Water pump voiceprint feature fusion and data set construction method based on auto-encoder

Abstract

The invention relates to a water pump voiceprint feature fusion and data set construction method based on an auto-encoder, and the method comprises the steps: 1, collecting the voiceprint data of a water pump, obtaining the features of N samples and the corresponding fault types, and enabling the features to comprise image features and numerical features; 2, designing an auto-encoder model which is divided into an image type feature encoder and a numerical value type feature encoder, and reconstructing image type features and numerical value type features; step 3, training the auto-encoder model by using a sample, and evaluating the sum of reconstruction errors of the image-class and numerical-class features by using a loss function; and 4, fusing the image type and numerical value type features by using the trained auto-encoder model to obtain a feature data set, and evaluating the feature data set. The operation state of the water pump is comprehensively captured by fusing multiple features, and the accuracy of fault diagnosis is improved; the self-encoder model is utilized to effectively process voiceprint data collected in a strong noise environment, and the stability and reliability of the model in practical application are ensured.

Description

Technical Field

[0001] The present invention relates to the field of water pump fault diagnosis, and in particular to a water pump voiceprint feature fusion and data set construction method based on an autoencoder. Background Art

[0002] In urban infrastructure, the submersible pump in the tunnel pump room is the core equipment of the drainage system. Its stable operation is crucial to the safety of the tunnel. During the operation of the submersible pump, common fault types include cavitation, foreign body impact, low speed and impeller damage. If these faults are not detected and handled in time, they may weaken the drainage capacity of the pump room. In severe cases, they may even cause the water level in the tunnel to rise, directly threatening driving safety and the integrity of the tunnel structure. Therefore, real-time fault detection and diagnosis of submersible pumps is of great significance.

[0003] Current pump fault detection technologies mostly use contact detection devices such as vibration sensors. Although these sensors can obtain certain vibration characteristics for analyzing the operating status of submersible pumps, their installation and maintenance costs are high, and they are limited by the pump room environment and are difficult to work effectively under strong noise conditions. In addition, the deployment of contact sensors often requires downtime for installation or regular maintenance, which is difficult to meet in some key application scenarios.

[0004] In contrast, the fault diagnosis method based on voiceprint is non-contact and non-invasive, which can avoid problems such as equipment shutdown and sensor contact wear, and is suitable for the complex noise environment of the pump room. The water pump will produce unique voiceprint features during operation. These voiceprint signals contain rich status information and can reflect the health of the water pump. However, the water pump voiceprint signal is usually interfered by complex background noise such as mechanical noise and water flow noise in the pump room. It is not easy to directly extract these features and analyze them. What is more complicated is that the water pump voiceprint features involve multi-dimensional information, including time domain features, frequency domain features, and resonance peak features. It is difficult to accurately identify the faults of submersible pumps by directly analyzing these features.

[0005] Therefore, spectrum analysis and feature engineering are commonly used in existing technologies to separate and extract the features of voiceprint data. However, these methods have great limitations when dealing with the complex acoustic environment and multiple fault modes in pump rooms. How to effectively extract key information from submersible pump voiceprint data under complex noise conditions in order to improve the accuracy of fault diagnosis is still one of the main challenges facing current technology. Summary of the invention

[0006] The present invention provides a method for fusion of water pump voiceprint features and data set construction based on an autoencoder to solve the above technical problems.

[0007] In order to solve the above technical problems, the present invention provides a method for fusion of water pump voiceprint features and data set construction based on autoencoder, comprising the following steps:

[0008] Step 1: Collect water pump voiceprint data to obtain features of N samples and corresponding fault types, where the features include image features and numerical features;

[0009] Step 2: Design an autoencoder model, wherein the autoencoder model includes an image feature encoder and a numerical feature encoder, and reconstructs image features and numerical features;

[0010] Step 3: Using the samples to train the autoencoder model, and using the loss function to evaluate the sum of the reconstruction errors of the image features and the numerical features;

[0011] Step 4: Use the trained autoencoder model to fuse the image features and numerical features, obtain the feature dataset and evaluate the feature dataset.

[0012] Preferably, step 1 also includes preprocessing the collected water pump voiceprint data.

