Carrier roller abnormal sound recognition method based on small sample number feature fusion

Abstract

The invention relates to the technical field of signal processing, and discloses a carrier roller abnormal sound recognition method based on small sample number feature fusion, which is used for improving the accuracy of carrier roller abnormal sound recognition. The carrier roller abnormal sound recognition method based on small sample number feature fusion comprises the following steps: collecting carrier roller operation audio and enhancing preprocessing; extracting time domain, frequency domain and Mel-frequency cepstral coefficient features to form a multi-domain feature vector; carrying out clustering optimization on the Mel-frequency cepstral coefficient, fusing the Mel-frequency cepstral coefficient with other features, and constructing an optimization feature adaptive to the small sample; inputting the optimized features into a pre-trained RBF-SVM classification model to identify a state; and judging an alarm condition in combination with sliding window counting and an alarm cooling mechanism, and outputting alarm information when the alarm condition is met.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a method for identifying abnormal noises from idler rollers based on feature fusion with a small number of samples. Background Technology

[0002] As a key load-bearing component of belt conveyors, the operating status of idler rollers directly affects the safety and efficiency of the entire conveying system. Initially, idler roller bearing wear and jamming often exhibit specific unusual noises. Therefore, online monitoring and intelligent diagnosis of idler roller operating sounds using audio analysis technology is an important means of achieving predictive maintenance and avoiding unplanned downtime.

[0003] Currently, machine learning-based audio recognition methods have been widely applied in the field of industrial equipment fault diagnosis. Existing technical solutions typically follow this process: acquiring audio from the equipment during operation, extracting acoustic features such as Mel-frequency cepstral coefficients and spectral centroids, constructing a high-dimensional feature set, and using deep learning models (such as convolutional neural networks) for pattern classification.

[0004] Due to the large number of idler roller devices and the complex installation environment, it is difficult to obtain sufficient and accurately labeled fault audio samples, resulting in an extreme scarcity of "abnormal sound" data for model training, i.e., the "small sample problem." Existing mainstream methods exhibit significant shortcomings under this constraint: on the one hand, deep learning models heavily rely on large datasets and are prone to overfitting on small samples, exhibiting poor generalization ability; on the other hand, directly using traditional feature extraction and fusion methods without optimizing for the sparse distribution and strong noise interference of small sample data results in insufficient discriminative power and low robustness in the constructed feature representations, leading to a sharp decline in model recognition accuracy and reliability. Summary of the Invention

[0005] This invention provides a method for identifying abnormal noises on idlers based on feature fusion with a small number of samples, in order to solve the problem in the prior art that the scarcity of fault audio samples leads to overfitting of traditional deep learning models and insufficient feature discrimination, resulting in low accuracy and reliability of fault identification.

[0006] The first aspect of this invention provides a method for identifying abnormal noise from idler rollers based on feature fusion with a small sample size, comprising: acquiring the original audio signal during idler roller operation; performing enhancement preprocessing on the original audio signal to obtain an enhanced audio signal; performing frame segmentation on the enhanced audio signal and extracting time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features from each frame to form a multi-domain feature vector; performing clustering optimization on the Mel-frequency cepstral coefficient features in the multi-domain feature vector, and fusing the optimization result with the time-domain features and the frequency-domain features to obtain an optimized feature vector adapted to a small sample scenario; inputting the optimized feature vector into a pre-trained RBF kernel support vector machine classification model to obtain a classification result of the idler roller state; and, based on the classification result, determining whether an alarm is triggered by combining a sliding window counting mechanism and an alarm cooling mechanism, and outputting alarm information when the alarm conditions are met.

[0007] In one feasible implementation, before inputting the optimized feature vector into the pre-trained RBF kernel support vector machine classification model, the method further includes: collecting and constructing a small sample training set, which includes three types of idler roller operation audio samples: normal, abnormal noise, and shutdown; performing enhancement processing, frame segmentation processing, multi-feature extraction, and Mel-frequency cepstral coefficient feature analysis on the audio samples in the small sample training set to obtain a training feature vector set; using the training feature vector set to train the RBF kernel support vector machine classification model, and optimizing the hyperparameters through cross-validation to obtain the RBF kernel support vector machine classification model.

[0008] In one feasible implementation, the step of performing frame segmentation processing on the enhanced audio signal and extracting time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features from each frame to form a multi-domain feature vector includes: performing frame segmentation processing on the enhanced audio signal to obtain several audio frames; extracting time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features from each audio frame; and concatenating the time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features of each audio frame to form a multi-domain feature vector. In one feasible implementation, the step of clustering and optimizing the Mel-Cepstral Coefficient features in the multi-domain feature vector, and fusing the optimization result with the time-domain features and the frequency-domain features to obtain an optimized feature vector adapted to small sample scenarios, includes: performing clustering operations on the Mel-Cepstral Coefficient features in the multi-domain feature vector to form at least two feature clusters; determining the main cluster with the largest sample size from the at least two feature clusters, and extracting representative feature values ​​of the main cluster to optimize the feature representation under small sample data, thereby obtaining optimized Mel-Cepstral Coefficient features; and concatenating the optimized Mel-Cepstral Coefficient features with the time-domain features and frequency-domain features in the multi-domain feature vector to form the optimized feature vector adapted to small sample scenarios.

