Method and system for diagnosing faults in vertical sieve-plate tower by using voiceprint detection

Abstract

The invention provides a method and a system for diagnosing faults in a vertical sieve-plate tower by utilizing voiceprint detection, and the method comprises the following steps: collecting voiceprint signal data of the vertical sieve-plate tower, and classifying the data according to a normal state and a fault state; the collected voiceprint signals in different states are preprocessed; extracting time domain, frequency domain and inverted frequency domain features in the preprocessed voiceprint data; fusing an SE module and an ECA module on the basis of a ResNet50D network model, and constructing a detection model called SECA-ResNet50D by modifying a feature map output by a deep-stage convolutional layer in a forward propagation part; voiceprint features extracted from the vertical sieve-plate tower voiceprint signals are used for constructing a data set, and the data set is used as the input of a ResNet50D network in an SECA-ResNet50D detection model; selecting a feature combination suitable for detection; and obtaining a fault state detection result by using the selected feature combination suitable for detection.

Description

Technical Field

[0001] The present invention relates to the technical field of vertical sieve plate tower fault detection, and more specifically to a method and system for diagnosing faults in a vertical sieve plate tower by using sound print detection. Background Art

[0002] The vertical sieve plate tower is an efficient gas-liquid or liquid-liquid separation equipment with the characteristics of large scale, continuity and multivariable. It is widely used in chemical production and separation processes. However, due to the complex balance between momentum, heat and mass transfer, the tower is difficult to operate stably. Liquid-gas two-phase flow plate towers are unstable under normal conditions. If the operating parameters are not adjusted properly, leakage or flooding will occur, which will not only endanger the safety of the processing equipment, but also may cause huge economic losses. Therefore, the development of a system for fault diagnosis in vertical sieve plate towers is crucial to reduce the probability of accidents.

[0003] The common method for fault diagnosis in vertical sieve plate towers is mainly the visual inspection method. The operator determines whether there is a fault by observing whether the tower's pressure, temperature and other parameters are normal, and by listening to the sound. However, these methods are not sensitive and rely heavily on the operator's experience. In addition, there is acoustic emission detection technology, but this is not an ideal method because it requires advanced technical equipment and professional operators, and the technical and labor costs are relatively high. In recent years, voiceprint detection methods have become a new method. Voiceprint signals are one-dimensional data, easy to collect, and have relatively low storage requirements. They also have the advantages of non-invasiveness and pattern recognition potential, which can provide strong support for the stable operation and optimized operation of equipment.

[0004] Although there are some methods that attempt to use complex algorithms and machine learning models for fault diagnosis, these methods usually have shortcomings in processing multi-angle sound data, feature weight optimization, and dynamic parameter adjustment. Therefore, there is an urgent need for a new method that can comprehensively address these problems and improve fault diagnosis accuracy and system adaptability. Summary of the invention

[0005] The present invention provides a method for diagnosing faults in a vertical sieve plate tower using soundprint detection, which can accurately monitor the fault state of the tower and prevent faults from occurring, thereby ensuring the safe operation of the vertical sieve plate tower. In order to achieve the above purpose, the present invention provides a method and system for diagnosing faults in a vertical sieve plate tower using soundprint detection. The technical solution is as follows:

[0006] A method for diagnosing faults in a vertical sieve plate tower using soundprint detection comprises the following steps:

[0007] S1. Collect the vertical sieve plate tower soundprint signal data and classify it according to normal state and fault state;

[0008] S2. Preprocess the collected voiceprint signals in different states;

[0009] S3. Extract the time domain, frequency domain and inverse frequency domain features from the preprocessed voiceprint data, the time domain includes short-time energy features and short-time zero-crossing rate features, the frequency domain includes spectral features, and the inverse frequency domain includes Mel-cepstral coefficient features;

