Fan abnormal sound intelligent discrimination and detection method based on time-frequency diagram and deep learning

Abstract

The invention relates to an intelligent fan abnormal sound distinguishing and detecting method and system based on a time-frequency diagram and deep learning. The method comprises the steps that a fan sound signal and a fan rotating speed signal are collected; identifying the fan sound signal and intercepting a stable section sound signal; a weighting processing is carried out on the sound signal in the stable section so as to obtain sound features reflecting loudness perception of human ears; performing short-time Fourier transform on the sound signal subjected to A weighting processing to obtain a time-frequency graph; and inputting the time-frequency diagram through a ResNet18 deep learning model to carry out fan abnormal sound classification so as to obtain an intelligent detection result of the fan operation state. Intelligent detection of the running state of the fan is achieved.

Description

Technical Field

[0001] The present invention relates to an intelligent identification and detection method for abnormal noise of a fan based on a time-frequency diagram and deep learning, and belongs to the field of fan detection. Background Art

[0002] In the manufacturing and quality control process of fans, abnormal noise detection of fans is a key link to ensure product quality and performance. The traditional manual listening detection method has been widely used for a long time. The inspectors rely on their experience and hearing to judge whether the fan has abnormal noise. However, this manual method has many limitations. On the one hand, manual detection is highly dependent on the professional skills, mental state and experience level of the inspectors. It is difficult for different inspectors to maintain completely consistent judgment standards, which can easily lead to subjective differences and misjudgments in the test results. On the other hand, with the continuous expansion of the scale of fan production and the increasing requirements for production efficiency, the low efficiency of manual detection has gradually become a bottleneck restricting the optimization of production processes and the improvement of production capacity. It cannot meet the needs of accurate and efficient detection of abnormal noise of fans in large-scale and rapid production environments, and the traceability is poor. Summary of the invention

[0003] The present invention provides an intelligent identification and detection method for abnormal fan noise based on time-frequency diagram and deep learning, aiming to solve at least one of the technical problems existing in the prior art.

[0004] The technical solution of the present invention relates to an intelligent identification and detection method for abnormal fan noise based on time-frequency diagram and deep learning. The method according to the present invention comprises the following steps:

[0005] S100, collecting fan sound signals and fan speed signals;

[0006] S200, identifying the fan sound signal and intercepting the steady-state sound signal; performing A-weighting processing on the steady-state sound signal to obtain a sound feature that reflects the loudness perception of the human ear;

[0007] S300, performing short-time Fourier transform on the sound signal after A-weighting processing to obtain a time-frequency diagram; inputting the time-frequency diagram into the ResNet18 deep learning model to classify the abnormal sound of the fan, so as to obtain an intelligent detection result of the operating status of the fan.

[0008] Further, in step S200:

[0009] The A-weighting process attenuates low-frequency sounds and enhances high-frequency sounds, wherein the frequency response function H of the A-weighting is A (f) is expressed as follows:

[0010]

[0011] Wherein, f represents frequency.

[0012] Further, in step S300:

[0013] The short-time Fourier transform STFT(t,f) is expressed as follows:

[0014]

[0015] In the formula, t represents time; x(t) represents the original signal; w(t-τ) represents the window function, where τ is the central time point of the window function, indicating that the spectrum analysis of the signal is performed at this time position; f represents frequency.

[0016] Further, in step S300, the ResNet18 deep learning model includes an input layer, an initial convolutional layer, a maximum pooling layer, a residual module, a flat pooling layer and a fully connected layer connected in sequence.

[0017] Furthermore, the residual module network structure of the ResNet18 deep learning model includes a first-stage module, a second-stage module, a third-stage module and a fourth-stage module: wherein,

[0018] The first-stage module contains two residual blocks, and the output channels are set to 64;

[0019] The second-stage module contains two residual blocks, and the output channels are set to 128;

[0020] The third stage module includes two residual blocks, and the output channel is set to 256;

[0021] The fourth stage module contains two residual blocks and the output channels are set to 512.

[0022] Furthermore, the residual block includes two convolutional layers, each of the convolutional layers uses a 3×3 convolution kernel, and each of the convolutional layers is respectively connected to a ReLU activation function and batch normalization.

