A diesel generator set fault detection method and system based on deep learning

Abstract

This invention relates to the field of fault detection technology and discloses a method and system for fault detection of diesel generator sets based on deep learning. The method includes: acquiring first vibration signal data and performing time-frequency decomposition to obtain dynamic change features; denoising the dynamic change features to obtain a time-frequency feature sequence; extracting peak energy distribution dataset and calculating the entropy value of each frequency band to obtain a frequency band entropy value sequence; classifying the frequency band entropy value sequence to determine the random fluctuation baseline mode and separating abnormal frequency components; calculating the spectral line spacing and amplitude ratio to obtain a spectral feature dataset and classifying it to determine the fault classification result; acquiring the current second vibration signal data and performing similarity calculation with a pre-established normal pattern library to output a fault feature vector; verifying the fault feature vector to obtain the final fault detection result. This method can realize closed-loop diagnosis from signal acquisition to fault classification, improving the fault detection accuracy of diesel generator sets.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, and in particular to a fault detection method and system for diesel generator sets based on deep learning. Background Technology

[0002] Diesel generator sets are critical equipment in industrial production, emergency power supply, and infrastructure operation. Their operational stability and reliability directly affect the continuity of energy supply and production safety. Given the continuous growth in electricity demand, timely and effective fault detection of generator sets is an important means of reducing downtime risks. Vibration signals can directly reflect the operating status of internal mechanical components and are a core analytical object in the field of fault detection.

[0003] In existing technologies, fault detection methods for vibration signals of diesel generator sets mostly rely on fixed threshold judgment or simple statistical analysis. These methods establish detection models based on single frequency or time-domain features, and determine whether the equipment is abnormal by setting a fixed threshold for vibration amplitude. However, the operating environment of diesel generator sets is complex, and vibration signals are easily affected by factors such as load fluctuations, environmental noise, and equipment aging, exhibiting significant nonlinearity and randomness. Fixed threshold methods are difficult to adapt to dynamically changing operating conditions, often misjudging normal random fluctuations as faults, or missing early weak fault signals due to excessively high threshold settings. At the same time, simple statistical analysis has limited ability to extract high-dimensional time-frequency features, and cannot fully capture the dynamic evolution of fault signals in the time and frequency domains, resulting in insufficient adaptability and accuracy of the detection system under complex operating conditions.

[0004] In summary, under dynamic operating conditions, the fault frequency components of diesel generator sets are easily submerged by random fluctuations, making them difficult to accurately separate and classify, resulting in insufficient accuracy in fault detection of diesel generator sets. Summary of the Invention

[0005] This invention provides a method and system for fault detection of diesel generator sets based on deep learning, so as to realize closed-loop diagnosis from signal acquisition to fault classification and improve the accuracy of fault detection.

[0006] In a first aspect, to address the aforementioned technical problems, this invention provides a deep learning-based fault detection method for diesel generator sets, comprising: The first vibration signal data of the diesel generator set during operation is acquired, and the first vibration signal data is decomposed into time and frequency to obtain dynamic change characteristics. Adaptive filtering is performed on the aforementioned dynamic change characteristics to obtain a denoised time-frequency feature sequence; The peak energy distribution dataset is extracted from the time-frequency feature sequence. Based on the peak energy distribution dataset, the entropy value of each frequency band is calculated to obtain the frequency band entropy value sequence. The frequency band entropy sequence is classified by k-means clustering algorithm to determine the random fluctuation baseline pattern. Based on the random fluctuation baseline pattern, the abnormal frequency components are separated from the peak energy distribution dataset. Based on the abnormal frequency components, the spectral line spacing and amplitude ratio are calculated to obtain a spectral line feature dataset. The spectral line feature dataset is then classified based on a preset fault mode library to determine the fault classification result. Acquire the current second vibration signal data, calculate the similarity between the second vibration signal data and the pre-established normal pattern library, and if the similarity is lower than the preset similarity threshold, output the corresponding fault feature vector according to the fault classification result. Based on the normal pattern library, the fault feature vector is verified using the random forest algorithm to obtain the final fault detection result.

[0007] Preferably, the step of performing time-frequency decomposition on the first vibration signal data to obtain dynamic change characteristics includes: The first vibration signal data is subjected to analog-to-digital conversion to obtain a digital vibration signal; Wavelet transform is used to perform time-frequency decomposition on the digital vibration signal to generate a time-spectrum diagram containing time and frequency information. The time-spectrum diagram is divided into preset time windows, and the frequency components of each time window are calculated by Fourier transform to obtain the dynamic change characteristics.

[0008] Preferably, the step of performing adaptive filtering on the dynamically changing features to obtain the denoised time-frequency feature sequence includes: To address the aforementioned dynamic change characteristics, a wavelet denoising method is used for preliminary denoising to obtain the first feature dataset after denoising. For the first feature dataset, discrete cosine transform is used to compress the data to obtain a denoised time-frequency feature sequence.

[0009] Preferably, the step of extracting the peak energy distribution dataset from the time-frequency feature sequence, and calculating the entropy value of each frequency band based on the peak energy distribution dataset to obtain a frequency band entropy value sequence includes: Based on the time-frequency feature sequence, the time-domain signal is converted into a frequency-domain signal using a fast Fourier transform, and the peak energy distribution features of the signal at different frequencies are extracted to obtain the peak energy distribution dataset. Based on the peak energy distribution dataset, the energy values ​​of each frequency band are normalized into a probability distribution. The Shannon entropy of each frequency band is calculated based on the probability distribution. The Shannon entropy is then arranged in order of frequency band to obtain a sequence of frequency band entropy values.

[0010] Preferably, the step of classifying the frequency band entropy sequence using the k-means clustering algorithm to determine a random fluctuation baseline pattern, and separating the abnormal frequency components from the peak energy distribution dataset based on the random fluctuation baseline pattern, includes: When the fluctuation amplitude of the frequency band entropy value sequence is higher than the preset fluctuation threshold, the frequency band entropy value sequence is classified by the k-means clustering algorithm to obtain the random fluctuation benchmark pattern; Based on the random fluctuation baseline pattern, a normal energy threshold is determined, and frequency bands whose peak energy exceeds the preset energy threshold are selected from the peak energy distribution dataset and identified as abnormal frequency components.

[0011] Preferably, the step of calculating the spectral line spacing and amplitude ratio based on the abnormal frequency components to obtain a spectral line feature dataset, classifying the spectral line feature dataset based on a preset fault mode library, and determining the fault classification result includes: An abnormal feature sequence is obtained from the abnormal frequency components, and the spectral line spacing and amplitude ratio of the abnormal feature sequence are calculated by fast Fourier transform to obtain the spectral line feature dataset. The spectral line feature dataset is normalized to obtain a normalized feature vector set; When the dimension of the normalized feature vector set meets the preset dimension threshold, the normalized feature vector set is divided by a support vector machine to obtain a fault type label set. The fault type label set is matched with the fault pattern library to obtain the fault classification result.