[0013] Preferably, the characteristics of the sample include at least time domain characteristics, frequency domain characteristics, harmonic characteristics and resonance peak characteristics of the water pump.

[0014] Preferably, the fault types include at least a normal state, a cavitation fault, a foreign body impact fault, a low speed fault and an impeller damage fault.

[0015] Preferably, the image-like feature encoder includes a convolution layer, an activation function and a pooling layer in sequence to map the image-like features to a hidden space.

[0016] Preferably, the numerical feature encoder includes a fully connected layer and an activation function to map the numerical features to a hidden space of the same dimension as the image features.

[0017] Preferably, the implicit representation of the image-like features and the implicit representation of the numerical-like features are weightedly fused in the hidden layer to form a comprehensive feature vector.

[0018] Preferably, the comprehensive feature vector is decoded back into a high-dimensional space of original features to generate reconstructed features.

[0019] Preferably, the quality of the feature data set is evaluated based on inter-class difference evaluation, intra-class consistency evaluation and data balance evaluation.

[0020] Preferably, the quality of the feature data set is characterized by a comprehensive indicator Q.

[0021]

[0022] Among them, α and β are weight coefficients, S B is the inter-class scatter matrix, S W is the intra-class scatter matrix, S B / S W It reflects the ratio between inter-class differences and intra-class consistency, and BI is the balance index.

[0023] Compared with the prior art, the method for water pump voiceprint feature fusion and data set construction based on autoencoder provided by the present invention has the following advantages:

[0024] 1. The present invention can comprehensively capture the operating status of the submersible pump by integrating multiple features (such as time domain features, frequency domain features and image features), thereby improving the accuracy of fault diagnosis; compared with a single feature, the use of comprehensive features significantly improves the model's ability to identify different fault types;

[0025] 2. The present invention utilizes the characteristics of the autoencoder model to effectively process voiceprint data collected in a strong noise environment, and ensures the stability and reliability of the model in practical applications through feature reconstruction and noise suppression;

[0026] 3. The autoencoder model in the present invention is trained by unsupervised learning, which can learn useful features from unlabeled data, reduce the dependence on a large amount of labeled data, and reduce the workload and cost of data labeling;

[0027] 4. The data set constructed by the present invention has good structured characteristics through standardized data collection, feature extraction and fusion steps, which is conducive to the subsequent deep learning model training;

[0028] 5. The present invention can monitor the status of the submersible pump in real time, detect the failure of the submersible pump in time and issue an early warning, which greatly improves the safe operation ability of the water pump and reduces potential equipment damage and economic losses;

[0029] 6. The detection method proposed in the present invention can be extended to other types of pumps, mechanical equipment and different fault diagnosis scenarios. It has strong adaptability and flexibility, and promotes the widespread application of voiceprint recognition technology in the industrial field. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flow chart of a method for water pump voiceprint feature fusion and data set construction based on an autoencoder in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to describe the technical solution of the above invention in more detail, specific embodiments are listed below to demonstrate the technical effects; it should be emphasized that these embodiments are used to illustrate the present invention but not to limit the scope of the present invention.

[0032] The method for fusion of water pump voiceprint features and data set construction based on autoencoder provided by the present invention is as follows: Figure 1 As shown, the following steps are included:

[0033] Step 1: Collect water pump voiceprint data to obtain features of N samples and corresponding fault types, where the features include image features and numerical features;

[0034] Step 2: Design an autoencoder model, wherein the autoencoder model includes an image feature encoder and a numerical feature encoder, and reconstructs image features and numerical features;

[0035] Step 3: Using the samples to train the autoencoder model, and using the loss function to evaluate the sum of the reconstruction errors of the image features and the numerical features;

[0036] Step 4: Use the trained autoencoder model to fuse the image features and numerical features, obtain the feature dataset and evaluate the feature dataset.

[0037] By integrating multiple features (such as time domain features, frequency domain features and image features), the present invention can comprehensively capture the operating status of the submersible pump and realize efficient, accurate and real-time diagnosis of submersible pump faults. Compared with a single feature, the use of comprehensive features significantly improves the model's ability to recognize different fault types. At the same time, the present invention utilizes the characteristics of the autoencoder model to effectively process voiceprint data collected in a strong noise environment, and ensures the stability and reliability of the model in practical applications through feature reconstruction and noise suppression.