[0010] S4. Based on the ResNet50D network model, the SE module and the ECA module are integrated. The feature map output by the deep convolution layer in the forward propagation part is modified by these two modules to construct a detection model called SECA-ResNet50D. The voiceprint features extracted from the vertical sieve plate tower voiceprint signal are used to construct a data set as the input of the ResNet50D network in the SECA-ResNet50D detection model. The training process of the SECA-ResNet50D detection model includes:

[0011] 1) SE module processing, including Squeeze, Excitatio and Recalibration operations: The data set constructed using voiceprint features is used as the input of the ResNet50D network; the SE module acts on the first feature map generated by the deep convolution layer of the ResNet50 network, and the Squeeze operation compresses the first feature map through global average pooling GAP; in GAP, a convolution kernel with the same size as the spatial size of the first feature map is used to perform a convolution operation on the first feature map, and a global average value is calculated for each channel of the first feature map, thereby obtaining a descriptor for each channel, forming a one-dimensional vector with the same number of channels as the first feature map;

[0012] In the Excitation operation, the obtained one-dimensional vector is nonlinearly transformed through a bottleneck structure consisting of two fully connected layers FC. The first FC layer is used as a dimensionality reduction layer to reduce the number of parameters and computational complexity; the second FC layer is used as a dimensionality increase layer to restore the feature dimension to the same number of input channels; after the second FC layer, the Sigmoid activation function is applied to compress the output value to between 0 and 1 to obtain the importance weight of each channel;

[0013] In the Recalibration operation, the channel weights generated by the SE module are multiplied by the first feature map channel by channel, and the weight of each channel is multiplied by the first feature map corresponding to the channel element by element to achieve channel-level recalibration of the feature map and output the second feature map. Through this operation, the importance of each channel is learned and adjusted, so that the network model can focus on the feature channels that have the greatest impact on the final task.

[0014] 2) ECA module: The output of the SE module is used as the input of the ECA module. The ECA module processes the second feature map output by the SE module using a one-dimensional convolution operation. The ECA module is used to capture the interaction information between local cross-channels and generate a set of channel weights.

[0015] 3) Feature fusion: Multiply a set of channel weights output by the ECA module with the second feature map channel by channel to achieve channel-level recalibration and obtain the third feature map after feature fusion;

[0016] 4) Output feature map: The third feature map after feature fusion will be used as the input of the subsequent convolutional layer;

[0017] 5) Continue to participate in the calculation of the ResNet50 network to perform feature extraction and classification;

[0018] S5. Combine different features and input them into the trained SECA-ResNet50D detection model to perform feature ablation and select the feature combination suitable for detection;

[0019] S6. Use the selected feature combination suitable for detection to obtain the fault state detection result.

[0020] Furthermore, the deep stage convolutional layer includes third and fourth stage convolutional layers.

[0021] Furthermore, in step S1, the fault conditions include at least four types: gas-liquid backmixing, liquid leakage, downcomer liquid flooding, and mist entrained liquid flooding.

[0022] Furthermore, in step S2, the preprocessing specifically includes pre-emphasis, frame division, windowing and noise reduction.

[0023] Furthermore, in step S5, the selected feature combination suitable for detection is a short-time zero-crossing rate feature, a spectral feature, and a Mel-frequency cepstral coefficient feature.

[0024] The present invention also provides a system for diagnosing faults in a vertical sieve plate tower by using soundprint detection to implement the above method, comprising:

[0025] A soundprint signal collection unit is used to collect soundprint signal data of the vertical sieve plate tower and classify them according to normal state and fault state;

[0026] A voiceprint signal preprocessing unit, used to preprocess the collected voiceprint signals in different states;

[0027] A feature extraction unit is used to extract time domain, frequency domain and inverse frequency domain features from the preprocessed voiceprint data, the time domain includes short-time energy features and short-time zero-crossing rate features, the frequency domain includes spectral features, and the inverse frequency domain includes Mel-cepstral coefficient features;