[0012] Preferably, the step of calculating the similarity between the second vibration signal data and a pre-established normal pattern library, and if the similarity is lower than a preset similarity threshold, then based on the fault classification result, outputting the corresponding fault feature vector, including: For the second vibration signal data, high-frequency noise above the preset frequency cutoff point is filtered out by Fourier transform, and low-frequency effective signals are retained. The low-frequency effective signals are then standardized to obtain a real-time signal feature set. Calculate and normalize the distance between the real-time signal feature set and each pattern in the normal pattern library to obtain a normalized distance set. Subtract each distance value in the normalized distance set from the value 1 to form a similarity set. If the proportion of similarity values ​​below a preset similarity threshold in the similarity set exceeds a preset ratio threshold, then based on the fault classification result, the abnormal mode of the real-time signal feature set is determined, and the corresponding fault feature vector is extracted.

[0013] Secondly, the present invention provides a diesel generator set fault detection system based on deep learning, comprising: The signal time-frequency decomposition module is used to acquire the first vibration signal data during the operation of the diesel generator set, and to perform time-frequency decomposition on the first vibration signal data to obtain dynamic change characteristics. The dynamic denoising module is used to perform adaptive filtering on the dynamically changing features to obtain a denoised time-frequency feature sequence. The frequency band entropy calculation module is used to extract the peak energy distribution dataset from the time-frequency feature sequence, and calculate the entropy value of each frequency band based on the peak energy distribution dataset to obtain the frequency band entropy value sequence. An abnormal frequency separation module is used to classify the frequency band entropy value sequence using a k-means clustering algorithm, determine a random fluctuation baseline pattern, and separate abnormal frequency components from the peak energy distribution dataset based on the random fluctuation baseline pattern. The fault classification determination module is used to calculate the spectral line spacing and amplitude ratio based on the abnormal frequency components to obtain a spectral line feature dataset, classify the spectral line feature dataset based on a preset fault mode library, and determine the fault classification result. The fault feature output module is used to acquire the current second vibration signal data, calculate the similarity between the second vibration signal data and the pre-established normal pattern library, and if the similarity is lower than the preset similarity threshold, output the corresponding fault feature vector according to the fault classification result. The random forest verification module is used to verify the fault feature vectors based on the normal pattern library using the random forest algorithm to obtain the final fault detection result.

[0014] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention obtains the dynamic change characteristics by acquiring the first vibration signal data of the diesel generator set during operation and performing time-frequency decomposition. Adaptive filtering is then applied to the dynamic change characteristics to obtain a denoised time-frequency feature sequence. This process can first comprehensively capture the dynamic information of the vibration signal in the time and frequency dimensions, and then dynamically suppress interference factors such as environmental noise and load fluctuations through adaptive filtering, avoiding noise from masking key fault characteristics. This makes the extracted vibration data more consistent with the actual operating state of the equipment, thereby improving the reliability of fault detection.

[0015] (2) This invention extracts a peak energy distribution dataset from the denoised time-frequency feature sequence, calculates the entropy value of each frequency band based on the peak energy distribution dataset to obtain a frequency band entropy value sequence, and then classifies the frequency band entropy value sequence using a k-means clustering algorithm to determine a random fluctuation benchmark pattern, and separates abnormal frequency components based on this pattern. This series of operations can quantify the energy distribution and randomness of each frequency band from the denoised features, establish a random fluctuation benchmark that fits the dynamic operating conditions through a clustering algorithm, accurately distinguish between normal fluctuations and abnormal frequencies, and effectively capture the frequency components corresponding to early weak faults.

[0016] (3) This invention obtains a spectral feature dataset by calculating the spectral line spacing and amplitude ratio based on abnormal frequency components, and classifies the spectral feature dataset based on a preset fault mode library to determine the fault classification result. This classification process can transform abstract abnormal frequency components into spectral features with clear fault orientation, and then achieve accurate correspondence of fault types by matching with the fault mode library, avoiding misjudgment caused by unclear correlation between features and faults, making the fault classification result more targeted.

[0017] (4) This invention acquires the current second vibration signal data, calculates its similarity with a pre-established normal pattern library, and outputs a fault feature vector if the similarity is lower than a preset similarity threshold. Then, it uses a random forest algorithm to verify the fault feature vector in conjunction with the normal pattern library to obtain the final fault detection result. The reliability of the abnormal signal is further confirmed through similarity comparison, and the multi-dimensional verification capability of the random forest algorithm is used to comprehensively evaluate the fault feature vector, reducing the limitations of single feature judgment and ensuring that the final output fault detection result has both accuracy and credibility. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an embodiment of the deep learning-based diesel generator set fault detection method provided by the present invention. Figure 2 This is a schematic diagram of an embodiment of the diesel generator set fault detection system based on deep learning provided by the present invention. Detailed Implementation

[0019] 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 some embodiments of the present invention, and not all embodiments. 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.

[0020] Reference Figure 1The first embodiment of the present invention provides a fault detection method for diesel generator sets based on deep learning, including steps S11 to S17, as follows: S11, acquire the first vibration signal data of the diesel generator set during operation, and perform time-frequency decomposition on the first vibration signal data to obtain dynamic change characteristics; S12, perform adaptive filtering on the dynamic change features to obtain the denoised time-frequency feature sequence; S13, extract the peak energy distribution dataset from the time-frequency feature sequence, and calculate the entropy value of each frequency band based on the peak energy distribution dataset to obtain the frequency band entropy value sequence; S14, the frequency band entropy sequence is classified by k-means clustering algorithm to determine the random fluctuation benchmark pattern, and the abnormal frequency components are separated from the peak energy distribution dataset based on the random fluctuation benchmark pattern; S15, Calculate the spectral line spacing and amplitude ratio based on the abnormal frequency components to obtain a spectral line feature dataset, classify the spectral line feature dataset based on a preset fault mode library, and determine the fault classification result; S16, acquire the current second vibration signal data, calculate the similarity between the second vibration signal data and the pre-established normal pattern library, and if the similarity is lower than the preset similarity threshold, output the corresponding fault feature vector according to the fault classification result. S17. Based on the normal pattern library, the fault feature vector is verified using the random forest algorithm to obtain the final fault detection result.

[0021] In step S11, first vibration signal data of the diesel generator set during operation is acquired, and time-frequency decomposition is performed on the first vibration signal data to obtain dynamic change characteristics, including: The first vibration signal data is subjected to analog-to-digital conversion to obtain a digital vibration signal; Wavelet transform is used to perform time-frequency decomposition on the digital vibration signal to generate a time-spectrum diagram containing time and frequency information. The time-spectrum diagram is divided into preset time windows, and the frequency components of each time window are calculated by Fourier transform to obtain the dynamic change characteristics.