[0009] In one feasible implementation, the step of performing clustering operations on the Mel-Cepstral Coefficient features in the multi-domain feature vector to form at least two feature clusters includes: performing clustering analysis on the Mel-Cepstral Coefficient features using the K-means clustering algorithm; determining the optimal number of clusters K based on the silhouette coefficient or cross-validation results; and performing clustering based on the number of clusters K to divide the Mel-Cepstral Coefficient features into K feature clusters, where K is an integer greater than or equal to 2.

[0010] In one feasible implementation, the step of extracting representative feature values ​​of the main cluster to optimize the feature representation under small sample data and obtain optimized Mel-Cepstral Coefficient features includes: calculating the median of all samples in the main cluster in each Mel-Cepstral Coefficient dimension; using the calculated median sequence as the representative feature value of the main cluster in the corresponding dimension; and combining the representative feature values ​​of each dimension to form the optimized Mel-Cepstral Coefficient features.

[0011] In one feasible implementation, the step of determining whether to trigger an alarm based on the classification result, combined with a sliding window counting mechanism and an alarm cooling mechanism, and outputting alarm information when the alarm conditions are met, includes: based on a preset sliding window containing multiple consecutive classification results, and counting the percentage of frames in the sliding window whose classification result is "abnormal sound"; if the percentage exceeds a preset alarm threshold, an alarm is triggered and a cooling timer is started, and new alarm triggers are blocked during the cooling period; when an alarm is triggered, at least one of the following operations is executed simultaneously: local sound and light alarm, abnormal audio data storage, and remote alarm information push.

[0012] A second aspect of the present invention provides a roller noise recognition device based on small sample size feature fusion, comprising: an acquisition module for acquiring the original audio signal during roller operation, performing enhancement preprocessing on the original audio signal to obtain an enhanced audio signal; an extraction module for performing frame-by-frame processing on the enhanced audio signal, and extracting time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features from each frame to form a multi-domain feature vector; an optimization module for performing cluster optimization on the Mel-frequency cepstral coefficient features in the multi-domain feature vector, and fusing the optimization result with the time-domain features and the frequency-domain features to obtain an optimized feature vector adapted to small sample scenarios; a classification module for inputting the optimized feature vector into a pre-trained RBF kernel support vector machine classification model to obtain a classification result of the roller state; and an alarm module for determining whether an alarm is triggered based on the classification result, combined with a sliding window counting mechanism and an alarm cooling mechanism, and outputting alarm information when the alarm conditions are met.

[0013] In one feasible implementation, the roller noise recognition device based on small sample feature fusion further includes: a training module for collecting and constructing a small sample training set, which includes three types of roller operation audio samples: normal, abnormal, and stopped; performing enhancement processing, frame segmentation processing, multi-feature extraction, and Mel-frequency cepstral coefficient feature analysis on the audio samples in the small sample training set to obtain a training feature vector set; using the training feature vector set to train an RBF kernel support vector machine classification model, and optimizing the hyperparameters through cross-validation to obtain the RBF kernel support vector machine classification model.

[0014] In one feasible implementation, the processing module is specifically used to: perform frame-segmentation processing on the enhanced audio signal to obtain several audio frames; extract time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features from each audio frame; and concatenate the time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features of each audio frame to form a multi-domain feature vector.

[0015] In one feasible implementation, the optimization module includes: a clustering unit, used to perform clustering operations on the Mel-Cepstral Coefficient features in the multi-domain feature vector to form at least two feature clusters; an optimization unit, used to determine the main cluster with the largest sample size from the at least two feature clusters, and extract representative feature values ​​of the main cluster to optimize the feature representation under small sample data, thereby obtaining optimized Mel-Cepstral Coefficient features; and a splicing unit, used to splice the optimized Mel-Cepstral Coefficient features with the time-domain features and frequency-domain features in the multi-domain feature vector to form the optimized feature vector adapted to the small sample scenario. In one feasible implementation, the clustering unit is specifically used to: perform cluster analysis on the Mel-means cepstral coefficient features using the K-means clustering algorithm; determine the optimal number of clusters K based on the silhouette coefficient or cross-validation results; and perform clustering based on the number of clusters K to divide the Mel-means cepstral coefficient features into K feature clusters, where K is an integer greater than or equal to 2.

[0016] In one feasible implementation, the optimization unit is specifically used to: calculate the median of all samples in the main cluster in each Mel-Cepstral Coefficient dimension; use the calculated median sequence as the representative feature value of the main cluster in the corresponding dimension; and combine the representative feature values ​​of each dimension to form the optimized Mel-Cepstral Coefficient feature.

[0017] In one feasible implementation, the alarm module is specifically used to: based on a preset sliding window containing multiple consecutive classification results, and to count the percentage of frames in the sliding window whose classification result is "abnormal sound"; if the percentage exceeds a preset alarm threshold, an alarm is triggered and a cooling timer is started, and new alarm triggers are blocked during the cooling period; when an alarm is triggered, at least one of the following operations is executed simultaneously: local sound and light alarm, abnormal audio data storage, and remote push of alarm information.

[0018] A third aspect of the present invention provides an electronic device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the electronic device to execute the above-described method for identifying abnormal noise of idler rollers based on feature fusion of small sample numbers.

[0019] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described method for identifying abnormal noises on idlers based on feature fusion with a small number of samples.