[0028] The SECA-ResNet50D detection model construction and training unit integrates the SE module and the ECA module on the basis of the ResNet50D network model, and uses these two modules to modify the feature map output by the deep convolution layer in the forward propagation part to build a detection model called SECA-ResNet50D. The voiceprint features extracted from the vertical sieve plate tower voiceprint signal are used to construct a data set as the input of the ResNet50D network in the SECA-ResNet50D detection model. The training process of the SECA-ResNet50D detection model includes:

[0029] 1) SE module processing, including Squeeze, Excitatio and Recalibration operations: The data set constructed using voiceprint features is used as the input of the ResNet50D network; the SE module acts on the first feature map generated by the deep convolution layer of the ResNet50 network, and the Squeeze operation compresses the first feature map through global average pooling GAP; in GAP, a convolution kernel with the same size as the spatial size of the first feature map is used to perform a convolution operation on the first feature map, and a global average value is calculated for each channel of the first feature map, thereby obtaining a descriptor for each channel, forming a one-dimensional vector with the same number of channels as the first feature map;

[0030] In the Excitation operation, the obtained one-dimensional vector is nonlinearly transformed through a bottleneck structure consisting of two fully connected layers FC. The first FC layer is used as a dimensionality reduction layer to reduce the number of parameters and computational complexity; the second FC layer is used as a dimensionality increase layer to restore the feature dimension to the same number of input channels; after the second FC layer, the Sigmoid activation function is applied to compress the output value to between 0 and 1 to obtain the importance weight of each channel;

[0031] In the Recalibration operation, the channel weights generated by the SE module are multiplied by the first feature map channel by channel, and the weight of each channel is multiplied by the first feature map corresponding to the channel element by element to achieve channel-level recalibration of the feature map and output the second feature map. Through this operation, the importance of each channel is learned and adjusted, so that the network model can focus on the feature channels that have the greatest impact on the final task.

[0032] 2) ECA module: The output of the SE module is used as the input of the ECA module. The ECA module processes the second feature map output by the SE module using a one-dimensional convolution operation. The ECA module is used to capture the interaction information between local cross-channels and generate a set of channel weights.

[0033] 3) Feature fusion: Multiply a set of channel weights output by the ECA module with the second feature map channel by channel to achieve channel-level recalibration and obtain the third feature map after feature fusion;

[0034] 4) Output feature map: The third feature map after feature fusion will be used as the input of the subsequent convolutional layer;

[0035] 5) Continue to participate in the calculation of the ResNet50 network to perform feature extraction and classification;

[0036] The detection feature combination unit is used to combine different features and input them into the trained SECA-ResNet50D detection model to perform feature ablation and select the feature combination suitable for detection;

[0037] The fault state detection unit uses the selected feature combination suitable for detection to obtain a fault state detection result.

[0038] It can be seen from the above technical solutions that the present invention provides a method and system for diagnosing faults in a vertical sieve plate tower using soundprint detection, which has the following beneficial effects compared with the prior art:

[0039] The present invention extracts multiple features from voiceprint signals in different states to improve the feature types and voiceprint types in the database, thereby making fault detection more complete.

[0040] The present invention improves the detection accuracy by preprocessing the voiceprint signal.

[0041] The present invention improves the relevant algorithms, further improves the speed and accuracy of detection, further improves the functional effect of fault detection, and improves the efficiency of later fault elimination. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of the method for diagnosing faults in a vertical sieve plate tower using soundprint detection provided by the invention

[0043] Figure 2 This is the structure diagram of the voiceprint collection system

[0044] Figure 3 Schematic diagram of extracting multiple features of voiceprint signals

[0045] Figure 4 This is a schematic diagram of the ResNet50D network structure

[0046] Figure 5 This is the schematic diagram of the Squeeze-and-Excitation module.

[0047] Figure 6This is the schematic diagram of the Efficient Channel Attention module

[0048] Figure 7 This is a comparison of the short-term energy features extracted under five conditions.

[0049] Figure 8 This is a comparison bar chart of the results after the feature ablation experiment.

[0050] Fig. 9 This is the accuracy and loss index diagram of the SECA-ResNet50D model after 20 trainings.

