Industrial air conditioner fault self-diagnosis method based on audio analysis

Abstract

The invention discloses an industrial air conditioner fault self-diagnosis method based on audio analysis, and the method comprises the steps: 1, setting an audio collection standard, and synchronously obtaining the operation sound of equipment in a normal state and a fault state; 2, a sample set with labels is established, and signal preprocessing is carried out through methods of silent segment filtering, data segmentation, wavelet denoising and the like; 3, performing feature extraction on the audio signals to construct a training data set; and 4, performing model training by using a neural network, and screening an optimal diagnosis model through an adaptive learning mechanism. According to the method, the characteristic difference between the normal signal and the abnormal signal in the audio signal can be effectively utilized, and the neural network is combined for efficient distinguishing, so that high-precision fault diagnosis can be realized, the diagnosis efficiency is improved, and the safety and reliability of the industrial air conditioner in actual motion are improved.

Description

Technical Field

[0001] The present invention relates to fault diagnosis of industrial air conditioners, including signal processing and deep learning fields, and adopts a method for realizing equipment status diagnosis by combining audio frequency domain features with deep learning. Background Art

[0002] In the tobacco manufacturing industry, the industrial air conditioning system is an essential production auxiliary system in production. Its main function is to ensure the temperature and humidity balance in the silk reeling workshop and the cigarette making and packing workshop to meet the temperature and humidity requirements of the production environment. When the air conditioning unit is operating, faults are inevitable. However, the inspection personnel do not have the skills to repair, and they need to notify the repairman. The repairman needs to conduct a manual fault investigation before carrying out the repair work, resulting in a long repair time and affecting the temperature and humidity balance in the production workshop. Summary of the Invention

[0003] The present invention is to solve the above-mentioned deficiencies existing in the prior art, and proposes a method for self-diagnosis of industrial air conditioner faults based on audio analysis, in order to improve the fault diagnosis efficiency and accuracy of industrial air conditioners, thereby ensuring the safety and reliability of industrial air conditioners in actual operation.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A method for self-diagnosis of industrial air conditioner faults based on audio analysis according to the present invention is characterized by including the following steps:

[0006] Step S1: Use the same device to collect the standard audio signal set S under the normal state of the industrial air conditioner std and the abnormal audio signal set S under the fault state abn , and form the original audio data set D raw ;

[0007] Step S2: Locate the original audio data set D raw , and delete the invalid audio segments in the equipment startup / shutdown stage to obtain the effective audio signal set D valid . Then, use the wavelet denoising algorithm to process the effective audio signal set D valid to obtain the denoised audio signal set D denoise ;

[0008] Step S3: Classify the denoised audio signal set D denoise according to the normal state and the fault state, and add the corresponding state category label L to obtain the labeled audio signal set D label ;

[0009] Step S4: Extract the audio features in the audio signal set D labe to obtain the audio enhanced feature matrix Xenhance :

[0010] Step S5: Construct an industrial air conditioner fault self-diagnosis network, which successively includes: a convolution operation module, a hybrid pooling module, and a coordinate attention mechanism module, and process X enhance to obtain a state discrimination vector Y containing fault characteristics pred ;

[0011] Step S6: Normalize the state discrimination vector Y pred to generate a diagnosis confidence vector S confidence . Construct a hybrid loss function from the diagnosis confidence vector S confidence and the true state label L, and optimize the network parameters through backpropagation until convergence to obtain a diagnosis model corresponding to the optimal parameter θ*, and this diagnosis model realizes multi-dimensional feature analysis and fault mode discrimination of the air conditioner operating state.

[0012] The characteristics of an industrial air conditioner fault self-diagnosis method based on audio analysis according to the present invention also lie in that step S4 includes the following steps:

[0013] Step S4.1: Segment the audio signals in D label to obtain an audio segment set S segment ;

[0014] Step S4.2: After successively performing pre-emphasis processing and frame segmentation on the audio segment set S segment , perform Hamming windowing on each frame of audio signal to generate a preprocessed audio signal set S pre ;

[0015] Step S4.3: Perform a fast Fourier transform on the preprocessed audio signal set S pre to obtain a frequency-domain signal set S f ;

[0016] Step S4.4: Use a set of Mel filter banks to filter and scale-transform the audio signals in different frequency bands in S f to obtain the Mel spectrum of each frame of audio signal in S f ;