[0022] It should be noted that the first vibration signal data is collected by accelerometers installed in key parts of the diesel generator set (such as bearings and the engine block). These sensors can directly capture the vibrations generated by the movement of mechanical components during generator operation. Under the conventional signal output standards of industrial-grade sensors, the amplitude of the output analog vibration signal is generally in the range of -5V to 5V, which can cover the vibration signal intensity of the diesel generator set from normal operation to minor fault conditions. An analog-to-digital converter (ADC) is a device that converts continuously changing analog signals into discrete digital signals. Its working process mainly includes three steps: sampling, quantization, and encoding. First, the analog vibration signal is sampled according to the sampling frequency to obtain a series of discrete time points of signal amplitude; then, these amplitudes are mapped to a finite number of discrete numerical levels; finally, the quantized values ​​are converted into binary code to form a digital vibration signal. To ensure that subtle features in the vibration signal are not lost, a 16-bit ADC with a resolution of 0.0001526V can be selected, which can divide the signal into 65536 levels and accurately characterize the minute fluctuations of the vibration signal. It should be noted that this voltage value is linearly related to the vibration acceleration value measured by the sensor. The signal amplitude mentioned in subsequent processing refers to this equivalent voltage value for relative comparison and analysis.

[0023] In this embodiment, wavelet transform is a mathematical method that can analyze signals simultaneously in the time and frequency domains. By convolving the signal with a series of wavelet basis functions at different scales and positions, multi-resolution analysis of the signal is achieved. In this invention, the Daubechies wavelet basis is selected to decompose the digital vibration signal into wavelet components of different scales, thereby generating a time-spectrum graph containing both time and frequency axes. The grayscale or color depth at different positions in the time-spectrum graph represents the signal energy intensity at the corresponding time point and frequency. For example, when performing wavelet transform on the digital vibration signal of a 500kW diesel generator set, the generated time-spectrum graph can cover the frequency range of 0-500Hz. If a significant high-energy concentration region appears near the 100Hz frequency in the time-spectrum graph within a certain time period, it can be preliminarily determined that the unit may have a rotor imbalance problem during that time period.

[0024] It is worth noting that the duration of the preset time window needs to be determined in conjunction with the sensor sampling frequency and the time scale of the unit's vibration characteristics. Taking a sampling frequency of 1000Hz as an example, a 1-second time window can be set, containing 1000 sampling points. This satisfies the data volume requirements of the Fourier transform while accurately reflecting the short-term frequency changes of the vibration signal, avoiding the obscuring of frequency change details due to an excessively long time window. The basic principle of the Fourier transform is to decompose a complex time-domain signal into a sum of simple harmonic components of different frequencies by calculating the correlation between the signal and sine or cosine functions of different frequencies. For a signal segment divided by a time window, the Fast Fourier Transform (FFT) can efficiently calculate the amplitude or energy of each frequency component in the segment, i.e., the frequency component, thereby quantifying the main frequency composition and intensity of each frequency of the vibration signal within the corresponding time window. By summarizing the frequency components of all time windows, the dynamic change characteristics are obtained, reflecting the frequency distribution changes of the vibration signal at different time points.

[0025] In step S12, adaptive filtering is performed on the dynamically changing features to obtain a denoised time-frequency feature sequence, including: To address the aforementioned dynamic change characteristics, a wavelet denoising method is used for preliminary denoising to obtain the first feature dataset after denoising. For the first feature dataset, discrete cosine transform is used to compress the data to obtain a denoised time-frequency feature sequence.

[0026] It should be noted that the core of wavelet denoising is to utilize the time-frequency localization capability of wavelet transform to separate noise and effective signal components in dynamic features. First, a wavelet basis suitable for the vibration signal characteristics of the diesel generator set is selected, and the dynamic features are decomposed into multiple wavelet components of different scales. High-frequency wavelet components mostly correspond to environmental noise (such as airflow noise during equipment operation, external mechanical interference, etc.), while low-frequency wavelet components correspond to the effective signal generated by the vibration of the unit's mechanical components. Then, the signal standard deviation is calculated. σ To obtain a robust threshold ,in N The number of sampling points for the signal segment is given. High-frequency wavelet components with signal amplitudes exceeding a robustness threshold are filtered out, retaining only the low-frequency components containing valid information, resulting in the first feature dataset after denoising. Taking a 250kW diesel generator set as an example, its vibration signal sampling frequency is 2000Hz, and the time-spectrum map corresponding to the dynamic change features covers the frequency range of 0-1000Hz. To retain more details of the non-stationary mechanical vibration signal, the feature is decomposed into 5 layers of wavelet components using the Symlet wavelet basis. The high-frequency components of the 4th and 5th layers are filtered out, and the resulting first feature dataset can significantly reduce noise interference.

[0027] In one feasible approach, the least mean square algorithm from adaptive filtering is used for further denoising. First, the first feature dataset obtained through wavelet denoising is used as the input signal. x ( n Then, starting from the signal's initial position, take all sampling points within the first window, calculate the arithmetic mean of these sampling points, and use this average as the output value corresponding to the center position of the window; then move the window one sampling point along the time axis, repeating the operation of calculating the average, until the entire input signal has been traversed. x ( n This approximates the pure effective vibration signal, which is then used as the desired signal. d ( n Then initialize the filter weight vector. w ( n Each vector element is randomly selected within a small range of -0.01 to 0.01 to avoid initial output distortion, and a convergence factor μ is set with a value of 0.001 to 0.01 to balance convergence speed and stability. Furthermore, the filter is based on the current weight vector w ( n For input signals x ( n Perform weighted calculations to obtain real-time output. y ( n )= w ( n )∙ x ( n (∙ denotes the vector dot product), calculate the error signal e ( n )= d ( n )- y ( n ), representing the deviation between the desired effective signal and the filter output; finally, through iterative formulas... Iteratively adjust the weight vector until the error... e ( n An amplitude less than 0.02V indicates that residual noise has been effectively suppressed, and the output second feature dataset mainly retains the effective vibration characteristics of the unit itself. n =0, 1,…,N-1, where N is the number of sampling points. For example, when the input signal… x ( n The 50Hz noise component in the signal significantly leads to e ( n When the value increases, the weights will be corrected in the direction of enhancing the attenuation of the 50Hz frequency; if the effective signal is excessively suppressed, the error feedback will cause the weights to be adjusted in the opposite direction to preserve the features.

[0028] It's worth noting that the denoised second feature dataset contains a large number of data points. Directly using it for subsequent analysis would increase computational load and data storage. The Discrete Cosine Transform (DCT), however, can significantly compress the data volume while preserving key features. Specifically, the DCT converts the time-domain signal of the second feature dataset into frequency-domain cosine coefficients, and the adaptive filtering output... y ( n The frequency domain coefficient sequence obtained after DCT transformation Y ( k )( k The formula for calculating (=0,1,…,N-1) is: ,in, For normalization coefficients, when k When =0, ;when k When ≥1, , The kernel function represents the cosine basis functions at different frequencies. k The smaller the value, the lower the corresponding frequency. Vibration signal energy has a concentrated characteristic. Key features of effective unit vibration, such as the 80Hz main frequency of bearing rotation and the 160Hz harmonic of rotor imbalance, correspond to low-frequency coefficients with high energy proportions; while high-frequency coefficients are mostly noise residues or redundant information with extremely low energy. Data compression can be achieved by retaining the top N low-frequency coefficients with an energy proportion exceeding 95% and discarding redundant high-frequency coefficients.