[0051] Fig.10 This is the confusion matrix of the SECA-ResNet50D model under STOZCR+MFCC+Spectrogram features DETAILED DESCRIPTION

[0052] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0053] like Figure 1 As shown, a method and system for diagnosing faults in a vertical sieve plate tower using soundprint detection disclosed in an embodiment of the present invention mainly include the following steps:

[0054] (1) Voiceprint data collection. This data is the fault data collected from the vertical sieve plate tower built by Chuangju Technology Company, such as Figure 2 (a) shows that the platform is mainly composed of a motor, a water pump, a sound sensor, and a vertical sieve plate tower.

[0055] The detailed information of the sensor is shown in Table 1. The sound sensor is placed above the vertical sieve plate tower. The five states in the test are as follows: Figure 2 (b)-(f) A total of 300 acoustic signals of five states (downcomer flooding, mist entrainment flooding, gas phase backmixing, leakage, and normal) were recorded. Each signal lasted for 20 seconds and was further divided into 10 shorter signals with a duration of 2 seconds.

[0056] Table 1

[0057]

[0058] (2) Voiceprint signal preprocessing: Preprocessing includes pre-emphasis, framing, windowing, and noise reduction.

[0059] Pre-emphasis: The pre-emphasis function is shown in Equation 1, where x(n) and y(n) are the sampled value and enhanced value of the signal at time n, respectively, and α is the pre-weighting coefficient, which is set to 0.97.26.

[0060] y(n)=x(n)-a·x(n-1) (1)

[0061] Framing: The continuous speech signal is divided into frames according to a certain frame length, and each frame contains 20 milliseconds of signal.

[0062] Windowing: Based on the framing, each frame signal is multiplied by a Hamming window to reduce the sudden change and energy leakage at the frame edge. The Hamming window function ω(n) is shown in Formula 2, where a is usually 0.4627 and L represents the window length.

[0063]

[0064] Noise reduction: A reasonable noise reduction method can reduce the impact of environmental noise on voiceprint signals. The environment in this experiment is a workshop environment, and most of the environmental noise is relatively stable. Therefore, a wavelet function with a "deterministic" threshold selection rule, a "soft threshold" function, and a lower decomposition level can be used to denoise the signal.

[0065] (3) Voiceprint feature extraction: Extract the time domain, frequency domain and inverse frequency domain features from the preprocessed voiceprint data. The time domain includes short-time energy features and short-time zero-crossing rate features, the frequency domain includes spectral features, and the inverse frequency domain includes Mel-frequency cepstral coefficient features.

[0066] The characteristics of time domain analysis are simplicity, low computational cost and strong interpretability. The energy changes under different fault conditions are different. Equation 3 is the definition of short-time energy, where h(n) = ω(n) 2 .

[0067]

[0068] The short-time zero-crossing rate indicates the number of times the waveform signal in a frame of speech crosses the zero value. The short-time zero-crossing rate Zn of the n-th frame speech signal Xn(m) is expressed as Equation 4.

[0069]

[0070] A spectrogram is a graph that shows the change of speech spectrum over time. The vertical axis is frequency and the horizontal axis is time. The strength of any given frequency component at a given moment is represented by the grayscale or hue of the corresponding point.

[0071] First, add windows and frames to become x(m,n) (Formula 6).

[0072] Then perform FFT transformation to obtain X(m,n) and make the periodogram Y(m,n) (Formula 7).

[0073]

[0074] The analysis of MFCC represents the static characteristics of the voiceprint signal and conforms to the auditory characteristics of the human ear. There is a nonlinear relationship between the Mel frequency and the actual frequency, and the mapping relationship is shown in Formula 8, where fmel is the Mel parameter and fs is the sampling frequency.

[0075]

[0076] MFCC is a spectrum characteristic calculated based on the correspondence between two frequencies. The power spectrum is filtered using a 24-channel Mel filter Hm(k) to obtain the Mel power value. The filter transfer function is shown in Formula 9.