[0017] Use a flipped Mel filter bank to filter and scale-transform the audio signals in different frequency bands in S f to obtain the flipped Mel spectrum of each frame of audio signal in S f ;

[0018] After calculating the logarithmic energy of the Mel spectrum and the logarithmic energy of the flipped Mel spectrum for each frame of the audio signal respectively, the logarithmic spectrum of the Mel spectrum and the logarithmic spectrum of the flipped Mel spectrum of each frame of the audio signal are obtained, so as to perform discrete cosine transform on the logarithmic spectrum of the Mel spectrum and the logarithmic spectrum of the flipped Mel spectrum of each frame of the audio signal respectively, and correspondingly obtain S f MFCC coefficients X of each frame of the audio signal in S MFCC and IMFCC coefficients X IMFCC ;

[0019] Step S4.5: Calculate the first-order difference feature ΔX of X MFCC of each frame of the audio signal respectively MFCC and the first-order difference feature ΔX of X IMFCC ; and after splicing ΔX IMFCC and ΔX MFCC as well as X IMFCC and X MFCC and X IMFCC together, an audio enhancement feature matrix X enhance is obtained.

[0020] Further, step S5 includes the following steps:

[0021] Step 5.1: The convolution operation module processes X enhance using a standard convolutional layer to obtain deep basic features C base ;

[0022] The convolution operation module uses a depthwise separable convolutional layer to decouple cross-channel features of C base to obtain lightweight detail features C detail ;

[0023] Step 5.2: The hybrid pooling module performs max-pooling operation and average-pooling operation on C detail respectively, and correspondingly obtains kurtosis features P max representing significant responses and mean features P avg representing distribution characteristics, so as to obtain hybrid response features F fusion with bimodal characteristics using Equation (1);

[0024]

[0025] In Equation (1), is a random value, ;

[0026] Step 5.3: The attention mechanism module performs pooling compression on F fusion through global average pooling to obtain horizontal incentive With the longitudinal excitation on and splicing, the spatial energy feature S is obtained energy ;

[0027] The spatial information excitation block S is decomposed into the lateral attention weight W energy through a 1D convolution and non-linear activation horizon and the longitudinal attention weight W vertical , and after expansion in the cross-channel dimension, they are respectively multiplied point-by-point with the spatial information excitation block S energy and spliced into the cross-dimensional attention feature F refined ;

[0028] Step 5.4: Generate the diagnostic decision weight σ through the Sigmoid function, and perform weighted fusion of F refined with σ, and output the state discrimination vector Y containing the fault characterization pred .

[0029] An electronic device according to the present invention includes a memory and a processor, characterized in that the memory is used to store a program for supporting the processor to execute the industrial air conditioner fault diagnosis method, and the processor is configured to execute the program stored in the memory.

[0030] A computer-readable storage medium according to the present invention, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is run by a processor, it executes the steps of the industrial air conditioner fault diagnosis method.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] 1. Traditional methods based on vibration or temperature sensors require invasive modification of the equipment, which is time-consuming and laborious. However, the present invention only needs to collect external audio and judge the types of faults that occur through the sound generated during the operation of the industrial air conditioner, effectively reducing the equipment maintenance cost and downtime.

[0033] 2. The MFCC and IMFCC feature fusion method used in the present invention enhances the feature signal through the logarithmic energy analysis of the Mel spectrum and the flipped Mel spectrum, combined with the first-order difference feature. Compared with the traditional method that only uses MFCC or a single feature, the present invention can capture the differences in audio signals under normal and fault states more comprehensively.

[0034] 3. In the industrial air conditioner fault self-diagnosis network constructed by the present invention, the convolution operation module adopts two convolution methods, namely the standard convolution layer and the depthwise separable convolution layer, and combines the method of the CA attention mechanism, which has not been used in the original audio anomaly detection. The standard convolution layer processes the audio enhanced feature matrix to obtain deep basic features, and the depthwise separable convolution layer decouples the cross-channel features of the deep basic features to obtain lightweight detailed features. The advantage of this structure is that it can fully extract features while significantly reducing the network parameters and computational complexity. The coordinate attention mechanism module further processes the mixed response features, which can accurately locate the key frequency bands related to faults and highlight important feature information. The combination of the network architecture and the attention mechanism of the present invention can more efficiently extract fault-related features from audio signals, improving the accuracy and reliability of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is the overall flowchart of the present invention;