[0029] For example, for a second feature dataset containing 10,000 data points obtained after adaptive filtering of a 250kW diesel generator set, 10,000 cosine coefficients can be generated by discrete cosine transform. The first 200 low-frequency coefficients contain more than 95% of the energy of the original signal, which is sufficient to characterize the core vibration characteristics of the generator set. At this time, only these 200 coefficients are retained, and the amount of data can be compressed to 2% of the original.

[0030] In step S13, a peak energy distribution dataset is extracted from the time-frequency feature sequence. Based on the peak energy distribution dataset, the entropy value of each frequency band is calculated to obtain a frequency band entropy value sequence, including: Based on the time-frequency feature sequence, the time-domain signal is converted into a frequency-domain signal using a fast Fourier transform, and the peak energy distribution features of the signal at different frequencies are extracted to obtain the peak energy distribution dataset. Based on the peak energy distribution dataset, the energy values ​​of each frequency band are normalized into a probability distribution. The Shannon entropy of each frequency band is calculated based on the probability distribution. The Shannon entropy is then arranged in order of frequency band to obtain a sequence of frequency band entropy values.

[0031] It should be noted that for the time-frequency characteristic sequence, the sequence is first padded with zeros to the nearest power of 2, denoted as Nn; then, it is split into two subsequences of length Nn / 2 (even-indexed part, low-frequency trend part) and odd-indexed subsequence (high-frequency fluctuation part); the discrete Fourier transform results are calculated for each subsequence, and then combined using a butterfly operation to obtain the final frequency domain signal. This frequency domain signal contains Nn / 2 effective frequency points, and the square of the amplitude of each frequency point is the energy value of the corresponding frequency. The average and standard deviation of the energy values ​​of all frequency points in the frequency domain signal are calculated, and the average value is summed with three times the standard deviation to obtain the peak threshold. Points with energy values ​​greater than the peak threshold are identified as peak points, and the corresponding frequencies and energy values ​​are extracted as peak features. The resulting peak energy distribution dataset can intuitively reflect the concentration of unit vibration energy at different frequencies.

[0032] Furthermore, the frequency range is divided into several continuous, equally wide frequency bands. The bandwidth is determined based on the structural parameters of the diesel generator set (such as bearing type, gear teeth number, and rotor speed) to calculate the characteristic frequencies of its key components (such as fundamental frequency, harmonics, and bearing passing frequency). Using these as the center, non-uniform frequency bands covering these characteristic frequencies are set for energy statistics. Selectable bandwidths include 10Hz, 20Hz, 50Hz, and 100Hz. For small generator sets below 50kW, the vibration characteristic frequency is usually concentrated between 0-500Hz, so the 20Hz and 50Hz bandwidths can be emphasized for testing. For large generator sets above 500kW, the vibration characteristic frequency can extend to 0-2000Hz, requiring the inclusion of wider candidate bandwidths such as 100Hz. Next, the sum of all peak energies within each frequency band is calculated to obtain the energy value of each band. Dividing the energy value of each band by the total energy value of all bands yields the proportion of that band's energy to the total energy, which is the probability distribution. These proportions together form a probability distribution sequence, whose sum is 1.

[0033] In this embodiment, Shannon entropy H The calculation formula is: ,in m For the number of frequency bands, For the first i The probability distribution of each frequency band. Shannon entropy is used to quantify the randomness and complexity of energy distribution in each frequency band. The higher the entropy value, the more uniform the energy distribution in each frequency band, and the stronger the signal randomness. Arranging the Shannon entropy values ​​of each frequency band in order of frequency frequency yields a sequence of frequency band entropy values, which reflects the complexity of vibration energy distribution in different frequency ranges.

[0034] For example, for the time-frequency characteristic sequence of a 250kW diesel generator set, FFT processing was performed using a sampling frequency of 2000Hz and a length of 2048 data points to obtain a frequency domain signal of 0-1000Hz. Peak energies at frequencies such as 80Hz, 160Hz, and 320Hz were extracted to form a peak energy distribution dataset. The 0-1000Hz range was divided into 10 frequency bands of 100Hz each. The energy percentage of the first frequency band (0-100Hz) was 0.4, the percentage of the second frequency band (100-200Hz) was 0.3, and the percentages of the remaining frequency bands were 0.1, 0.05, 0.05, 0.03, 0.02, 0.02, 0.02, and 0.01, respectively. After calculating the entropy of each frequency band, the frequency band entropy value sequence was obtained by arranging the frequency bands in order. The frequency bands with low entropy values ​​in this sequence correspond to the energy concentration areas, reflecting the main vibration characteristics of the unit during normal operation.

[0035] In step S14, the frequency band entropy sequence is classified using the k-means clustering algorithm to determine a random fluctuation baseline pattern. Based on the random fluctuation baseline pattern, anomalous frequency components are separated from the peak energy distribution dataset, including: When the fluctuation amplitude of the frequency band entropy value sequence is higher than the preset fluctuation threshold, the frequency band entropy value sequence is classified by the k-means clustering algorithm to obtain the random fluctuation benchmark pattern; Based on the random fluctuation baseline pattern, a normal energy threshold is determined, and frequency bands whose peak energy exceeds the preset energy threshold are selected from the peak energy distribution dataset and identified as abnormal frequency components.

[0036] It should be noted that the fluctuation amplitude is quantified by the standard deviation of the frequency band entropy value sequence. The preset fluctuation threshold needs to be determined based on historical normal operation data. By analyzing a large set of standard deviations of frequency band entropy value sequences under normal operating conditions, the average and standard deviation of this set are calculated. The average plus twice the standard deviation is used as the preset fluctuation threshold, which can be set to 0.5 for conventional units. When the standard deviation of the entropy value sequence calculated in real time exceeds this threshold, it reflects an abnormal change in the dynamic characteristics of the vibration system, triggering k-means clustering analysis.

[0037] It's worth noting that the k-means clustering algorithm is an unsupervised learning model used to cluster frequency band entropy value sequences with similar characteristics into one class. The training data consists of the frequency band entropy value sequence triggered by clustering and historical entropy value sequences from the past 10 minutes. Each sequence contains entropy values ​​from m frequency bands, forming an m-dimensional feature vector. The value of k needs to be determined in conjunction with common operating states of the unit (such as stable operation, load switching transition, startup phase, etc.). Initially, the range of k is set to 3-5. Then, by calculating the clustering error (the sum of squared distances from all samples to their cluster centers) under different k values, a curve showing the error changing with k is plotted. The k value corresponding to the inflection point where the curve slope gradually decreases is taken as the optimal number of clusters.