[0020] The technical solution provided by this invention involves acquiring the original audio signal of the idler roller during operation, performing enhancement preprocessing on the original audio signal to obtain an enhanced audio signal, performing frame segmentation on the enhanced audio signal, and extracting time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features from each frame to form a multi-domain feature vector, performing clustering optimization on the Mel-frequency cepstral coefficient features in the multi-domain feature vector, and fusing the optimization results with the time-domain features and the frequency-domain features to obtain an optimized feature vector adapted to small sample scenarios, inputting the optimized feature vector into a pre-trained RBF kernel support vector machine classification model to obtain the classification result of the idler roller state, and determining whether an alarm is triggered based on the classification result, combined with a sliding window counting mechanism and an alarm cooling mechanism, and outputting alarm information when the alarm conditions are met. In this embodiment of the invention, the original audio signal-to-noise ratio is enhanced through preprocessing. Multi-domain feature extraction in the time domain, frequency domain, and MFCC is performed, and clustering optimization of MFCC features is carried out to address the sparsity of small sample data. This effectively fuses feature representations with higher discriminative power and stronger noise resistance. Combined with a pre-trained RBF kernel support vector machine model, the fault identification accuracy and generalization ability can be significantly improved in small sample scenarios where idler fault audio data is scarce. At the same time, the sliding window counting and alarm cooling mechanism effectively reduce false alarms, realizing reliable online monitoring and intelligent early warning of idler operating status. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of an embodiment of the roller noise recognition method based on feature fusion with a small number of samples in this invention. Figure 2This is a schematic diagram of another embodiment of the roller noise recognition method based on feature fusion with a small number of samples in this invention; Figure 3 This is a schematic diagram of an embodiment of the roller noise recognition device based on feature fusion with a small number of samples in this invention; Figure 4 This is a schematic diagram of another embodiment of the roller noise recognition device based on feature fusion with a small number of samples in this invention; Figure 5 This is a schematic diagram of one embodiment of the electronic device in this invention. Detailed Implementation

[0022] This invention provides a method for identifying abnormal noises from idlers based on feature fusion with a small sample size. By enhancing the signal, optimizing features, and fusing a classification decision mechanism, the method improves the accuracy, generalization ability, and alarm reliability of idler fault identification in small sample scenarios.

[0023] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] It is understood that the executing entity of this invention can be a roller noise recognition device based on feature fusion with a small sample size, or it can be a terminal or a server; the specific implementation is not limited here. This embodiment of the invention will be described using a server as an example.

[0025] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the roller noise recognition method based on small sample size feature fusion in this invention includes: 101. Obtain the original audio signal during the operation of the idler roller, perform enhancement preprocessing on the original audio signal to obtain the enhanced audio signal; A microphone is fixedly mounted on the inspection robot to collect the raw audio signal during the operation of the idler roller in real time. The signal is then transmitted to an industrial server for real-time caching via a shielded network cable. The raw signal is then enhanced and preprocessed using the maximum correlation kurtosis deconvolution (MCKD) algorithm. The filter coefficients are iteratively optimized to maximize the correlation kurtosis of the periodic impact components, thereby improving the signal-to-noise ratio of the abnormal sound signal. Band-stop filtering and spectral attenuation are then used to further suppress environmental noise, resulting in the enhanced audio signal.

[0026] 102. Perform frame segmentation on the enhanced audio signal and extract time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features from each frame to form a multi-domain feature vector; When performing frame segmentation on the enhanced audio signal, the signal is divided into continuous segments according to a fixed duration window, with a certain overlap ratio maintained between adjacent segments to ensure information continuity. Then, three types of features are extracted from each frame: in the time domain, statistics including root mean square, kurtosis, peak factor, and zero-crossing rate are calculated to describe the signal energy and impulse characteristics. Specifically: For a signal x(n) of length N, its root mean square (RMS) is calculated as follows:

[0027] For a signal x(n), its kurtosis (minus 3 with a relative normal distribution as the zero baseline) is calculated using the following formula:

[0028] in, The mean of the signal. The standard deviation is denoted as .

[0029] For a signal x(n), the formula for calculating its peak factor is:

[0030] in, This represents the peak value of the signal.

[0031] For a signal x(n), the zero-crossing rate is calculated using the following formula:

[0032] In the frequency domain, the frequency distribution characteristics are characterized by calculating the high-to-low frequency energy ratio and the spectral centroid after obtaining the spectrum through Fourier transform. Specifically: Set frequency point The corresponding energy is Then the centroid SC of the spectrum is:

[0033] In the time-frequency domain, Mel-based cepstral coefficients are extracted using a Mel filter bank to capture the spectral texture changes of the signal at the Mel scale. Specifically: Frame the signal and calculate the power spectrum , where m is the frame index and k is the frequency point.

[0034] Through 26 Mel filter banks Calculate the Mel spectrum:

[0035] Taking the logarithm and performing a discrete cosine transform, retaining the first 13 coefficients yields the Mel-frequency cepstral coefficient (MFCC) characteristics:

[0036] Where j = 1, 2, ..., 13.

[0037] Finally, these three types of features are spliced ​​and fused frame by frame to form a unified multi-domain feature vector.

[0038] 103. Cluster optimization is performed on the Mel-Cepstral coefficient features in the multi-domain feature vector, and the optimization results are fused with the time-domain features and frequency-domain features to obtain an optimized feature vector suitable for small sample scenarios; The Mel-Cepstral Coefficient (MCC) features in the multi-domain feature vector are optimized using a clustering method. Specifically, the K-means algorithm is used to group the MCCs of all samples, determine the main clusters based on the sample distribution, and select the median of each dimension in the cluster as the representative feature value to suppress outlier interference and extract stable core features. Subsequently, the optimized MCC features are merged with the time-domain features and frequency-domain features to form an optimized feature vector with an overall dimension. This vector not only retains the complementarity of multi-domain information but also improves the robustness of expression and classification adaptability on small sample datasets through feature denoising and agglomeration.