[0038] In some embodiments, step 1 is used to collect water pump voiceprint data. Specifically, the voiceprint data of the submersible pump can be collected through a variety of sensors. The time domain features (such as short-time energy, short-time zero-crossing rate, etc.), frequency domain features (such as short-time Fourier transform, power spectral density, etc.) and image features (such as spectrogram, spectrum graph) of the water pump are extracted. In some embodiments, step 1 also includes preprocessing the collected water pump voiceprint data to remove noise and irrelevant information to obtain a clear voiceprint data feature input.

[0039] Assume that the collected data set D contains N samples, each sample x i Corresponding to the voiceprint data of a submersible pump and its fault category y i The dataset can be formally expressed as formula (1).

[0040] D={(x i ,y i )|i=1,2,...,N} (1)

[0041] Among them, x i Represents the feature vector of the i-th sample, which is composed of time domain, frequency domain and other high-order features, y i Indicates the category label to which the i-th sample belongs.

[0042] y i ∈{y normal ,y cavitation ,y impact ,y low-speed ,y impeller-damage} (2)

[0043] The corresponding fault types include at least: normal state, cavitation, foreign body impact, low speed, impeller damage, etc.

[0044] The feature matrix X contains the feature combinations of all samples and can be expressed as formula (3).

[0045] X = [x1, x2, ..., x N ] T (3)

[0046] Among them, each sample x i The feature vector of is composed of time domain features, frequency domain features, harmonic features and resonance peak features, and is in the form of formula (4).

[0047] x i =[x i，time , x i，freq , x i，harmonic , x i，formant ,…] (4)

[0048] Class label y i It represents the fault type of each sample, which is defined as formula (5).

[0049]

[0050] The category label y in the data set consists of K categories, which represent the normal and faulty states of the water pump, respectively. In this embodiment, K=5. Of course, each parameter can be modified or expanded according to the actual application, so that the detection method proposed in the present invention can be extended to other types of pumps, mechanical equipment and different fault diagnosis scenarios, with strong adaptability and flexibility, which promotes the widespread application of voiceprint recognition technology in the industrial field.

[0051] In some embodiments, step 2 is used for the design and feature fusion of the autoencoder model. The water pump voiceprint features include image features and numerical features, and two sub-networks are designed to process image features and numerical features respectively. Step 2 specifically includes:

[0052] Step 2.1: Construct an encoder for image-like features. Use multiple convolutional layers to extract features layer by layer, reduce feature dimensions through pooling layers, and finally map image-like features to low-dimensional latent representations. The typical structure includes a combination of convolutional layer + activation function + pooling layer.

[0053] The image-like features are extracted through convolution operation, as shown in formula (6).

[0054]

[0055] Among them, x i，j is the pixel value of the input image, w k，l is the convolution kernel weight, σ is the activation function, and b is the bias.

[0056] Step 2.2: Construct an encoder for numerical features. Use a fully connected layer to process numerical inputs. Use multiple layers of linear transformations and nonlinear activation functions to reduce the dimension of numerical features and compress them to an implicit representation of the same dimension as image features. Numerical features are compressed by linear transformations of the fully connected layer, as shown in Equation (7).

[0057] z num =σ(W num ·x num +b num ) (7)

[0058] Among them, W num is the weight matrix, x num is a numerical feature, b num is the bias and σ is the activation function.

[0059] Step 2.3: Feature fusion. After the image features and numerical features are processed by their respective encoders, they are compressed into the latent space of the same dimension. In the latent layer, the latent representation of the image features and the latent representation of the numerical features are weighted and fused to form a comprehensive feature vector, as shown in formula (8).

[0060] z fused =αz img +(1-α)z num (8)

[0061] Among them, α is a hyperparameter that controls the fusion ratio of image features and numerical features.

[0062] Step 2.4: Decoder design. The decoder is responsible for decoding the fused implicit representation back to the high-dimensional space of the original features and generating reconstructed features. The structure of the decoder is similar to that of the encoder, but the operation is opposite, reconstructing the original image features and numerical features respectively.