[0023] Furthermore, the ResNet18 deep learning model uses a cross entropy loss function to calculate the difference between the predicted value and the true value. The cross entropy loss function is expressed as follows:

[0024]

[0025] In the formula, y i represents the true category label, p i Represents the probability of the model output.

[0026] Furthermore, the ResNet18 deep learning model uses the Adam optimizer to update the network weights, and the update rule of the Adam optimizer is set as follows:

[0027]

[0028] In the formula, t represents time, represents the current gradient, m t and v t They represent the estimates of the first-order moment and the second-order moment of the gradient at time t, respectively. and They represent the first-order moment after bias correction and the second-order moment after bias correction respectively; β 1 and β 2 is the attenuation factor, and They represent the attenuation factor β 1 The exponential power and decay factor β at step t 2 The exponential power at step t; α is the learning rate; θ t represents the model parameters at the tth step, θ t+1 represents the model parameters after the next update; ∈ is a constant to prevent division by zero errors.

[0029] The technical solution of the present invention also relates to a computer-readable storage medium on which program instructions are stored, and the above-mentioned method is implemented when the program instructions are executed by a processor.

[0030] The technical solution of the present invention also relates to an intelligent identification and detection system for abnormal fan noise based on time-frequency diagram and deep learning, wherein the system includes a computer device, which contains the above-mentioned computer-readable storage medium.

[0031] The beneficial effects of the present invention are as follows:

[0032] In order to overcome the shortcomings of traditional manual listening detection, the present invention proposes an intelligent detection method for abnormal fan noise based on short-time Fourier transform and deep learning model. By performing short-time Fourier transform on the collected variable speed signal and generating a time-frequency graph, and using the ResNet18 deep learning model to train and classify the time-frequency graph, the intelligent detection of the fan operation status is realized. The present invention uses data preprocessing and short-time Fourier transform (STFT) to extract the time-frequency characteristics of the fan signal, which can significantly reduce the influence of background noise. Combined with the ResNet18 deep learning model, through data-driven objective judgment, it can effectively get rid of manual dependence and reduce manual errors, which is conducive to ensuring the stability and reliability of the detection results, while improving the detection efficiency, and realizing data traceability and quality control enhancement. It can fully record the detection data including the original sound signal, abnormal sound characteristics, prediction results and time, and can trace the quality of a specific fan or batch, trace back to locate the root cause when there is a problem, and assist in quality improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1is a basic flow chart of the method according to the present invention.

[0034] Figure 2 : It is a neural network structure diagram of the ResNet18 deep learning model according to the method of the present invention. DETAILED DESCRIPTION

[0035] The concept, specific structure and technical effects of the present invention will be clearly and completely described below in combination with the embodiments and drawings to fully understand the purpose, scheme and effect of the present invention.

[0036] It should be noted that, unless otherwise specified, when a feature is referred to as being "fixed" or "connected" to another feature, it may be directly fixed or connected to another feature, or it may be indirectly fixed or connected to another feature. The singular forms "a", "said" and "the" used herein are also intended to include the plural forms, unless the context clearly indicates otherwise. In addition, unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art. The terms used in this specification are intended only to describe specific embodiments and are not intended to limit the invention. The term "and / or" used herein includes any combination of one or more of the related listed items.

[0037] It should be understood that, although the term first, second, third etc. may be adopted to describe various elements in the present disclosure, these elements should not be limited to these terms. These terms are only used to distinguish the same type of elements from each other. For example, without departing from the scope of the present disclosure, the first element may also be referred to as the second element, and similarly, the second element may also be referred to as the first element. The use of any and all examples or exemplary language ("for example", "such as" etc.) provided herein is only intended to better illustrate embodiments of the present invention, and unless otherwise required, the scope of the present invention will not be limited.

[0038] Reference Figure 1 to Figure 2 In some embodiments, the method for intelligently distinguishing and detecting abnormal fan noise based on time-frequency diagram and deep learning according to the present invention comprises at least the following steps:

[0039] S100, collecting fan sound signals and fan speed signals;

[0040] S200, identifying the fan sound signal and intercepting the steady-state sound signal; performing A-weighting processing on the steady-state sound signal to obtain a sound feature that reflects the loudness perception of the human ear;

[0041] S300, performing short-time Fourier transform on the sound signal after A-weighting processing to obtain a time-frequency diagram; inputting the time-frequency diagram into the ResNet18 deep learning model to classify the abnormal fan noise, so as to obtain an intelligent detection result of the fan operation status.