[0077]

[0078] Finally, the Mel spectrum is converted to the time domain by discrete cosine transform to obtain the 12th order MFCC. Then, the MFCC can be calculated to obtain 12 first-order difference coefficients (△MFCC) to reflect the dynamic characteristics of the audio characteristics. The difference calculation formula is shown in Formula 10

[0079]

[0080] Where d(n) is the nth first-order difference, c(n+i) represents the order of the (n+i)th cepstral coefficient, and k represents the interval between different frames.

[0081] (4) The SECA-ResNet50D network module is constructed by fusing the SE module and the ECA module based on the ResNet50D network model for training. The features extracted from the voiceprint signal are constructed into a data set as the input of the ResNet50D network model. In ResNet50D, the input of the network first passes through a starting stage (stage0), which usually contains a convolution layer, a batch normalization layer, and an activation function layer for preliminary processing of the input data. Then, the data will enter a series of stages containing residual blocks (stage1 to stage4), each of which contains multiple residual blocks with the same structure. In each residual block, the input data is first processed by two or three convolution layers, then added to the original input through a short-circuit connection, and finally passed through a ReLU activation function to obtain the output. The main difference of the model proposed in this paper is that the SE module and the ECA module are introduced and combined with the ResNet50D model. In ResNet50D, there is a convolution layer in the initial stage, and there are two or three convolution layers in each residual block in the stage of the residual block (stage1 to stage4). The third and fourth stages (stage3 and stage4) play a more important role in the final classification task. The fusion of the SE module and the ECA module after the convolution layer of the third and fourth stages will produce a more significant effect. The present invention adds the fusion SE module and the ECA module to the convolution layer of the third and fourth stages. The SE module and the ECA module mainly deal with the weight problem of the channel in the feature map output by the convolution layer, and perform channel-level recalibration on the feature map. The feature map is a three-dimensional array output by the convolution layer in the convolutional neural network, usually expressed as [height, width, number of channels]. The channel refers to one of the dimensions of the feature map, which is the number of color channels of the image. Each convolution layer uses a set of convolution kernels to perform convolution operations on the input data and generate multiple new feature maps. These feature maps are then used as input to the next convolutional layer, or passed to subsequent layers after operations such as pooling and activation.

[0082] The detailed steps are as follows:

[0083] 1) SE module (Squeeze-and-Excitation) such as Figure 5 The processing is mainly divided into three steps: Squeeze, Excitatio and Recalibration operations.

[0084] First, the dataset constructed using voiceprint features is used as the input of the SECA-ResNet50 network model. After the deeper convolutional layer, the Squeeze operation compresses the feature map through global average pooling (GAP). In GAP, a convolution kernel with the same size as the spatial size of the feature map (that is, the width and height of the convolution kernel are equal to the width and height of the feature map, respectively), and the stride is also set to 1, and the feature map is convolved. Since the size of the convolution kernel is the same as the spatial size of the feature map, and the stride is 1, this operation is actually equivalent to calculating the global average value of each channel of the feature map, thereby obtaining the "descriptor" or "global feature" of each channel. These descriptors form a one-dimensional vector with the same number of channels as the input feature map.

[0085] In the Excitation operation, the one-dimensional vector obtained above is transformed nonlinearly through a bottleneck structure (Bottleneck) containing two fully connected layers (FC). These two FC layers can be regarded as special 1x1 convolution layers. The first FC layer (dimensionality reduction layer) is used to reduce the number of parameters and computational complexity, while the second FC layer (dimensionality increase layer) restores the feature dimension to the same number of input channels. After the second FC layer, the Sigmoid activation function is applied to compress the output value to between 0 and 1 to obtain the importance weight of each channel.

[0086] In the Recalibration operation, the channel weights generated by the SE module are multiplied channel by channel with the original feature map. This operation is a channel-level weighted operation and does not involve convolution operations in the spatial dimension. Specifically, the weight of each channel is multiplied element by element with the feature map corresponding to the channel to achieve channel-level recalibration of the feature map. Through this operation, the model can automatically learn and adjust the importance of each channel, so that the network can focus on the feature channels that have the greatest impact on the final task.