[0039] 104. Input the optimized feature vector into the pre-trained RBF kernel support vector machine classification model to obtain the classification result of the idler roller state; The optimized feature vectors are input into a pre-trained radial basis function (RBF) kernel support vector machine (SVM) classification model. This model maps low-dimensional features to a high-dimensional separable space through kernel functions and constructs a classification hyperplane based on the principle of minimizing structural risk. During the training phase, cross-validation and grid search strategies are used to optimize the model's hyperparameters. At the same time, a sample weight adjustment mechanism is introduced to address the class imbalance problem in small sample scenarios. The model performs real-time inference on the input optimized feature vectors and outputs three discrimination results corresponding to the normal, abnormal noise, and stopped states of the idler rollers, thereby achieving accurate classification and recognition of the idler roller operating status in small sample scenarios.

[0040] 105. Based on the classification results, combined with the sliding window counting mechanism and alarm cooling mechanism, determine whether an alarm is triggered, and output alarm information when the alarm conditions are met.

[0041] Based on the real-time acquired roller status classification results, a sliding window counting mechanism is used to continuously count the occurrence ratio of abnormal noise categories in the most recent frames. When the ratio exceeds a preset threshold, it is initially determined to be an abnormal event. At the same time, an alarm cooling mechanism is introduced, which starts a silent period after each alarm is triggered. During this period, even if the alarm conditions are met again, it will not be triggered again, and only the duration of the abnormal state will be updated. Finally, when the abnormal noise ratio exceeds the standard and is not in the cooling period, the system officially triggers the alarm and performs the linkage operation of local audible and visual warning, abnormal noise period data storage, and alarm information push to the remote monitoring platform.

[0042] In this embodiment of the invention, the original audio signal-to-noise ratio is enhanced through preprocessing. Multi-domain feature extraction in the time domain, frequency domain, and MFCC is performed, and clustering optimization of MFCC features is carried out to address the sparsity of small sample data. This effectively fuses feature representations with higher discriminative power and stronger noise resistance. Combined with a pre-trained RBF kernel support vector machine model, the fault identification accuracy and generalization ability can be significantly improved in small sample scenarios where idler fault audio data is scarce. At the same time, the sliding window counting and alarm cooling mechanism effectively reduce false alarms, realizing reliable online monitoring and intelligent early warning of idler operating status.

[0043] Please see Figure 2 Another embodiment of the roller noise recognition method based on small sample size feature fusion in this invention includes: 201. Obtain the original audio signal during the operation of the idler roller, perform enhancement preprocessing on the original audio signal to obtain the enhanced audio signal; 202. Perform frame segmentation on the enhanced audio signal and extract time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features from each frame to form a multi-domain feature vector; The enhanced audio signal is segmented into frames to obtain several audio frames. Time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features are extracted from each audio frame. The time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features of each audio frame are concatenated to form a multi-domain feature vector.

[0044] The window size and step size are determined based on the signal sampling rate and preset duration. Taking a sampling rate of 44.1kHz as an example, the window duration is set to 3 seconds, corresponding to a window size of 132,300 sampling points; the step size is set to half the sampling rate, i.e., 22,050 sampling points, to achieve a 50% overlap rate between frames, ensuring that each audio frame can capture complete roller rotation cycle information and maintain temporal continuity. During framing, a Hamming window function is used to weight each window to reduce spectral leakage, ultimately dividing the continuously enhanced audio signal into several overlapping audio frames. Next, multi-dimensional feature extraction is performed on each audio frame: in the time domain, four key statistics are calculated—the root mean square (RMS) of the signal energy, used to characterize the overall energy level; the kurtosis, reflecting the signal pulse intensity, obtained by calculating the ratio of the fourth-order central moment of the signal to the fourth power of the standard deviation; the peak factor, describing the relationship between the peak value and the effective value, i.e., the ratio of the absolute peak value of the signal to the RMS; and the zero-crossing rate, characterizing the frequency of the signal crossing zeros, obtained by normalizing the number of times the signal crosses zero. In the frequency domain, a Fast Fourier Transform (FFT) is first performed on each frame of the signal to obtain the spectrum. Two core indicators are calculated: the high-frequency energy ratio (HFR) and the low-frequency energy ratio (FLFR). By dividing the signal at 1500Hz, the energy values ​​of the high-frequency and low-frequency bands are calculated separately, and the ratio is obtained to highlight the characteristics of the high-frequency components of the abnormal sound. The spectral centroid is also calculated, and its position is determined by the frequency position of the weighted average of the spectral amplitude to reflect the energy concentration area. In the time-frequency domain, the Mel frequency cepstral coefficient (MFCC) analysis method is used: First, the power spectrum of each frame of the signal is calculated and filtered by a Mel filter bank composed of 26 triangular bandpass filters to obtain the Mel spectrum. After taking the natural logarithm of this spectrum, a Discrete Cosine Transform (DCT) is performed, retaining the first 13 coefficients to form the MFCC feature vector. These coefficients can effectively characterize the spectral envelope shape and time-frequency texture features of the sound at the Mel frequency scale. Finally, the time-domain 4-dimensional feature vector, the frequency-domain 2-dimensional feature vector extracted from each audio frame, and the 13-dimensional feature vector of the MFCC are concatenated in a predetermined order to form a unified 19-dimensional multi-domain feature vector. This vector comprehensively integrates the complementary information of the signal in the time, frequency, and time-frequency domains.