[0042] The intelligent detection method and system of abnormal fan noise based on short-time Fourier transform and deep learning model of the present invention realizes intelligent detection of the fan operating status by performing short-time Fourier transform on the collected variable speed signal and generating a time-frequency graph, and using the ResNet18 deep learning model to train and classify the time-frequency graph.

[0043] In some embodiments, when collecting fan sound signals, the present invention places the fan in an existing soundproof box to reduce the interference of environmental noise, and uses a high-sensitivity microphone to collect sound signals during operation. The microphone is fixedly installed perpendicular to the fan, 5 cm above the center of the fan, and points to the center of the fan. The system sampling frequency is set to 48kHz to ensure that the high-frequency components of the signal are fully captured, and the analog signal is converted into a digital signal through a high-precision A / D conversion module to ensure the integrity and accuracy of the data. Furthermore, the system also synchronously collects fan speed pulse signals, where 2 pulses are recorded per revolution, so as to achieve precise alignment of the sound signal with the mechanical state of the fan in subsequent analysis. After the system completes the collection of the fan sound signal, the collected data is transmitted to the data preprocessing unit for further processing.

[0044] In some embodiments, the present invention identifies and intercepts a sound signal segment of the fan in a stable operating state from the collected fan sound signal. It should be noted that during the start and stop process of the fan, the sound signal will be affected by a variety of unstable factors, such as starting current shock, speed change, etc. The present invention uses a signal of a stable speed segment to better reflect the sound characteristics of the fan during normal operation, which is beneficial to subsequent abnormal sound detection.

[0045] In some embodiments, the present invention performs A-weighting processing on the obtained fan sound signal in the stable section, so that the processed signal is more consistent with the human ear's subjective perception of sound loudness, which helps to extract sound features that are more practical and relevant. It should be noted that the A-weighting processing method adopted by the present invention can simulate the human ear's perception of the loudness of sounds of different frequencies and weight them. Specifically, based on the equal loudness curve, in acoustic measurement, by weighting the different frequency components of the sound signal, the measurement results are more consistent with the subjective perception of the human ear. Furthermore, the A-weighting processing method of the present invention attenuates low-frequency sounds to a certain extent and appropriately enhances high-frequency sounds, thereby reflecting the characteristics of the human ear in loudness perception.

[0046] Specifically, the A-weighted frequency response function H of the present invention is A(f) is expressed as follows:

[0047]

[0048] Wherein, f represents frequency, and its unit is Hz.

[0049] In some embodiments, the present invention extracts and screens the fan abnormal sound features from the sound signal after A-weighting processing. Specifically, in the detection of fan abnormal sound, feature extraction and screening play a vital role. The fan abnormal sound features of the present invention mainly include 1 / 12 octave, 15 constant bandwidth energy, average spectrum, sound pressure level, loudness, etc.

[0050] In one application embodiment, for the 1 / 12 octave fan abnormal sound feature extraction, the 1 / 12 octave is an analysis method that divides the frequency range into finer intervals, and has important applications in the field of fan abnormal sound detection. The 1 / 12 octave divides the audible frequency range (usually 20Hz-20kHz) into smaller frequency bands. Compared with traditional octave analysis, it can more finely reveal the characteristic changes of sound at different frequencies. This is very important for detecting fan abnormal sounds, because the abnormal sound of the fan may appear in a specific narrow frequency range, and the 1 / 12 octave can more accurately locate these abnormal frequency areas.

[0051] The center frequency f of the 1 / 12 octave band ci According to a specific geometric series distribution, the formula is:

[0052] f ci =f 0 2 i / 12

[0053] In the formula, f 0 is the initial frequency, which is set to 20Hz in this project; i is the octave number.

[0054] The frequency range of each octave is determined by the upper and lower frequency limits f li and f ui The specific formula is:

[0055] f li =f ci 2 -1 / 24

[0056] f ui =f ci 2 1 / 24

[0057] The upper and lower frequency limits are based on the center frequency and are proportional to 2 ±1 / 24 Calculate to ensure that the width of each frequency band is symmetrical on a logarithmic scale. The octave frequency bandwidth is:

[0058] BW i =f ui -f li

[0059] Energy is usually the integral (or accumulation in the case of discrete sampling) of the signal's power spectral density (PSD):

[0060]

[0061] If discrete spectrum representation is used (power spectrum P(fk) calculated by FFT):

[0062]

[0063] Where P(fk)=|X(fk)| 2 is the power value of the spectrum; Δf=fs / N is the spectrum resolution, and N is the number of FFT points.