[0038] During training, k feature vectors are first randomly selected from the training data as initial cluster centers. Then, the Euclidean distance between each sample and each cluster center is calculated, and the sample is assigned to the nearest cluster. The mean of each cluster is recalculated as the new cluster center. This assignment and update process is repeated until the change in cluster centers is less than a preset convergence threshold. Through this process, the frequency band entropy sequence is divided into k classes. The class with the highest sample size and the smallest entropy fluctuation is determined as the random fluctuation baseline pattern. This pattern represents the typical energy distribution fluctuation characteristics of the unit under current operating conditions, providing a reference benchmark for anomaly detection.

[0039] Furthermore, peak energy distribution data corresponding to the random fluctuation baseline model is extracted, and the historical maximum value or specific high quantile (such as the 99th percentile) of energy at each frequency point is calculated. This value is set as the normal energy threshold. Frequency bands that exceed the normal energy threshold of the corresponding frequency are considered to deviate from the typical range of the baseline model and are identified as abnormal frequency components, corresponding to early faults of the diesel generator set (such as characteristic frequencies generated by bearing wear) or sudden interference (such as impact vibrations caused by loose components).

[0040] For example, for a certain diesel generator set's frequency band entropy value sequence, its standard deviation was calculated to be 0.6, which is higher than the preset fluctuation threshold of 0.5, triggering k-means clustering analysis. 200 entropy value sequences from the most recent 10 minutes were selected as training data, each sequence containing entropy values ​​from 10 frequency bands. With k=3, the analysis converged after 15 iterations, resulting in three clusters: Cluster 1 (60%), with small entropy value fluctuations (standard deviation 0.2), corresponding to a stable operating mode; Cluster 2 (30%), with moderate entropy value fluctuations; and Cluster 3 (10%), with large entropy value fluctuations. Cluster 1 was determined as the random fluctuation baseline mode. In its corresponding peak energy distribution, the energy mean at the 80Hz frequency was 0.02V², with a standard deviation of 0.003V², and the calculated normal energy threshold was 0.029V². The peak energy distribution dataset at 80 Hz showed a peak energy of 0.035 V², which exceeded the normal energy threshold. Therefore, 80 Hz was identified as an abnormal frequency component, suggesting possible rotor imbalance or premature bearing wear.

[0041] In step S15, based on the abnormal frequency components, the spectral line spacing and amplitude ratio are calculated to obtain a spectral line feature dataset. The spectral line feature dataset is then classified based on a preset fault mode library to determine the fault classification result, including: An abnormal feature sequence is obtained from the abnormal frequency components, and the spectral line spacing and amplitude ratio of the abnormal feature sequence are calculated by fast Fourier transform to obtain the spectral line feature dataset. The spectral line feature dataset is normalized to obtain a normalized feature vector set; When the dimension of the normalized feature vector set meets the preset dimension threshold, the normalized feature vector set is divided by a support vector machine to obtain a fault type label set. The fault type label set is matched with the fault pattern library to obtain the fault classification result.

[0042] It should be noted that the anomalous feature sequence refers to a time-domain vibration signal segment within the frequency band of the anomalous frequency, containing the vibrational details that cause the energy anomaly. Processing the anomalous feature sequence using a Fast Fourier Transform (FFT) yields its corresponding frequency-domain spectral lines. A frequency-domain spectral line is a set of spectral lines consisting of discrete frequency points and their amplitudes corresponding to energy peaks in the frequency domain. Each spectral line represents the vibrational energy characteristic at a specific frequency, where the horizontal axis represents frequency and the vertical axis represents the energy amplitude at the corresponding frequency. The spectral line spacing refers to the frequency difference between adjacent anomalous frequency peaks; for example, the spectral line spacing between 100Hz and 200Hz is 100Hz. This value is related to parameters such as the rotational speed and number of teeth of mechanical components. The amplitude ratio is the ratio of the amplitude of the larger peak to the amplitude of the smaller peak between two anomalous frequency peaks, used to characterize the relative intensity of each anomalous component. The spectral feature dataset is obtained by integrating these components in the order of "spectral line spacing + amplitude ratio".

[0043] Furthermore, to eliminate dimensional differences between different features, the spectral feature dataset is normalized to uniformly map the data to the 0-1 interval. The normalization method uses min-max standardization, and the formula is as follows: ,in, x These are the original eigenvalues. For normalized eigenvalues, and These represent the maximum and minimum values ​​of the feature in historical data, respectively. The normalized feature vector set, after standardization, consists of multiple normalized spectral line spacings and amplitude ratios for each vector. The preset dimensionality threshold can be determined through validation experiments. On historical datasets containing various fault types, the classification accuracy of the support vector machine under different dimensionality thresholds is tested. The dimensionality range that achieves peak and stable accuracy is selected as the threshold standard. At the same time, excessively high dimensionality should be avoided to prevent increased model complexity. In the scenario of diesel generator set fault diagnosis, a threshold of 8-10 is suitable. A higher dimensionality in the normalized feature vector set indicates more anomalous feature information contained within it.

[0044] It's worth noting that Support Vector Machines (SVMs) are supervised learning models. They use kernel functions to map low-dimensional features to a high-dimensional space, constructing an optimal hyperplane to achieve linear separation of different categories. A commonly used kernel function is the Radial Basis Function (RBF), which effectively handles nonlinear feature relationships. The training data consists of normalized spectral feature vectors from historical fault cases, with each vector corresponding to a known fault type label (such as rotor imbalance, bearing wear, etc.). The training data is input into the SVM. The search range of the penalty parameter C is preset to [0.1, 1, 10, 100], covering the order of magnitude from low penalty to high penalty; the search range of the kernel function parameter γ is [0.01, 0.1, 1, 10], corresponding to the kernel function width from narrow to wide. Then, for each (C, γ) parameter combination, a 5-fold cross-validation method is used to divide the training data into 5 parts, and 4 parts are used for training and 1 part for validation in turn to obtain the average cross-validation accuracy. Finally, the parameter combination with the highest cross-validation accuracy is selected so that the model has the highest classification accuracy on the training set, and the optimal hyperplane is finally determined.

[0045] Finally, the input normalized feature vector set is partitioned using this hyperplane, and the fault type label corresponding to each vector is output, forming a fault type label set. Fault data verified through disassembly from historical fault cases is collected. For each fault type, such as bearing outer ring wear, rotor imbalance, and blade cracks, processes S11 to S14 are executed to extract abnormal frequency components from the vibration signal, and the spectral line spacing, amplitude ratio, and amplitude change trend of these frequencies as the fault develops are recorded. Similarly, these features are subjected to min-max normalization to form structured typical spectral feature templates. Each template corresponds to a feature vector, and these are finally classified and stored according to fault type, forming a fault pattern library. The fault type label set is matched with the preset fault pattern library, and the cosine similarity of the angle between the two vectors is calculated. The fault type corresponding to the template with the highest cosine similarity is taken as the fault classification result.

