that, In step S3, the fitness function that maximizes classification accuracy takes the following form: ； In the formula, It is about optimizing the combination of parameters. Indicates the length of the validation set. It is an indicator function; a correct classification is 1, and a wrong classification is 0. and These are the actual class and the predicted class of the i-th sample, respectively.

9. A diesel engine cylinder fault diagnosis method based on MMFE-PSO-KELM according to claim 1 or 2, characterized in that, In step S4, the specific steps for constructing the KELM diagnostic model are as follows: S41. Output-based regularization parameters and kernel function parameters The optimal combination is used to construct the original KELM fault diagnosis model; S42. Divide the preprocessed diesel engine cylinder head surface vibration data into training set and test set, and calculate and construct MMFE feature vector; S43. The original KELM fault diagnosis model is trained and tested using the training set and the test set until the diagnostic performance of the original KELM fault diagnosis model meets the preset requirements, and then used as the optimized KELM diagnosis model.

10. A diesel engine cylinder fault diagnosis system based on MMFE-PSO-KELM, characterized in that, include: The data acquisition module is used to collect vibration signals from the surface of the diesel engine cylinder head and crankshaft speed signals. The data preprocessing module is used to calculate the cylinder working cycle information, and to perform noise reduction, smoothing and working cycle division on the vibration signal to obtain single working cycle data samples. The feature extraction module is used to process the preprocessed single working cycle vibration signal using the MMFE algorithm to construct an MMFE feature vector containing high and low frequency fault mode information. The diagnostic model optimization module is used to search for the optimal combination of the regularization parameter C and the kernel function parameter σ of the KELM model using the PSO algorithm. The fault diagnosis module is used to construct a KELM diagnostic model based on the optimal parameter combination and to realize diesel engine cylinder fault diagnosis using MMFE feature vectors as input.

Images