[0064] In some embodiments, the present invention performs a short-time Fourier transform (STFT) on the sound signal after A-weighting processing to obtain a time-frequency diagram. It is understandable that many actual sound signals are non-stationary, that is, the frequency components of the sound signal will change over time. In order to effectively analyze these non-stationary sound signals, the present invention uses a short-time Fourier transform (STFT) for time-frequency analysis, and converts the sound signal from the time domain to the time-frequency domain, thereby obtaining the distribution information of the signal in time and frequency, and through the STFT transformation method, the frequency characteristics of the signal at different time points can be obtained at the same time, revealing the trend of the signal spectrum over time. Among them, the present invention performs a short-time Fourier transform on the signal after A-weighting processing to obtain a time-frequency diagram, and the image size of the time-frequency diagram is 224×224×3.

[0065] It should be noted that the short-time Fourier transform (STFT) used in the present invention can be used to analyze non-stationary signals, convert the signal from the time domain to the time-frequency domain, and simultaneously obtain the frequency components of the signal at different time points, thereby showing the change of the spectrum of the signal over time. Among them, the STFT transformation of the present invention performs a window function weighting on the signal, divides the signal into multiple short time periods, and performs a Fourier transform in each short time period, so that its spectrum can be analyzed.

[0066] Specifically, the signal is represented by x(t), the window function is represented by w(t), and the short-time Fourier transform STFT(t,f) is expressed as follows:

[0067]

[0068] In the formula, t represents time; x(t) represents the original signal; w(t-τ) represents the window function, where τ is the central time point of the window function, indicating that the spectrum analysis of the signal is performed at this time position; f represents frequency.

[0069] Further, in the spectrum diagram, the frequency resolution Δf=Fs / N=5Hz, where Fs represents the sampling frequency, the Fs value can be 48kHz, N is the window length, i.e., the number of FFT points, and the overlap ratio can be set to 90%. It should be noted that the window function type used in the present invention can be a Blackman-Harris window function.

[0070] In some embodiments, the present invention constructs a ResNet18 deep learning model to train and classify the time-frequency graph, thereby realizing intelligent detection of the fan operating status. It can be understood that ResNet18, as a deep residual network, has powerful feature extraction and classification capabilities. The present invention can realize effective analysis and classification of fan sound signals by constructing a neural network structure of the ResNet18 deep learning model. Specifically, ResNet18 solves the gradient vanishing and representation bottleneck problems in the deep network training process by introducing residual connections, wherein the residual connections allow the network to more easily propagate gradients and information during the training process, thereby enabling the network to be trained deeper and more efficiently.

[0071] Specifically, the network structure of ResNet18 of the present invention consists of multiple parts, including the initial convolutional layer, the residual block and the fully connected layer, see Figure 2 , its network architecture is as follows:

[0072] Input layer (image): receives image input, where the input frequency spectrum map size is 224×224×3, where 3 represents the RGB channels.

[0073] Initial convolution layer (conv): The convolution kernel size is 7×7, the stride is 2, and the number of output channels is 64. The first convolution layer uses the ReLU activation function and is normalized to batch normalization. The ReLU function is expressed as follows:

[0074]

[0075] Where x is the input signal and f(x) is the output signal after activation.

[0076] Max pooling layer (maxpool): The pooling kernel size is 3×33 and the stride is 2.

[0077] Residual module: Each residual block contains two convolutional layers, each of which uses a 3×3 convolution kernel, and each convolutional layer is connected to the ReLU activation function and batch normalization. The jump connection directly connects the input and output through the identity mapping to alleviate the gradient disappearance problem.

[0078] Avgpool and FC: The input feature count of the fully connected layer is 512 and the output feature count is 1000

[0079] In an application embodiment, the residual module network structure of the present invention is composed of four stage modules, wherein the first stage module (Stage 1) includes 2 residual blocks and the number of output channels is 64; the second stage module (Stage 2) includes 2 residual blocks and the number of output channels is 128; the third stage module (Stage 3) includes 2 residual blocks and the number of output channels is 256; the fourth stage module (Stage 4) includes 2 residual blocks and the number of output channels is 512.