[0046] For example, an abnormal frequency component of 60Hz, 120Hz, and 240Hz was detected in a diesel generator set during operation, and the corresponding abnormal feature sequences were extracted. The sampling frequency was set to 1600Hz, and the number of time-domain sampling points was 2048. After Fast Fourier Transform (FFT), three main frequency domain spectral lines were obtained: 60Hz (amplitude 0.4V), 120Hz (amplitude 0.25V), and 240Hz (amplitude 0.1V). The spectral line spacing between 120Hz and 60Hz was calculated to be 60Hz, with an amplitude ratio of 1.6; the spectral line spacing between 240Hz and 120Hz was 120Hz, with an amplitude ratio of 2.5; and the amplitude ratio between 60Hz and 240Hz was 4. Integrating these features in the order of "spectral line spacing + amplitude ratio," the resulting spectral feature dataset is [60Hz, 120Hz, 1.6, 2.5, 4]. Given that the maximum spectral line spacing in historical data is 200Hz and the minimum is 50Hz, and the maximum amplitude ratio is 5 and the minimum is 1, the normalized feature vector set is [0.07, 0.47, 0.15, 0.38, 0.75], with a dimension of 5, which does not meet the preset dimension threshold of 8. Therefore, weak peaks near other frequencies such as 300Hz are included to supplement more adjacent spectral line features, expanding the vector dimension to 9, thus meeting the threshold requirement.

[0047] Next, the expanded normalized feature vector is input into the trained SVM model, which outputs the fault type label "1", corresponding to "rotor imbalance". The feature vector corresponding to this label is matched with the fault mode library. The typical template for rotor imbalance in the library is "spectral line spacing is an integer multiple of the fundamental frequency (such as 60Hz, 120Hz, 240Hz, etc.), and the amplitude ratio increases with frequency". The calculated cosine similarity is 0.93, which is the highest value obtained by matching all templates. Therefore, the fault classification result is determined to be rotor imbalance.

[0048] In step S16, the current second vibration signal data is acquired, and the similarity between the second vibration signal data and the pre-established normal pattern library is calculated. If the similarity is lower than a preset similarity threshold, the corresponding fault feature vector is output according to the fault classification result, including: For the second vibration signal data, high-frequency noise above the preset frequency cutoff point is filtered out by Fourier transform, and low-frequency effective signals are retained. The low-frequency effective signals are then standardized to obtain a real-time signal feature set. Calculate and normalize the distance between the real-time signal feature set and each pattern in the normal pattern library to obtain a normalized distance set. Subtract each distance value in the normalized distance set from the value 1 to form a similarity set. If the proportion of similarity values ​​below a preset similarity threshold in the similarity set exceeds a preset ratio threshold, then based on the fault classification result, the abnormal mode of the real-time signal feature set is determined, and the corresponding fault feature vector is extracted.

[0049] It should be noted that the second vibration dataset consists of real-time collected vibration signals from the generator set, used for dynamic monitoring of the unit's current operating status. The preset frequency cutoff point needs to be set in conjunction with the characteristic frequency range of the unit's key components. The effective vibration signals of diesel generator sets are mostly concentrated in the 0-1000Hz range, while high-frequency components above 1000Hz are mostly environmental interference (such as electromagnetic noise, airflow noise, etc.). Therefore, the frequency cutoff point can be set to 1000Hz, retaining only the low-frequency effective signals in the 0-1000Hz range. Using the z-score standardization method, the amplitude of the low-frequency effective signal is subtracted from the mean amplitude of historical normal signals, and then divided by the standard deviation of historical normal signal amplitudes to obtain the standardized low-frequency effective signal. This standardized signal is then combined according to the time series to obtain the real-time signal feature set.

[0050] In this embodiment, a normal mode library is constructed based on historical data of the unit under multiple operating conditions in a healthy state. First, vibration signals of the unit under different normal operating conditions (such as 20% load, 50% load, 80% load, etc.) are collected, with each condition continuously collected for at least 24 hours to ensure coverage of the complete operating cycle. Then, the original signals are sliced ​​according to a fixed time window (the same as the time window in step S11) to form an initial signal segment set. The frequency band entropy sequence of each data segment is calculated, and z-score normalization is used to eliminate the amplitude magnitude difference of the initial signal segment set. Subsequently, the core features of each segment are extracted, including the main frequency components (such as 50Hz fundamental frequency, 100Hz harmonics, etc.), the amplitude change trend of characteristic frequencies over time, and the statistical characteristics of signal energy distribution (such as variance, peak factor, etc.). These features are combined into feature vectors according to the time series. All feature vectors are clustered using the K-means clustering algorithm to obtain several cluster centers. These cluster centers are the patterns in the normal mode library, forming the normal mode library.

[0051] It should be noted that when calculating the distance between the real-time signal feature set and each pattern in the normal pattern library, feature vectors identical to those in the normal pattern library (e.g., root mean square, kurtosis, spectral peak value, etc.) need to be extracted from the real-time signal feature set. The Euclidean distance between the real-time feature vector and the feature vector in the pattern library is then calculated. The smaller the distance value, the more similar the two sequences are. By dividing by the maximum possible distance, the global distance is normalized to the 0-1 interval, resulting in a normalized distance set. Subtracting each normalized distance value from 1 yields the similarity set, ranging from 0 to 1, where 1 represents complete similarity and 0 represents complete dissimilarity.

[0052] It is worth noting that the preset similarity threshold is used to determine whether the real-time signal deviates from the normal pattern, and its value needs to be determined based on the similarity distribution of historical normal data. By statistically analyzing the similarity between a large number of real-time signals during normal operation and the normal pattern library, it was found that more than 95% of the similarities were greater than 0.8, so the preset similarity threshold can be set to 0.8. The preset proportion threshold is derived from the statistical analysis of historical normal operation data of diesel generator sets. Vibration signal data of the unit under typical normal operating conditions at different loads and speeds were selected, and the similarity between each operating condition signal and all patterns in the normal pattern library was calculated. The distribution of the proportion of similarities below the similarity threshold under normal conditions was statistically analyzed. It can be seen that in more than 95% of normal operating conditions, the proportion of similarities below the similarity threshold is less than 20%, so the preset proportion threshold can be set to 20%.

[0053] Finally, based on the fault classification results obtained in step S15, the abnormal modes related to the fault in the real-time signal feature set (such as the abnormal increase in the amplitude of a specific frequency corresponding to rotor imbalance) are determined, and the abnormal frequency, amplitude, duration and other feature parameters corresponding to the abnormal mode are extracted to form a fault feature vector.