[0087] 2) ECA module (Efficient Channel Attention) Figure 6 deal with:

[0088] The output of the SE module is used as the input of the ECA module. In the ECA module, the size of the one-dimensional convolution kernel is not fixed, but is adaptively selected according to the number of channels of the input feature map. This adaptive strategy is based on an assumption: the importance of local cross-channel interaction information is related to the size of the convolution kernel, and there is a mapping relationship, so that for different numbers of channels, a suitable one-dimensional convolution kernel size can be found to effectively capture this interaction information.

[0089] After determining the size of the one-dimensional convolution kernel, the ECA module processes the output feature map of the SE module using a one-dimensional convolution operation. This one-dimensional convolution operation is performed on the channel dimension, which is equivalent to convolving all channels at each position (spatial position). In this way, the ECA module can capture the interaction information between local cross-channels and generate a set of channel weights.

[0090] 3) Feature fusion, multiply the channel weights output by the ECA module by the original feature map channel by channel to achieve channel-level recalibration. The combination of the SE module and the ECA module with ResNet50D can bring out their respective advantages. The SE module is responsible for recalibrating the feature channels as a whole, while the ECA module performs more detailed local feature adjustments on this basis. This collaborative working method helps the model better focus on the features that have the greatest impact on the final task, while reducing the interference of unimportant features. In this way, the entire system can achieve higher accuracy and better generalization performance when performing fault classification.

[0091] 4) Output feature map: The feature map after feature fusion will be used as the input of the subsequent convolution layer and continue to participate in the network calculation. Since the feature map has been recalibrated at the channel level, the subsequent convolution layer will be able to better use this key information for feature extraction and classification.

[0092] (5) In order to detect which feature is most suitable for the fault detection device, the commonly used indicators in the detection field, such as accuracy, recall rate, and F1, are used for evaluation.

[0093] A total of 12 groups of experiments were conducted. Table 2 shows the experimental results. The most important one is the accuracy. In order to see the relationship between the accuracy more intuitively, a bar chart of the accuracy is drawn, as shown in the figure below. Figure 8 As can be seen from the figure, STAE has the lowest accuracy of 0.933441, while STOZCR+MFCC+Spectrogram has the highest accuracy of 0.996717. The experimental results with STAE are generally low, which shows that STAE features are not as suitable for identifying vertical sieve plate tower faults as the other three features. For a single feature, MFCC has the highest accuracy of 0.995344. This accuracy can basically ensure that equipment failures can be discovered in a timely manner.

[0094] Table 2

[0095]

[0096]

[0097] In the SECA-ResNet50D architecture, we also use the STOZCR+MFCC+Spectrogram with the highest accuracy to illustrate. Fig. 9 The training and verification results of SECA-ResNet50D show that the SECA-ResNet50D network tends to be stable when it iterates to the tenth generation, and its accuracy and loss rate are better than those of the ResNet50D network.

[0098] Fig.10 is the confusion matrix of SECA-ResNet50D. It can be seen from the confusion matrix that the recognition accuracy of SECA-ResNet50D for leakage, normal, gas phase backmixing, mist entrainment liquid flooding, and downcomer liquid flooding are 99.72%, 99.44%, 100%, 99.74% and 99.45% respectively. Among them, 0.28% of leakage was misdiagnosed as downcomer liquid flooding, 0.28% of normal was misdiagnosed as leakage and mist entrainment liquid flooding, 0.26% of mist entrainment liquid flooding was misdiagnosed as downcomer liquid flooding, and 0.27% of downcomer liquid flooding was misdiagnosed as normal and mist entrainment liquid flooding. The results show that the SECA-ResNet50D model has high recognition accuracy and strong stability. The model has high fault recognition accuracy for vertical sieve plate towers.