[0045] 203. Construct an RBF kernel support vector machine classification model; A small training set was collected and constructed, which included audio samples of three types of idler operation: normal, abnormal, and stopped. Enhancement processing, frame segmentation processing, multi-feature extraction, and Mel-frequency cepstral coefficient feature analysis were performed on the audio samples in the small training set to obtain a training feature vector set. The RBF kernel support vector machine classification model was trained using the training feature vector set, and the hyperparameters were optimized through cross-validation to obtain the RBF kernel support vector machine classification model.

[0046] When constructing a small sample training set, it is necessary to systematically record the raw audio data of the idler rollers under different operating conditions in the industrial field through microphones deployed on inspection robots or fixed monitoring points. This ensures that the three core operating conditions are covered: normal operation, typical abnormal noise faults, and equipment shutdown. The number of samples collected for each type should be balanced according to the ease of obtaining samples in the actual scenario, but it should be ensured that each type has at least dozens of representative independent audio segments, and the duration of each audio segment should be sufficient to contain the complete characteristics of the equipment operation cycle. Subsequently, a systematic preprocessing and feature engineering process was performed on these original audio samples: First, each sample was enhanced by using the maximum correlation kurtosis deconvolution algorithm to highlight periodic impact components and suppress environmental noise; then, the enhanced audio signal was divided into frames, weighted using a fixed-duration Hamming window, and a certain overlap ratio was set to maintain continuity; then, multi-dimensional features were extracted in parallel for each frame of audio signal, including root mean square, kurtosis, peak factor and zero-crossing rate in the time domain, high-low frequency energy ratio and spectral centroid in the frequency domain, and the Mel cepstral coefficient sequence obtained by Mel filter bank and discrete cosine transform; then, the features of all frames of each sample were aggregated, for example, the statistics of each frame feature were taken as sample-level feature representation to form a structured training feature vector set. When training the radial basis function support vector machine model using this feature vector set, a grid search strategy is used to systematically explore the preset hyperparameter space. The model performance under different hyperparameter combinations is evaluated by combining cross-validation method. Finally, the hyperparameter combination with the best classification accuracy and generalization ability on the validation set is selected to complete the training and optimization of the model, thereby obtaining a highly reliable classification model that is suitable for small sample scenarios and can stably distinguish between normal roller status, abnormal noise and shutdown status.

[0047] 204. Perform clustering operations on the Mel-Cepstral coefficients in the multi-domain feature vectors to form at least two feature clusters; The K-means clustering algorithm is used to perform cluster analysis on the Mel-Cepstral Coefficient features. The optimal number of clusters K is determined based on the silhouette coefficient or cross-validation results. Clustering is performed based on the number of clusters K to divide the Mel-Cepstral Coefficient features into K feature clusters, where K is an integer greater than or equal to 2.

[0048] When using the K-means clustering algorithm to perform cluster analysis on the Mel-Cepstral Coefficient (MCC) features extracted from the training samples, the MCC features of all samples are first organized into a feature matrix, where each row represents a multi-dimensional MFCC feature vector of a sample. To determine the optimal number of clusters K, the clustering quality of multiple candidate K values ​​needs to be systematically evaluated: the silhouette coefficient corresponding to each candidate K value is calculated, which comprehensively measures the compactness of samples of the same class and the separation of samples of different classes; the closer the value is to positive one, the better the clustering effect. Simultaneously, cross-validation can be used to divide the feature data into multiple subsets, perform multiple clustering based on different K values, and evaluate the stability of intra-class and inter-class distances. Finally, the K value that maximizes the silhouette coefficient and has consistent cross-validation results is selected as the optimal number of clusters. After determining the K value, K cluster centers are initialized. The sample allocation and center point update steps are iteratively executed until the change in cluster centers is below a threshold or the maximum number of iterations is reached, thereby dividing all MCC features into K independent feature clusters, forming structured feature groupings.

[0049] 205. Determine the main cluster with the largest sample size from at least two feature clusters, and extract the representative feature values ​​of the main cluster to optimize the feature representation under small sample data, and obtain the optimized Mel-Cepstral Coefficient Features; Calculate the median of all samples in the main cluster in each Mel-Cepstral Coefficient dimension; use the calculated median sequence as the representative feature value of the main cluster in the corresponding dimension; combine the representative feature values ​​of each dimension to form the optimized Mel-Cepstral Coefficient feature.

[0050] After completing the cluster analysis of the Mel-Cepstral Coefficient (MCC) features and identifying the main cluster with the largest sample size, the feature data of all samples within this cluster needs to be refined to extract representative features. First, feature values ​​for all samples within the main cluster are extracted along each MCC dimension, forming a numerical set. Next, the numerical sets for each dimension are sorted, and the median is selected as the representative feature value for that dimension. The selection of the median effectively resists the interference of extreme outliers, ensuring the robustness of the feature representation. Subsequently, the median values ​​calculated along all MCC dimensions are concatenated in the original dimensional order to form a complete median sequence. This sequence represents the optimized MCC features extracted from the main cluster. This process not only preserves the core statistical characteristics of most samples within the main cluster but also significantly reduces the impact of random fluctuations and noise in small sample data through median aggregation, thereby generating a stable and highly representative MCC feature vector.