[0036] Figure 2 is the flowchart of the feature extraction step;

[0037] Figure 3 is the schematic diagram of the neural network used;

[0038] Figure 4 is the schematic diagram of the attention mechanism module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] In this embodiment, a method for self-diagnosing industrial air conditioner faults based on audio analysis, as Figure 1 shown, includes the following steps:

[0040] Step S1. The present invention first establishes the audio signal acquisition specification. With the help of existing audio acquisition devices such as mobile phones and recording pens, the user uses the same device to collect the standard audio signal set S std under the normal state of the industrial air conditioner and the abnormal audio signal set S abn under the fault state, and constitutes the original audio data set D raw . To ensure the acquisition quality, the sampling frequency is specifically set to 2500 Hz, so as to ensure the accuracy and clarity of the audio signal.

[0041] Step S2. Locate the original audio data set D raw , and delete the invalid audio segments in the device startup / stop stage to obtain the effective audio signal set D valid . Then, use the db4 wavelet basis for 3-layer decomposition, and process the detailed coefficients through the soft threshold function, and retain the low-frequency approximation coefficients to process the effective audio signal set D valid to obtain the denoised audio signal set Ddenoise .

[0042] Step S3: denoising the audio signal set D according to the normal state and the fault state denoise Classify and add the corresponding state category label L to obtain the labeled audio signal set D label .

[0043] Step S4: feature extraction steps Figure 2 As shown, extract the audio signal set D labe The audio features in the audio enhancement matrix X are obtained enhance :

[0044] Step S4.1: To expand the training data pair D label The audio signal in is segmented into 2s segments to obtain the audio segment set S segment ;

[0045] Step S4.2: For the audio segment set S segment After pre-emphasis processing and framing (frame length 0.1s frame shift 0.05s) processing, each frame of audio signal is subjected to Hamming windowing processing to generate a pre-processed audio signal set S pre ;

[0046] Step S4.3: preprocess the audio signal set S pre Perform fast Fourier transform to obtain the frequency domain signal set S f ,The FFT points are 256 points and the resolution is 9.77 Hz / bin to meet the bandwidth requirements of industrial fault characteristics.

[0047] Step S4.4: Use a set of Mel filters to filter S f The audio signals of different frequency bands are filtered and scaled to obtain S f Mel spectrum of each frame of audio signal in S; use a reversed Mel filter bank to f The audio signals of different frequency bands are filtered and scaled to obtain S f The flipped Mel spectrum of each frame of audio signal in the audio signal is obtained; after respectively calculating the logarithmic energy of the Mel spectrum of each frame of audio signal and the logarithmic energy of the flipped Mel spectrum, the logarithmic spectrum of the Mel spectrum of each frame of audio signal and the logarithmic spectrum of the flipped Mel spectrum are obtained, and then the logarithmic spectrum of the Mel spectrum of each frame of audio signal and the logarithmic spectrum of the flipped Mel spectrum are respectively subjected to discrete cosine transformation, and S is obtained accordingly. f The MFCC coefficients of each frame of audio signal X MFCC and IMFCC coefficient X IMFCC; among them, the Mel filter contains 40 triangular filters with a frequency range of 0 - 1250 Hz.

[0048] Step S4.5: Calculate the first-order difference feature ΔX MFCC of X MFCC and the first-order difference feature ΔX IMFCC of X IMFCC , and splice ΔX MFCC and ΔX IMFCC as well as X MFCC and X IMFCC to obtain the audio enhancement feature matrix X enhance .

[0049] Step S5: Construct an industrial air conditioner fault self-diagnosis network, as Figure 3 shown, which successively includes: a convolution operation module, a hybrid pooling module, and a coordinate attention mechanism module:

[0050] Step 5.1: The convolution operation module processes X enhance using a standard convolution layer to obtain the deep basic feature C base ;

[0051] The convolution operation module uses a depthwise separable convolution layer to perform cross-channel feature decoupling on C base to obtain the lightweight detail feature C detail ;

[0052] The standard convolution layer uses 32 convolution kernels with a kernel size of 3×3, a stride of 1, a padding method of "same", and the activation function is ReLU; the depthwise separable convolution layer first performs depth convolution using 32 convolution kernels with a kernel size of 3×3, a stride of 1, and a padding method of "same"; then it performs pointwise convolution using 64 convolution kernels with a kernel size of 1×1, and the activation function is ReLU.