[0080] In an application embodiment, the ResNet18 deep learning model of the present invention. Specifically, after short-time Fourier transform processing, the generated time-frequency graph is used as the input of the deep learning model ResNet18 for fan abnormal sound classification. The resolution of the time-frequency graph is set to 224×224 pixels, and the RGB three-channel format is used to meet the input requirements of ResNet18. The label of the image is annotated based on the manual listening results, where "normal fan" is defined as category 1 and "abnormal fan" is defined as category 0.

[0081] See also Figure 2 The ResNet18 deep learning model of the present invention is based on the ResNet residual network, which is mainly composed of a residual module group consisting of a convolutional layer. The input of the residual module is directly added to the output of the module through an identity mapping. The deep residual network is composed of multiple residual modules. For example, the model of the present invention is composed of four stage modules. The residual network of each stage model can fit the error of the previous classifier (see Figure 2 The basic structure of the residual module) is used to improve the classification ability.

[0082] Specifically, the ResNet18 deep learning model of the present invention is composed of 17 convolutional layers and 1 fully connected layer, wherein the 17 convolutional layers are composed of an independent convolution and 4 residual modules, namely, the initial convolutional layer, the first stage module, the second stage module, the third stage module and the fourth stage module. The network is composed of residual modules of multiple stages, and the residual modules of each stage merge the input and output information of the module by means of identity mapping. The last residual module is connected to the fully connected layer (FC) to reduce the fitting error, and finally the probability of the category is obtained through the Softmax operation. See Figure 2 The network structure of the ResNet18 deep learning model of the present invention, and its specific parameters are shown in Table I. Furthermore, the network parameter settings of the ResNet18 deep learning model of the present invention are shown in Table II.

[0083]

[0084] Table I

[0085]

[0086]

[0087]

[0088]

[0089]

[0090]

[0091]

[0092] Table II

[0093] In some embodiments, in order to ensure the training effect and generalization ability of the model, all images and their corresponding labels are organized into a standard data set format and randomly shuffled before use. Subsequently, the data set is divided into a training set and a test set at a ratio of 80% and 20%. The training set is used to optimize the model parameters and improve the classification performance by minimizing the loss function; the test set is used to evaluate the generalization ability of the model and adjust and verify the hyperparameters during the training process. Through this standardized processing, the scientific nature of the data and the reliability of the experimental results are ensured.

[0094] Specifically, the loss function of the ResNet18 deep learning model of the present invention is a cross entropy loss function, which is used to calculate the difference between the predicted value and the true value, and its formula is:

[0095]

[0096] Among them, y i is the true category label, p i is the probability of the model output.

[0097] Furthermore, the present invention uses an optimization algorithm to update the network weights. Specifically, the present invention uses the Adam optimizer, which has the characteristics of adaptive learning rate, can accelerate convergence and prevent overfitting. The update rule of the Adam optimizer is set as follows:

[0098]

[0099] In the formula, t represents time, represents the current gradient, m t and vt They represent the estimates of the first-order moment and the second-order moment of the gradient at time t, respectively. and They represent the first-order moment after bias correction and the second-order moment after bias correction respectively; β 1 and β 2 is the attenuation factor, and They represent the attenuation factor β 1 The exponential power and decay factor β at step t 2 The exponential power at step t; α is the learning rate; θ t represents the model parameters at the tth step, θ t+1 represents the model parameters after the next update; ∈ is a constant to prevent division by zero errors.

[0100] Furthermore, the learning rate can be set to 0.001. It is understandable that the learning rate determines the step size of the model parameters at each update. A smaller learning rate can ensure a smoother convergence process, but may result in longer training time. Furthermore, the batch size can be set to 64. Setting it to 64 means that 64 samples will be used to calculate the gradient and update the network parameters each time the optimization is performed. A smaller batch size will result in unstable training, but the computational efficiency is higher. Furthermore, the number of training epochs can be set to 50.