[0051] 206. The optimized Mel-frequency cepstral coefficient features are concatenated with the time-domain and frequency-domain features in the multi-domain feature vector to form an optimized feature vector adapted to small sample scenarios. First, ensure that the dimensions and order of the three types of features strictly correspond to each original sample frame. Then, according to the predetermined feature arrangement logic, the statistical sequence representing time-domain characteristics, the energy and centroid indices reflecting frequency-domain distribution, and the cluster-optimized Mel-frequency cepstral coefficient median sequence are sequentially concatenated into a unified high-dimensional feature vector. This concatenation process not only maintains the physical independence of each feature domain but also enhances the integrity and discriminative power of feature expression through a structured fusion method. The resulting optimized feature vector fully integrates the complementary information of the signal in the time, frequency, and time-frequency domains, while leveraging the median representation to improve the stability and generalization ability of the features in small sample data environments.

[0052] 207. Input the optimized feature vector into the pre-trained RBF kernel support vector machine classification model to obtain the classification result of the idler roller state; The optimized feature vector is used as input and loaded into a radial basis function kernel function support vector machine classification model that has been trained and saved. This model uses the kernel function to map the low-dimensional nonlinear features of the input to a high-dimensional space and makes classification decisions through the optimal hyperplane that minimizes structural risk. During inference, the model uses the learned support vectors and the similarity weights calculated by the kernel function to perform distance measurement and category discrimination on the input features in the high-dimensional space. Finally, it outputs classification labels corresponding to the running status of the idler roller, including three categories of recognition results: normal operation, abnormal noise fault, and equipment shutdown. This enables accurate and real-time status identification of the audio features of the idler roller in small sample scenarios.

[0053] 208. Based on the classification results, combined with the sliding window counting mechanism and alarm cooling mechanism, determine whether an alarm is triggered, and output alarm information when the alarm conditions are met.

[0054] Based on a preset sliding window containing multiple consecutive classification results, the percentage of frames with the classification result of "abnormal sound" within the sliding window is counted. If the percentage exceeds a preset alarm threshold, an alarm is triggered and a cooling timer is started, during which new alarm triggers are blocked. When an alarm is triggered, at least one of the following operations is executed simultaneously: local audible and visual alarm, abnormal audio data storage, and remote alarm information push.

[0055] In this embodiment of the invention, a feature vector with high discriminative power and strong robustness is constructed through signal enhancement preprocessing, MFCC feature optimization based on clustering screening, and time / frequency domain multi-feature fusion. Combined with the RBF kernel support vector machine model, the classification accuracy and generalization performance of small samples are effectively improved. Through the dual decision-making mechanism of sliding window counting and alarm cooling, the false alarm rate is significantly reduced while ensuring a high detection rate, realizing reliable and stable online intelligent monitoring and early warning of the idler roller's operating status.

[0056] The above describes the roller noise recognition method based on small sample number feature fusion in the embodiments of the present invention. The following describes the roller noise recognition device based on small sample number feature fusion in the embodiments of the present invention. Please refer to... Figure 3 One embodiment of the roller noise recognition device based on small sample number feature fusion in this invention includes: The acquisition module 301 is used to acquire the original audio signal during the operation of the idler roller, perform enhancement preprocessing on the original audio signal, and obtain the enhanced audio signal. The extraction module 302 is used to perform frame-by-frame processing on the enhanced audio signal and extract time-domain features, frequency-domain features and Mel-frequency cepstral coefficient features from each frame to form a multi-domain feature vector. The optimization module 303 is used to perform clustering optimization on the Mel-Cepstral coefficient features in the multi-domain feature vector, and to fuse the optimization results with the time-domain features and frequency-domain features to obtain an optimized feature vector suitable for small sample scenarios. The classification module 304 is used to input the optimized feature vector into the pre-trained RBF kernel support vector machine classification model to obtain the classification result of the roller state; The alarm module 305 is used to determine whether to trigger an alarm based on the classification results, combined with the sliding window counting mechanism and the alarm cooling mechanism, and to output alarm information when the alarm conditions are met.

[0057] In this embodiment of the invention, the original audio signal-to-noise ratio is enhanced through preprocessing. Multi-domain feature extraction in the time domain, frequency domain, and MFCC is performed, and clustering optimization of MFCC features is carried out to address the sparsity of small sample data. This effectively fuses feature representations with higher discriminative power and stronger noise resistance. Combined with a pre-trained RBF kernel support vector machine model, the fault identification accuracy and generalization ability can be significantly improved in small sample scenarios where idler fault audio data is scarce. At the same time, the sliding window counting and alarm cooling mechanism effectively reduce false alarms, realizing reliable online monitoring and intelligent early warning of idler operating status.

[0058] Please see Figure 4 Another embodiment of the roller noise recognition device based on small sample number feature fusion in this invention includes: The acquisition module 301 is used to acquire the original audio signal during the operation of the idler roller, perform enhancement preprocessing on the original audio signal, and obtain the enhanced audio signal. The extraction module 302 is used to perform frame-by-frame processing on the enhanced audio signal and extract time-domain features, frequency-domain features and Mel-frequency cepstral coefficient features from each frame to form a multi-domain feature vector. The optimization module 303 is used to perform clustering optimization on the Mel-Cepstral coefficient features in the multi-domain feature vector, and to fuse the optimization results with the time-domain features and frequency-domain features to obtain an optimized feature vector suitable for small sample scenarios. The classification module 304 is used to input the optimized feature vector into the pre-trained RBF kernel support vector machine classification model to obtain the classification result of the roller state; The alarm module 305 is used to determine whether to trigger an alarm based on the classification results, combined with the sliding window counting mechanism and the alarm cooling mechanism, and to output alarm information when the alarm conditions are met.