Claims

1. A method for diagnosing faults in a vertical sieve plate tower using soundprint detection, comprising the following steps: S1. Collect the vertical sieve plate tower soundprint signal data and classify it according to normal state and fault state; S2. Preprocess the collected voiceprint signals in different states; S3. Extract the time domain, frequency domain and inverse frequency domain features from the preprocessed voiceprint data, the time domain includes short-time energy features and short-time zero-crossing rate features, the frequency domain includes spectral features, and the inverse frequency domain includes Mel-cepstral coefficient features; S4. Based on the ResNet50D network model, the SE module and the ECA module are integrated. The feature map output by the deep convolution layer in the forward propagation part is modified by these two modules to construct a detection model called SECA-ResNet50D. The voiceprint features extracted from the vertical sieve plate tower voiceprint signal are used to construct a data set as the input of the ResNet50D network in the SECA-ResNet50D detection model. The training process of the SECA-ResNet50D detection model includes: 1) SE module processing, including Squeeze, Excitatio and Recalibration operations: The data set constructed using voiceprint features is used as the input of the ResNet50D network; the SE module acts on the first feature map generated by the deep convolution layer of the ResNet50 network, and the Squeeze operation compresses the first feature map through global average pooling GAP; in GAP, a convolution kernel with the same size as the spatial size of the first feature map is used to perform a convolution operation on the first feature map, and a global average value is calculated for each channel of the first feature map, thereby obtaining a descriptor for each channel, forming a one-dimensional vector with the same number of channels as the first feature map; In the Excitation operation, the obtained one-dimensional vector is nonlinearly transformed through a bottleneck structure consisting of two fully connected layers FC. The first FC layer is used as a dimensionality reduction layer to reduce the number of parameters and computational complexity; the second FC layer is used as a dimensionality increase layer to restore the feature dimension to the same number of input channels; after the second FC layer, the Sigmoid activation function is applied to compress the output value to between 0 and 1 to obtain the importance weight of each channel; In the Recalibration operation, the channel weights generated by the SE module are multiplied by the first feature map channel by channel, and the weight of each channel is multiplied by the first feature map corresponding to the channel element by element to achieve channel-level recalibration of the feature map and output the second feature map. Through this operation, the importance of each channel is learned and adjusted, so that the network model can focus on the feature channels that have the greatest impact on the final task. 2) ECA module: The output of the SE module is used as the input of the ECA module. The ECA module processes the second feature map output by the SE module using a one-dimensional convolution operation. The ECA module is used to capture the interaction information between local cross-channels and generate a set of channel weights. 3) Feature fusion: Multiply a set of channel weights output by the ECA module with the second feature map channel by channel to achieve channel-level recalibration and obtain the third feature map after feature fusion; 4) Output feature map: The third feature map after feature fusion will be used as the input of the subsequent convolutional layer; 5) Continue to participate in the calculation of the ResNet50 network to perform feature extraction and classification; S5. Combine different features and input them into the trained SECA-ResNet50D detection model to perform feature ablation and select the feature combination suitable for detection; S6. Use the selected feature combination suitable for detection to obtain the fault state detection result.

2. The method for diagnosing faults in a vertical sieve plate tower using soundprint detection according to claim 1 is characterized in that: The deep stage convolutional layers include third and fourth stage convolutional layers.

3. The method for diagnosing faults in a vertical sieve plate tower using soundprint detection according to claim 1 is characterized in that: In step S1, the fault conditions include at least four types: gas-liquid backmixing, liquid leakage, downcomer liquid flooding, and mist entrained liquid flooding.

4. The method for diagnosing faults in a vertical sieve plate tower using soundprint detection according to claim 1, characterized in that: In step S2, the preprocessing specifically includes pre-emphasis, frame division, windowing and noise reduction.

5. The method for diagnosing faults in a vertical sieve plate tower using soundprint detection according to claim 1, characterized in that: In step S5, the selected feature combination suitable for detection is a short-time zero-crossing rate feature, a spectral feature, and a Mel-frequency cepstral coefficient feature.