[0053] Step 5.2: The hybrid pooling module performs maximum pooling operation and average pooling operation on C detail respectively, and correspondingly obtains the kurtosis feature P max representing the significant response and the mean feature P avg representing the distribution characteristics, so as to obtain the hybrid response feature F fusion with bimodal characteristics using Equation (1);

[0054]

[0055] In Equation (1), is a random value, ; In this embodiment, the pooling window of the max pooling layer is 2×2, and the stride is 2; the pooling window of the average pooling layer is 2×2, and the stride is 2.

[0056] Step 5.3: The attention mechanism module is CA as Figure 4 shown, and F is pooled and compressed through global average pooling to obtain the horizontal fusion incentive and the vertical incentive on. After splicing, the spatial energy feature S is obtained ; The spatial information excitation block S is decomposed into the horizontal attention weight W energy through a 1D convolution and non-linear activation energy and the vertical attention weight W horizon . After expanding the cross-channel dimension, they are multiplied element-wise with the spatial information excitation block S vertical and then spliced into the cross-dimensional attention feature F energy ; The 1D convolution uses 16 convolutional kernels, the kernel size is 3, and the stride is 1. refined

[0057] Step 5.4: The diagnostic decision weight σ is generated through the Sigmoid function, and F refined is weighted and fused with σ to output the state discrimination vector Y containing the fault characterization pred ;

[0058] Specifically, in the implementation, the cross-dimensional attention feature F refined is input into the Sigmoid function to obtain the diagnostic decision weight σ. The output range of the Sigmoid function is between 0 and 1, representing the importance weight of each feature. F refined is multiplied element-wise with σ for weighted fusion to obtain the state discrimination vector Y containing the fault characterization pred ;

[0059] Step 5.5: The state discrimination vector Y pred is normalized to generate the diagnostic confidence vector S based on confidence . The diagnostic confidence vector S confidence and the true state label L are used to construct a hybrid loss function. The network parameters are optimized through backpropagation until convergence to obtain the diagnostic model corresponding to the optimal parameter θ*, which is used to realize the multi-dimensional feature analysis and fault mode discrimination of the air conditioner operating state.

[0060] In specific implementation, the Stochastic Gradient Descent (SGD) optimizer is used, with the learning rate set to 0.001 and the momentum parameter set to 0.9. Through the backpropagation algorithm, the gradient is calculated according to the hybrid loss function to update the parameters in the network. During the training process, the dataset is divided into a training set and a validation set in a ratio of 8:2. After each epoch of training, validation is performed on the validation set. When the value of the loss function on the validation set no longer decreases for 10 consecutive epochs, it is determined that the network converges. At this time, the diagnostic model corresponding to the optimal parameter θ* is obtained, and this diagnostic model is used to perform multi-dimensional feature analysis and fault mode discrimination on the operating state of the industrial air conditioner;

[0061] In summary, a self-diagnosis method for industrial air conditioner faults based on time-frequency features according to the present invention combines signal processing and neural network technologies, achieving a complete set of air conditioner fault diagnosis solutions from signal acquisition to signal processing and then to model training. When an air conditioner operation fault occurs, the fault location can be quickly determined, saving emergency repair time, optimizing the stability of the industrial air conditioner during production, and ensuring that the temperature and humidity balance in the production workshop meets the standards.

[0062] In this embodiment, an electronic device includes a memory and a processor. The memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0063] In this embodiment, a computer-readable storage medium stores a computer program, and when the computer program is run by a processor, it executes the steps of the above method.

Claims

1. An industrial air conditioner fault self-diagnosis method based on audio analysis, characterized in that, It includes the following steps: Step S1: Use the same device to collect the standard audio signal set S under the normal state of the industrial air conditioner std and the abnormal audio signal set S under the fault state abn , and form the original audio data set D raw ; Step S2: Locate the original audio dataset D raw and delete the invalid audio segments in the device startup / stop phase to obtain the effective audio signal set D valid After that, use the wavelet denoising algorithm to process the effective audio signal set D valid to obtain the denoised audio signal set D denoise ; Step S3: Classify the denoised audio signal set D according to the normal state and the fault state denoise and add the corresponding state category label L to obtain the labeled audio signal set D label ; Step S4: Extract the audio features from the audio signal set D labe to obtain the audio enhancement feature matrix X enhance : Step S5: Construct an industrial air conditioner fault self-diagnosis network, which successively includes: a convolution operation module, a hybrid pooling module, and a coordinate attention mechanism module, and process X enhance to obtain a state discrimination vector Y containing fault characteristics pred ; Step S6: Normalize the state discrimination vector Y pred to generate a diagnosis confidence vector S confidence . Construct a hybrid loss function from the diagnosis confidence vector S confidence and the true state label L, and optimize the network parameters through backpropagation until convergence to obtain a diagnosis model corresponding to the optimal parameter θ*. This diagnosis model realizes multi-dimensional feature analysis and fault mode discrimination of the air conditioner operating state.