[0059] Optionally, the roller noise recognition device based on small sample size feature fusion also includes: Training module 306 is used to collect and construct a small sample training set, which includes three types of idler operation audio samples: normal, abnormal, and stopped. Enhancement processing, frame segmentation processing, multi-feature extraction, and Mel-frequency cepstral coefficient feature analysis are performed on the audio samples in the small sample training set to obtain a training feature vector set. The RBF kernel support vector machine classification model is trained using the training feature vector set, and the hyperparameters are optimized through cross-validation to obtain the RBF kernel support vector machine classification model.

[0060] Optionally, the processing module can be specifically used for: The enhanced audio signal is segmented into frames to obtain several audio frames. Time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features are extracted from each audio frame. The time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features of each audio frame are concatenated to form a multi-domain feature vector.

[0061] Optionally, optimization module 303 includes: Clustering unit 3031 is used to perform clustering operations on the Mel-Cepstral coefficient features in the multi-domain feature vector to form at least two feature clusters; The optimization unit 3032 is used to determine the main cluster with the largest sample size from at least two feature clusters and extract the representative feature values ​​of the main cluster to optimize the feature representation under small sample data and obtain the optimized Mel-Cepstral Coefficient Features. The splicing unit 3033 is used to splice the optimized Mel-frequency cepstral coefficient features with the time-domain and frequency-domain features in the multi-domain feature vector to form an optimized feature vector adapted to small sample scenarios. Optionally, clustering unit 3031 can be specifically used for: The K-means clustering algorithm is used to perform cluster analysis on the Mel-Cepstral Coefficient features. The optimal number of clusters K is determined based on the silhouette coefficient or cross-validation results. Clustering is performed based on the number of clusters K to divide the Mel-Cepstral Coefficient features into K feature clusters, where K is an integer greater than or equal to 2.

[0062] Optionally, the optimization unit 3032 can be specifically used for: Calculate the median of all samples in the main cluster in each Mel-Cepstral Coefficient dimension; use the calculated median sequence as the representative feature value of the main cluster in the corresponding dimension; combine the representative feature values ​​of each dimension to form the optimized Mel-Cepstral Coefficient feature.

[0063] Optionally, the alarm module 305 can be specifically used for: Based on a preset sliding window containing multiple consecutive classification results, the percentage of frames with the classification result of "abnormal sound" within the sliding window is counted. If the percentage exceeds a preset alarm threshold, an alarm is triggered and a cooling timer is started, during which new alarm triggers are blocked. When an alarm is triggered, at least one of the following operations is executed simultaneously: local audible and visual alarm, abnormal audio data storage, and remote alarm information push.

[0064] In this embodiment of the invention, a feature vector with high discriminative power and strong robustness is constructed through signal enhancement preprocessing, MFCC feature optimization based on clustering screening, and time / frequency domain multi-feature fusion. Combined with the RBF kernel support vector machine model, the classification accuracy and generalization performance of small samples are effectively improved. Through the dual decision-making mechanism of sliding window counting and alarm cooling, the false alarm rate is significantly reduced while ensuring a high detection rate, realizing reliable and stable online intelligent monitoring and early warning of the idler roller's operating status.

[0065] above Figure 3 and Figure 4 The roller noise recognition device based on small sample number feature fusion in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The electronic equipment in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0066] See Figure 5 As shown, the electronic device includes a processor 500 and a memory 501. The memory 501 stores machine-executable instructions that can be executed by the processor 500. The processor 500 executes the machine-executable instructions to implement the above-described roller noise recognition method based on small sample number feature fusion.

[0067] Furthermore, Figure 5 The electronic device shown also includes a bus 502 and a communication interface 503. The processor 500, the communication interface 503 and the memory 501 are connected via the bus 502.

[0068] The memory 501 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 503 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 502 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0069] The processor 500 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 500 or by instructions in software form. The processor 500 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this disclosure. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this disclosure can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules may reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 501. The processor 500 reads the information in memory 501 and, in conjunction with its hardware, completes the method steps of the aforementioned embodiment.

[0070] The present invention also provides an electronic device, the computer device including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor performs the steps of the roller noise recognition method based on small sample number feature fusion described in the above embodiments. The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions, and when the instructions are executed on a computer, the computer performs the steps of the roller noise recognition method based on small sample number feature fusion.

[0071] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0072] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0073] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying abnormal noise from idler rollers based on feature fusion with a small sample size, characterized in that, The roller noise recognition method based on small sample size feature fusion includes: The original audio signal of the idler roller during operation is acquired, and the original audio signal is enhanced by preprocessing to obtain the enhanced audio signal. The enhanced audio signal is processed by frame segmentation, and time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features are extracted from each frame to form a multi-domain feature vector. Clustering optimization is performed on the Mel-Cepstral coefficient features in the multi-domain feature vector, and the optimization results are fused with the time-domain features and the frequency-domain features to obtain an optimized feature vector suitable for small sample scenarios; The optimized feature vector is input into a pre-trained RBF kernel support vector machine classification model to obtain the classification result of the idler roller state; Based on the classification results, combined with the sliding window counting mechanism and the alarm cooling mechanism, it is determined whether an alarm is triggered, and alarm information is output when the alarm conditions are met.