6. A system for diagnosing faults in a vertical sieve plate tower using soundprint detection, comprising: A soundprint signal collection unit is used to collect soundprint signal data of the vertical sieve plate tower and classify them according to normal state and fault state; A voiceprint signal preprocessing unit, used to preprocess the collected voiceprint signals in different states; A feature extraction unit is used to extract time domain, frequency domain and inverse frequency domain features from the preprocessed voiceprint data, the time domain includes short-time energy features and short-time zero-crossing rate features, the frequency domain includes spectral features, and the inverse frequency domain includes Mel-cepstral coefficient features; The SECA-ResNet50D detection model construction and training unit integrates the SE module and the ECA module on the basis of the ResNet50D network model, and uses these two modules to modify the feature map output by the deep convolution layer in the forward propagation part to build a detection model called SECA-ResNet50D. The voiceprint features extracted from the vertical sieve plate tower voiceprint signal are used to construct a data set as the input of the ResNet50D network in the SECA-ResNet50D detection model. The training process of the SECA-ResNet50D detection model includes: 1) SE module processing, including Squeeze, Excitatio and Recalibration operations: The data set constructed using voiceprint features is used as the input of the ResNet50D network; the SE module acts on the first feature map generated by the deep convolution layer of the ResNet50 network, and the Squeeze operation compresses the first feature map through global average pooling GAP; in GAP, a convolution kernel with the same size as the spatial size of the first feature map is used to perform a convolution operation on the first feature map, and a global average value is calculated for each channel of the first feature map, thereby obtaining a descriptor for each channel, forming a one-dimensional vector with the same number of channels as the first feature map; In the Excitation operation, the obtained one-dimensional vector is nonlinearly transformed through a bottleneck structure consisting of two fully connected layers FC. The first FC layer is used as a dimensionality reduction layer to reduce the number of parameters and computational complexity; the second FC layer is used as a dimensionality increase layer to restore the feature dimension to the same number of input channels; after the second FC layer, the Sigmoid activation function is applied to compress the output value to between 0 and 1 to obtain the importance weight of each channel; In the Recalibration operation, the channel weights generated by the SE module are multiplied by the first feature map channel by channel, and the weight of each channel is multiplied by the first feature map corresponding to the channel element by element to achieve channel-level recalibration of the feature map and output the second feature map. Through this operation, the importance of each channel is learned and adjusted, so that the network model can focus on the feature channels that have the greatest impact on the final task. 2) ECA module: The output of the SE module is used as the input of the ECA module. The ECA module processes the second feature map output by the SE module using a one-dimensional convolution operation. The ECA module is used to capture the interaction information between local cross-channels and generate a set of channel weights. 3) Feature fusion: Multiply a set of channel weights output by the ECA module with the second feature map channel by channel to achieve channel-level recalibration and obtain the third feature map after feature fusion; 4) Output feature map: The third feature map after feature fusion will be used as the input of the subsequent convolutional layer; 5) Continue to participate in the calculation of the ResNet50 network to perform feature extraction and classification; The detection feature combination unit is used to combine different features and input them into the trained SECA-ResNet50D detection model to perform feature ablation and select the feature combination suitable for detection; The fault state detection unit uses the selected feature combination suitable for detection to obtain a fault state detection result.

7. The system for diagnosing faults in a vertical sieve plate tower using soundprint detection according to claim 6 is characterized in that: The deep stage convolutional layers include third and fourth stage convolutional layers.

8. The method for diagnosing faults in a vertical sieve plate tower using soundprint detection according to claim 6, characterized in that: There are at least four types of fault conditions: gas-liquid backmixing, liquid leakage, downcomer flooding, and mist entrainment and liquid flooding.

9. The method for diagnosing faults in a vertical sieve plate tower using soundprint detection according to claim 6, characterized in that: Preprocessing specifically includes pre-emphasis, framing, windowing and noise reduction.

10. The method for diagnosing faults in a vertical sieve plate tower using soundprint detection according to claim 6, characterized in that: The selected feature combination suitable for detection is short-time zero-crossing rate feature, spectral feature and Mel-frequency cepstral coefficient feature.

Images