2. The industrial air conditioner fault self-diagnosis method based on audio analysis according to claim 1, characterized in that, Step S4 includes the following steps: Step S4.1: Segment the audio signal in D label to obtain an audio segment set S segment ; Step S4.2: For the audio segment set S segment After performing pre-emphasis processing and framing processing on each one in sequence, perform Hamming windowing on each frame of audio signal, thereby generating a preprocessed audio signal set S pre ; Step S4.3: Perform a fast Fourier transform on the preprocessed audio signal set S pre to obtain a frequency-domain signal set S f ; Step S4.4: Use a set of Mel filter banks to filter and scale-transform the audio signals in different frequency bands in S f to obtain the Mel spectrogram of each frame of audio signal in S f ; Filter and scale-transform the audio signals in different frequency bands of S using an inverted Mel filter bank to obtain the inverted Mel spectrum of each frame of audio signal in S f f ;​ After calculating the logarithmic energy of the Mel spectrum and the logarithmic energy of the flipped Mel spectrum for each frame of the audio signal respectively, the logarithmic spectrum of the Mel spectrum and the logarithmic spectrum of the flipped Mel spectrum of each frame of the audio signal are obtained. Thus, the discrete cosine transform is respectively performed on the logarithmic spectrum of the Mel spectrum and the logarithmic spectrum of the flipped Mel spectrum of each frame of the audio signal, and the MFCC coefficients X of each frame of the audio signal in S f are obtained MFCC and the IMFCC coefficients X IMFCC ; Step S4.5: Calculate the first-order difference feature ΔX of X for each frame of audio signal MFCC and the first-order difference feature ΔX of X MFCC and X IMFCC and the first-order difference feature ΔX of X IMFCC , and after concatenating ΔX MFCC and ΔX IMFCC and X MFCC and X IMFCC , an audio enhancement feature matrix X enhance is obtained.

3. The industrial air conditioner fault self-diagnosis method based on audio analysis according to claim 2, characterized in that, Step S5 includes the following steps: Step 5.1: The convolution operation module processes X using a standard convolution layer enhance to obtain the deep basic feature C base ; The convolution operation module uses a depthwise separable convolutional layer to perform cross-channel feature decoupling on C base to obtain lightweight detail features C detail ; Step 5.2: The hybrid pooling module performs max pooling operation and average pooling operation on C detail respectively, and correspondingly obtains the kurtosis feature P max representing the significant response and the mean feature P avg representing the distribution characteristics, so as to obtain the hybrid response feature F fusion with bimodal characteristics by using Equation (1); (1) In formula (1), is a random value, ; Step 5.3: The attention mechanism module performs pooling compression on F fusion through global average pooling to obtain horizontal incentives and vertical incentives and after splicing, the spatial energy feature S is obtained energy ; The spatial information excitation block S is decomposed into horizontal attention weights W energy and vertical attention weights W horizon through a 1D convolution and non-linear activation, and after extension across the channel dimension, they are multiplied element-wise with the spatial information excitation block S vertical respectively and concatenated into cross-dimensional attention features F energy ; refined ​ Step 5.4: Generate the diagnostic decision weight σ through the Sigmoid function, and perform weighted fusion on F refined and σ, and output the state discrimination vector Y containing the fault characteristics pred .

4. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor to execute any one of the industrial air conditioner fault diagnosis methods recited in claims 1-3, and the processor is configured to execute the program stored in the memory.

5. A computer-readable storage medium, on which a computer program is stored, characterized in that, When the computer program is run by the processor, it executes the steps of any one of the industrial air conditioner fault diagnosis methods recited in claims 1-3.

Images