2. The method for identifying abnormal noise of idlers based on feature fusion with a small sample size according to claim 1, characterized in that, Before inputting the optimized feature vector into the pre-trained RBF kernel support vector machine classification model, the following steps are also included: Collect and construct a small sample training set, which includes three types of idler roller operation audio samples: normal, abnormal noise, and shutdown. Enhancement processing, frame segmentation processing, multi-feature extraction, and Mel-frequency cepstral coefficient feature analysis are performed on the audio samples in the small training set to obtain a training feature vector set; The RBF kernel support vector machine classification model is trained using the training feature vector set, and the hyperparameters are optimized through cross-validation to obtain the RBF kernel support vector machine classification model.

3. The method for identifying abnormal noise of idlers based on feature fusion with a small sample size according to claim 1, characterized in that, The enhanced audio signal is processed by frame segmentation, and time-domain features, frequency-domain features, and Mel-frequency cepstral coefficient features are extracted from each frame to form a multi-domain feature vector, including: The enhanced audio signal is processed into frames to obtain several audio frames; Extract time-domain features, frequency-domain features, and Mel-spectral coefficient features from each audio frame; The temporal, frequency, and Mel-frequency cepstral coefficient features of each audio frame are concatenated to form a multi-domain feature vector.

4. The method for identifying abnormal noise of idlers based on feature fusion with a small sample size according to claim 1, characterized in that, The step of clustering and optimizing the Mel-frequency cepstral coefficient features in the multi-domain feature vector, and fusing the optimization results with the time-domain features and the frequency-domain features to obtain an optimized feature vector suitable for small sample scenarios, includes: Clustering operations are performed on the Mel-Cepstral coefficient features in the multi-domain feature vector to form at least two feature clusters; From the at least two feature clusters, determine the main cluster with the largest sample size, and extract the representative feature values ​​of the main cluster to optimize the feature representation under small sample data, and obtain the optimized Mel-Cepstral Coefficient Features; The optimized Mel-frequency cepstral coefficient features are concatenated with the time-domain and frequency-domain features in the multi-domain feature vector to form the optimized feature vector adapted to small sample scenarios.

5. The roller noise recognition method based on small sample size feature fusion according to claim 4, characterized in that, The clustering operation on the Mel-frequency cepstral coefficient features in the multi-domain feature vector to form at least two feature clusters includes: The K-means clustering algorithm was used to perform cluster analysis on the Mel-frequency cepstral coefficient features; Determine the optimal number of clusters K based on the silhouette coefficient or cross-validation results; Clustering is performed based on the number of clusters K, dividing the Mel-Cepstral coefficient features into K feature clusters, where K is an integer greater than or equal to 2.

6. The roller noise recognition method based on small sample size feature fusion according to claim 4, characterized in that, The step of extracting representative feature values ​​of the main cluster to optimize the feature representation under small sample data and obtain optimized Mel-frequency cepstral coefficient features includes: Calculate the median of all samples in the main cluster along each Mel-Cepstral Coefficient dimension; The calculated median sequence is used as the representative feature value of the main cluster in the corresponding dimension; The combination of representative feature values ​​from each dimension forms the optimized Mel-Cepstral coefficient features.

7. The method for identifying abnormal noise of idlers based on feature fusion with a small sample size according to any one of claims 1-6, characterized in that, Based on the classification results, combined with the sliding window counting mechanism and the alarm cooling mechanism, it is determined whether an alarm is triggered, and alarm information is output when the alarm conditions are met, including: Based on a preset sliding window containing multiple consecutive classification results, the percentage of frames with the classification result of "abnormal sound" within the sliding window is counted. If the percentage exceeds the preset alarm threshold, an alarm is triggered and a cooling timer is started, during which new alarm triggers are blocked. When an alarm is triggered, at least one of the following operations is executed simultaneously: local audible and visual alarm, abnormal audio data storage, and remote alarm information push.

8. A roller noise recognition device based on feature fusion with a small sample size, characterized in that, The roller noise recognition device based on small sample size feature fusion includes: The acquisition module is used to acquire the original audio signal during the operation of the idler roller, perform enhancement preprocessing on the original audio signal, and obtain the enhanced audio signal. The extraction module is used to perform frame-by-frame processing on the enhanced audio signal and extract time-domain features, frequency-domain features and Mel-frequency cepstral coefficient features from each frame to form a multi-domain feature vector. An optimization module is used to perform cluster optimization on the Mel-frequency cepstral coefficient features in the multi-domain feature vector, and fuse the optimization results with the time-domain features and the frequency-domain features to obtain an optimized feature vector adapted to small sample scenarios. The classification module is used to input the optimized feature vector into a pre-trained RBF kernel support vector machine classification model to obtain the classification result of the roller state; The alarm module is used to determine whether to trigger an alarm based on the classification results, combined with the sliding window counting mechanism and the alarm cooling mechanism, and to output alarm information when the alarm conditions are met.

9. A roller noise recognition device based on small sample size feature fusion, characterized in that, The roller noise recognition device based on small sample number feature fusion includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the roller noise recognition device based on small sample number feature fusion to perform the roller noise recognition method based on small sample number feature fusion as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instruction is executed by the processor, it implements the roller noise recognition method based on feature fusion with a small number of samples as described in any one of claims 1-7.

Images