[0054] For example, the second vibration signal data of a diesel generator set, after Fourier transform, filters out high-frequency noise above 1000Hz, retains low-frequency signals of 0-1000Hz, and obtains a real-time signal feature set through z-score normalization. The normal pattern library contains three typical patterns: low load mode (20% load), medium load mode (50% load), and high load mode (80% load). By calculating the global distance between the real-time feature set and each pattern, the maximum possible distance is 1, resulting in a normalized distance set of [0.3, 0.4, 0.5] and a similarity set of [0.7, 0.6, 0.5]. Since the preset similarity threshold is 0.8, and all similarities are below the threshold, combined with the fault classification result "rotor imbalance", it was determined that the amplitude of the 60Hz and 120Hz frequencies in the real-time signal was abnormally high, 30% higher than the normal mode. The amplitude, duration and other parameters of these frequencies were extracted to form a fault feature vector [60Hz, 0.35V, 120Hz, 0.22V, duration 10min]. This vector was output to guide further maintenance. It should be noted that the fault feature vector can be selected from one or more of the abnormal frequency, amplitude, duration and other features according to actual needs, and can also be further processed by normalization before being input into the random forest model.

[0055] In step S17, the fault feature vector is verified using the random forest algorithm based on the normal pattern library to obtain the final fault detection result.

[0056] It's important to note that random forests are supervised learning models, composed of multiple independent decision trees. Each tree is built based on a random subset of the training data and a random subset of features. The final output is determined by majority voting, which improves the accuracy and robustness of fault diagnosis and avoids potential misjudgments that can occur with a single model. The normal pattern library serves as a reference benchmark, providing the distribution range of features such as abnormal frequency, corresponding amplitude, and duration under normal operating conditions. Abnormal frequency points to potentially faulty components (e.g., bearing failures often correspond to specific high frequencies), amplitude reflects the severity of the fault (the larger the amplitude, the more severe the fault may be), and duration reflects the stability of the fault. These features can be used to assist the decision tree in determining whether a feature is abnormal.

[0057] In this embodiment, the training data for the random forest consists of feature vectors corresponding to historical fault cases and normal operation data. Each vector is labeled with either "fault" or "normal" and indicates the specific fault type. First, multiple sample sets with replacement are generated from the training data through random sampling. For each sample set, a fixed number of features are randomly selected (e.g., 5 out of 10 features) to construct a decision tree. Each node of the decision tree is split using the feature with the highest information gain. Information gain measures the degree to which a feature reduces the uncertainty of classifying the sample set; it is calculated as the information entropy before splitting minus the information entropy after splitting. The lower the information entropy, the clearer the sample classification. For example, "whether the 60Hz amplitude exceeds 0.3V" can be used as the splitting condition. This process is repeated to generate 100 decision trees, forming the random forest.

[0058] It's important to note that as the number of decision trees increases to a certain point, the rate of accuracy improvement slows significantly. Further increasing the number of trees at this point only increases computational costs (including memory usage and training time) without any noticeable benefit. In the diesel generator set fault diagnosis scenario, 100 decision trees is a reasonable value obtained after testing the model's validation accuracy with different numbers of trees. This value mitigates the overfitting risk of a single tree through independent judgments from multiple trees while controlling computational time during real-time detection. After training, each tree independently judges the input fault feature vector and outputs a classification result. The final validation result is the judgment of the majority of trees.

[0059] For example, a fault feature vector is [60Hz, 0.35V, 120Hz, 0.22V, lasting 10min], which is input into a random forest composed of 100 decision trees. 85 trees identify it as "rotor imbalance," 10 trees as "normal fluctuation," and 5 trees as "bearing wear." The probability of "rotor imbalance" is the highest, at 85%, and this result is consistent with the typical characteristics of rotor imbalance in the normal pattern library (increased amplitude and long duration of 60Hz and its harmonics). Therefore, the final fault detection result is confirmed as "rotor imbalance," which is consistent with the actual disassembly and inspection results.

[0060] In summary, this invention discloses a fault detection method for diesel generator sets based on deep learning, comprising: acquiring first vibration signal data of the diesel generator set during operation; performing time-frequency decomposition on the first vibration signal data to obtain dynamic change features; performing adaptive filtering on the dynamic change features to obtain a denoised time-frequency feature sequence; extracting a peak energy distribution dataset from the time-frequency feature sequence; calculating the entropy value of each frequency band based on the peak energy distribution dataset to obtain a frequency band entropy value sequence; classifying the frequency band entropy value sequence using a k-means clustering algorithm to determine a random fluctuation benchmark pattern; and based on the random fluctuation benchmark pattern... Abnormal frequency components are separated from the peak energy distribution dataset. Based on the abnormal frequency components, the spectral line spacing and amplitude ratio are calculated to obtain a spectral feature dataset. The spectral feature dataset is classified based on a preset fault mode library to determine the fault classification result. The current second vibration signal data is acquired, and the similarity between the second vibration signal data and the pre-established normal mode library is calculated. If the similarity is lower than a preset similarity threshold, the corresponding fault feature vector is output based on the fault classification result. The fault feature vector is verified using a random forest algorithm based on the normal mode library to obtain the final fault detection result. This invention integrates time-frequency analysis technology and pattern matching methods to construct a multi-level fault detection system, achieving efficient capture of abnormal features and accurate identification of fault types in the vibration signals of diesel generator sets.

[0061] Reference Figure 2 The second embodiment of the present invention provides a diesel generator set fault detection system based on deep learning, comprising: The signal time-frequency decomposition module is used to acquire the first vibration signal data during the operation of the diesel generator set, and to perform time-frequency decomposition on the first vibration signal data to obtain dynamic change characteristics. The dynamic denoising module is used to perform adaptive filtering on the dynamically changing features to obtain a denoised time-frequency feature sequence. The frequency band entropy calculation module is used to extract the peak energy distribution dataset from the time-frequency feature sequence, and calculate the entropy value of each frequency band based on the peak energy distribution dataset to obtain the frequency band entropy value sequence. An abnormal frequency separation module is used to classify the frequency band entropy value sequence using a k-means clustering algorithm, determine a random fluctuation baseline pattern, and separate abnormal frequency components from the peak energy distribution dataset based on the random fluctuation baseline pattern. The fault classification determination module is used to calculate the spectral line spacing and amplitude ratio based on the abnormal frequency components to obtain a spectral line feature dataset, classify the spectral line feature dataset based on a preset fault mode library, and determine the fault classification result. The fault feature output module is used to acquire the current second vibration signal data, calculate the similarity between the second vibration signal data and the pre-established normal pattern library, and if the similarity is lower than the preset similarity threshold, output the corresponding fault feature vector according to the fault classification result. The random forest verification module is used to verify the fault feature vectors based on the normal pattern library using the random forest algorithm to obtain the final fault detection result.

[0062] It should be noted that the deep learning-based diesel generator set fault detection system provided in this embodiment of the invention is used to execute all the process steps of the deep learning-based diesel generator set fault detection method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0063] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an adaptive filtering program. When the processor executes the computer program, it implements the steps described in the various embodiments of the deep learning-based diesel generator set fault detection method, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as the signal time-frequency decomposition module.