[0101] After the model training is completed, the model performance is tested and evaluated. Specifically, during the model evaluation process, the test set is input into the trained model, and the classification accuracy, confusion matrix and loss function are used as the main evaluation indicators. Among them, the classification accuracy is used to measure the proportion of samples correctly classified by the model, and is an intuitive indicator for evaluating the overall performance of the model. The confusion matrix provides detailed information on the classification results, which includes true positive examples, false positive examples, true negative examples and false negative examples to reflect the classification ability of the model from multiple dimensions. The loss function is used to quantify the difference between the model prediction and the actual label. The smaller the value, the closer the model's prediction result is to the true value. The present invention can comprehensively evaluate the classification performance of the model and its optimization effect by comprehensively analyzing these three indicators. It can be understood that after obtaining the optimal training parameters, the deep learning model based on ResNet18 together with its network structure and weight parameters will be saved to ensure that the model can be directly loaded and used in the subsequent detection process to avoid the waste of time and resources caused by repeated training.

[0102] In the actual detection task, the present invention needs to perform standardized preprocessing and feature conversion on the newly collected fan sound signal. First, the sound signal is intercepted, and the signal in the stable operation stage is retained, and then the signal is A-weighted to be more in line with the auditory perception characteristics of the human ear. Then, the signal is converted into a time-frequency spectrum using the short-time Fourier transform (STFT), and it is used as the input data form of the model. These time-frequency spectrum graphs need to go through the same processing steps as the training stage to ensure that the format of the input data is consistent with the input requirements of the ResNet18 model, thereby maximizing the detection accuracy and stability of the model. After completing the feature conversion, the generated time-frequency spectrum graph is input into the saved and trained ResNet18 deep learning model. The model extracts multi-level spatial features from the input image through its convolutional layer, and uses its pre-trained feature representation and optimized fully connected layer to complete the classification prediction. In the final decision layer of the model, combined with the weight parameters learned during training, the category of the input data is probabilistically predicted. By analyzing the output probability distribution, it can quickly determine whether the input sample belongs to a normal fan or an abnormal sound fan.

[0103] The fan abnormal noise intelligent discrimination detection method and system based on time-frequency diagram and deep learning of the embodiment of the present invention are verified by blind side. Specifically, after the ResNet18 model training is completed and saved, three independent blind test verifications are performed to comprehensively evaluate the classification ability and generalization performance of the model in practical applications. Among them, the blind test data is completely independent of the training set and the test set, and strictly simulates the real detection scene to ensure the scientificity and objectivity of the evaluation results.

[0104] Furthermore, in each blind test, the newly collected fan sound data is first preprocessed to generate a time-frequency graph of the same size as that during training. Subsequently, the time-frequency graph is input into the trained ResNet18 deep learning model for classification prediction. The results of the three blind tests are shown in Tables 1, 2, and 3. According to the tabular data, the constructed ResNet18 model not only has a high classification accuracy, but also exhibits good robustness and generalization capabilities, and can effectively cope with complex scenarios in actual industrial detection. The successful verification of the blind test results further proves the scientificity and rationality of the feature selection and parameter optimization strategies used in the model design of the present invention, and provides a reference for the practical application of the model in industrial fan abnormal sound detection.

[0105]

[0106] Table 1

[0107]

[0108] It should be appreciated that the method steps in the embodiments of the present invention can be implemented or implemented by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer readable memory. The method can use standard programming techniques. Each program can be implemented in a high-level process or object-oriented programming language to communicate with a computer system. However, if necessary, the program can be implemented in an assembly or machine language. In any case, the language can be a compiled or interpreted language. In addition, the program can be run on a programmed ASIC for this purpose.

[0109] Furthermore, the operations of the processes described herein may be performed in any suitable order unless otherwise indicated herein or otherwise clearly contradicted by context. The processes described herein (or variations and / or combinations thereof) may be performed under the control of one or more computer systems configured with executable instructions, and may be implemented as code (e.g., executable instructions, one or more computer programs, or one or more applications) that is executed collectively on one or more processors, by hardware, or a combination thereof. The computer program includes a plurality of instructions that may be executed by one or more processors.