[0064] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0065] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0066] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0067] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0068] If the modules / units integrated into the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0069] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0070] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A deep learning-based diesel generator set fault detection method, characterized in that, include: The first vibration signal data of the diesel generator set during operation is acquired, and the first vibration signal data is decomposed into time and frequency to obtain dynamic change characteristics. Adaptive filtering is performed on the aforementioned dynamic change characteristics to obtain a denoised time-frequency feature sequence; Extracting peak energy distribution datasets from the time-frequency feature sequence, and calculating the entropy values ​​of each frequency band based on the peak energy distribution datasets to obtain a frequency band entropy value sequence, includes: converting the time-domain signal into a frequency-domain signal using a fast Fourier transform based on the time-frequency feature sequence, extracting the peak energy distribution features of the signal at different frequencies to obtain a peak energy distribution dataset; normalizing the energy values ​​of each frequency band to a probability distribution based on the peak energy distribution dataset, calculating the Shannon entropy of each frequency band based on the probability distribution, and arranging the Shannon entropy in frequency band order to obtain a frequency band entropy value sequence; The frequency band entropy sequence is classified using the k-means clustering algorithm to determine a random fluctuation baseline pattern. Based on the random fluctuation baseline pattern, abnormal frequency components are separated from the peak energy distribution dataset. The frequency band entropy sequence is divided into k classes, and the class with the highest sample size and the smallest entropy fluctuation is determined as the random fluctuation baseline pattern. Based on the abnormal frequency components, the spectral line spacing and amplitude ratio are calculated to obtain a spectral line feature dataset. The spectral line feature dataset is then classified based on a preset fault mode library to determine the fault classification result. Acquire the current second vibration signal data, calculate the similarity between the second vibration signal data and the pre-established normal pattern library, and if the similarity is lower than the preset similarity threshold, output the corresponding fault feature vector according to the fault classification result. Based on the normal pattern library, the fault feature vector is verified using the random forest algorithm to obtain the final fault detection result.

2. The deep learning-based diesel generator set fault detection method according to claim 1, characterized in that, The step of performing time-frequency decomposition on the first vibration signal data to obtain dynamic change characteristics includes: The first vibration signal data is subjected to analog-to-digital conversion to obtain a digital vibration signal; Wavelet transform is used to perform time-frequency decomposition on the digital vibration signal to generate a time-spectrum diagram containing time and frequency information. The time-spectrum diagram is divided into preset time windows, and the frequency components of each time window are calculated by Fourier transform to obtain the dynamic change characteristics.

3. The deep learning-based diesel genset fault detection method of claim 1, wherein, The adaptive filtering performed on the dynamically changing features to obtain the denoised time-frequency feature sequence includes: To address the aforementioned dynamic change characteristics, a wavelet denoising method is used for preliminary denoising to obtain the first feature dataset after denoising. For the first feature dataset, discrete cosine transform is used to compress the data to obtain a denoised time-frequency feature sequence.

4. The deep learning-based diesel genset fault detection method of claim 1, wherein, The step of classifying the frequency band entropy sequence using the k-means clustering algorithm to determine a random fluctuation baseline pattern, and then separating abnormal frequency components from the peak energy distribution dataset based on the random fluctuation baseline pattern, includes: When the fluctuation amplitude of the frequency band entropy value sequence is higher than the preset fluctuation threshold, the frequency band entropy value sequence is classified by the k-means clustering algorithm to obtain the random fluctuation benchmark pattern; Based on the random fluctuation baseline pattern, a normal energy threshold is determined, and frequency bands whose peak energy exceeds the preset energy threshold are selected from the peak energy distribution dataset and identified as abnormal frequency components.

5. The deep learning-based diesel genset fault detection method of claim 1, wherein, The step involves calculating the spectral line spacing and amplitude ratio based on the abnormal frequency components to obtain a spectral line feature dataset. This dataset is then classified based on a pre-defined fault mode library to determine the fault classification result, including: An abnormal feature sequence is obtained from the abnormal frequency components, and the spectral line spacing and amplitude ratio of the abnormal feature sequence are calculated by fast Fourier transform to obtain the spectral line feature dataset. The spectral line feature dataset is normalized to obtain a normalized feature vector set; When the dimension of the normalized feature vector set meets the preset dimension threshold, the normalized feature vector set is divided by a support vector machine to obtain a fault type label set. The fault type label set is matched with the fault pattern library to obtain the fault classification result.

6. The deep learning-based diesel genset fault detection method of claim 1, wherein, The similarity between the second vibration signal data and the pre-established normal pattern library is calculated. If the similarity is lower than a preset similarity threshold, the corresponding fault feature vector is output based on the fault classification result, including: For the second vibration signal data, high-frequency noise above the preset frequency cutoff point is filtered out by Fourier transform, and low-frequency effective signals are retained. The low-frequency effective signals are then standardized to obtain a real-time signal feature set. Calculate and normalize the distance between the real-time signal feature set and each pattern in the normal pattern library to obtain a normalized distance set. Subtract each distance value in the normalized distance set from the value 1 to form a similarity set. If the proportion of similarity values ​​below a preset similarity threshold in the similarity set exceeds a preset ratio threshold, then based on the fault classification result, the abnormal mode of the real-time signal feature set is determined, and the corresponding fault feature vector is extracted.

7. A deep learning-based diesel generator set fault detection system, characterized by, include: The signal time-frequency decomposition module is used to acquire the first vibration signal data during the operation of the diesel generator set, and to perform time-frequency decomposition on the first vibration signal data to obtain dynamic change characteristics. The dynamic denoising module is used to perform adaptive filtering on the dynamically changing features to obtain a denoised time-frequency feature sequence. The frequency band entropy calculation module is used to extract peak energy distribution datasets from the time-frequency feature sequence, and calculate the entropy values ​​of each frequency band based on the peak energy distribution datasets to obtain a frequency band entropy value sequence. This includes: converting the time-domain signal into a frequency-domain signal using a fast Fourier transform based on the time-frequency feature sequence; extracting the peak energy distribution features of the signal at different frequencies to obtain a peak energy distribution dataset; normalizing the energy values ​​of each frequency band to a probability distribution based on the peak energy distribution dataset; calculating the Shannon entropy of each frequency band based on the probability distribution; and arranging the Shannon entropy values ​​in frequency band order to obtain a frequency band entropy value sequence. An abnormal frequency separation module is used to classify the frequency band entropy value sequence using a k-means clustering algorithm, determine a random fluctuation baseline pattern, and separate abnormal frequency components from the peak energy distribution dataset based on the random fluctuation baseline pattern; wherein, the frequency band entropy value sequence is divided into k classes, and the class with the highest sample number and the smallest entropy value fluctuation is determined as the random fluctuation baseline pattern; The fault classification determination module is used to calculate the spectral line spacing and amplitude ratio based on the abnormal frequency components to obtain a spectral line feature dataset, classify the spectral line feature dataset based on a preset fault mode library, and determine the fault classification result. The fault feature output module is used to acquire the current second vibration signal data, calculate the similarity between the second vibration signal data and the pre-established normal pattern library, and if the similarity is lower than the preset similarity threshold, output the corresponding fault feature vector according to the fault classification result. The random forest verification module is used to verify the fault feature vectors using the random forest algorithm based on the normal pattern library, and obtain the final fault detection result.

Images