[0110] Further, the method can be implemented in any type of computing platform that is operably connected to a suitable computer, including but not limited to a personal computer, a minicomputer, a mainframe, a workstation, a network or distributed computing environment, a separate or integrated computer platform, or in communication with a charged particle tool or other imaging device, etc. Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, an optical read and / or write storage medium, an RSM, a ROM, etc., so that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the process described herein. In addition, the machine-readable code, or portions thereof, can be transmitted via a wired or wireless network. When such media includes instructions or programs that implement the steps described above in conjunction with a microprocessor or other data processor, the invention described herein includes these and other different types of non-transitory computer-readable storage media. When programmed according to the methods and techniques of the present invention, the present invention can also include the computer itself.

[0111] The computer program can be applied to input data to perform the functions described herein, thereby converting the input data to generate output data stored in a non-volatile memory. The output information can also be applied to one or more output devices such as a display. In a preferred embodiment of the present invention, the converted data represents physical and tangible objects, including specific visual depictions of physical and tangible objects produced on the display.

[0112] The above is only a preferred embodiment of the present invention. The present invention is not limited to the above implementation. As long as the technical effect of the present invention is achieved by the same means, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of protection of the present invention. Within the scope of protection of the present invention, its technical scheme and / or implementation method may have various modifications and changes.

Claims

1. Intelligent identification and detection method of abnormal fan noise, characterized in that: The method comprises the following steps: S100, collecting fan sound signals and fan speed signals; S200, identifying the fan sound signal and intercepting the steady-state sound signal; performing A-weighting processing on the steady-state sound signal to obtain a sound feature that reflects the loudness perception of the human ear; S300, performing short-time Fourier transform on the sound signal after A-weighting processing to obtain a time-frequency diagram; inputting the time-frequency diagram into the ResNet18 deep learning model to classify the abnormal sound of the fan to obtain the detection result of the fan operation status.

2. The method according to claim 1, characterized in that In step S200: The A-weighting process attenuates low-frequency sounds and enhances high-frequency sounds, wherein the frequency response function H of the A-weighting is A (f) is expressed as follows: Wherein, f represents frequency.

3. The method according to claim 1, characterized in that In step S300: The short-time Fourier transform STFT(t,f) is expressed as follows: In the formula, t represents time; x(t) represents the original signal; w(t-τ) represents the window function, where τ is the central time point of the window function, indicating that the spectrum analysis of the signal is performed at this time position; f represents frequency.

4. The method according to claim 1, characterized in that In step S300, the ResNet18 deep learning model includes an input layer, an initial convolutional layer, a maximum pooling layer, a residual module, a flat pooling layer and a fully connected layer connected in sequence.

5. The method according to claim 4, characterized in that The residual module network structure of the ResNet18 deep learning model includes a first-stage module, a second-stage module, a third-stage module and a fourth-stage module: wherein, The first-stage module contains two residual blocks, and the output channels are set to 64; The second-stage module contains two residual blocks, and the output channels are set to 128; The third stage module includes two residual blocks, and the output channel is set to 256; The fourth stage module contains two residual blocks and the output channels are set to 512.

6. The method according to claim 5, characterized in that The residual block includes two convolutional layers, each of which uses a 3×3 convolution kernel, and each of which is connected to a ReLU activation function and batch normalization.

7. The method according to claim 1, characterized in that The ResNet18 deep learning model uses a cross entropy loss function to calculate the difference between the predicted value and the true value. The cross entropy loss function is expressed as follows: In the formula, y i represents the true category label, p i Represents the probability of the model output.

8. The method according to claim 1, characterized in that The ResNet18 deep learning model uses the Adam optimizer to update the network weights. The update rule of the Adam optimizer is set as follows: In the formula, t represents time, represents the current gradient, m t and v t They represent the estimates of the first-order moment and the second-order moment of the gradient at time t, respectively. and They represent the first-order moment after bias correction and the second-order moment after bias correction respectively; β1 and β2 are attenuation factors, and They represent the exponential power of the decay factor β1 at the tth step and the exponential power of the decay factor β2 at the tth step respectively; α is the learning rate; θ t represents the model parameters at the tth step, θ t+1 represents the model parameters after the next update; ∈ is a constant to prevent division by zero errors.

9. A computer-readable storage medium having program instructions stored thereon, wherein the program instructions are executed by a processor to implement the method according to any one of claims 1 to 8.

10. Fan abnormal sound intelligent identification and detection system, characterized by: include: A computer device comprising a computer readable storage medium according to claim 9.

Images