[0063] The decoding of image features gradually reconstructs the low-dimensional implicit representation into the original image-like features (such as spectrograms, etc.) through transposed convolution layers or up-sampling.

[0064] The decoding of numerical features reconstructs the implicit representation into the original numerical features layer by layer through the fully connected layer.

[0065] The data set constructed by the present invention through standardized data collection, feature extraction and fusion steps has good structured characteristics, which is conducive to the subsequent deep learning model training.

[0066] In some embodiments, step 3 is used for joint training of the autoencoder. The sum of the reconstruction errors of the image features and the numerical features is evaluated using a loss function to optimize the autoencoder model. The loss function of the autoencoder considers the reconstruction errors of both the image and numerical features at the same time, and usually optimizes the overall model by weighted summing the reconstruction losses of the two features. The total loss function can be expressed as formula (9).

[0067]

[0068] Where: L img is the reconstruction loss of image features (such as mean square error MSE); L num is the reconstruction loss of numerical features; λ1 and λ2 are weight coefficients that balance the reconstruction errors of images and numerical features.

[0069] Through joint optimization, it is possible to ensure that the autoencoder retains the key information of both types of features while fusing them.

[0070] The autoencoder model in the present invention is trained by unsupervised learning, and can learn useful features from unlabeled data, thereby reducing the dependence on a large amount of labeled data and reducing the workload and cost of data labeling.

[0071] In some embodiments, step 4 is used for data set construction and quality assessment, and specifically includes:

[0072] Step 4.1: Dataset construction. The image features (such as spectrograms and spectrograms) and numerical features (such as time domain and frequency domain features) are fused through the autoencoder to obtain the implicit feature representation z i , the implicit eigenvector z i The image and numerical features are combined and expressed as formula (10).

[0073] z i =AutoEncoder(x i,image , x i，numeric ) (10)

[0074] Among them, x i，image is the image feature, x i，numeric It is a numerical feature.

[0075] Finally, the dataset after feature fusion is formula (11).

[0076]

[0077] Among them, z i is the fused feature representation, y i is the corresponding category label.

[0078] Step 4.2: Build a data quality evaluation system. After completing the construction of the water pump voiceprint feature dataset based on the autoencoder, the dataset needs to be further evaluated to ensure that it has good performance in terms of discrimination, consistency and category balance between different categories.

[0079] The present invention proposes a data quality assessment system based on inter-class difference assessment, intra-class consistency assessment and data balance assessment, comprising the following steps:

[0080] Step 4.2.1: Inter-class difference evaluation The purpose of inter-class difference evaluation is to measure the distribution differences of different categories in the feature space to ensure that there is sufficient distinction between different categories.

[0081] For each category C k , calculate the characteristic mean μ of its sample k , the formula is (12).

[0082]

[0083] Among them, N k is the number of samples of category k, z i is the implicit feature representation of the i-th sample.

[0084] Calculate the feature mean distance between different categories, formula is (13).

[0085]

[0086] The larger the mean distance between classes, the stronger the difference between classes.

[0087] Calculate the inter-class scatter matrix S B , which is used to quantify the difference in category mean vectors, as shown in (14).

[0088]

[0089] Where μ is the overall mean of all samples, as shown in (15).

[0090]

[0091] Step 4.2.2: Intra-class consistency assessment.

[0092] Intra-class consistency evaluation is used to measure the concentration of samples within the same category in the feature space to ensure consistency within the category.

[0093] For each category C k , calculate the intra-class variance The formula is (16).

[0094]

[0095] The smaller the intra-class variance, the more concentrated the distribution of samples within the class in the feature space.

[0096] Calculate the intra-class scatter matrix S W , which is used to quantify the distribution of samples within a class. The formula is (17).

[0097]

[0098] S W The smaller it is, the higher the consistency within the class.

[0099] Calculate the average distance d within the class intra-class , the formula is (18).

[0100]

[0101] The smaller the intra-class average distance is, the better the intra-class consistency of the category is.

[0102] Step 4.2.3: Data balance assessment.

[0103] Data balance assessment is used to ensure that samples of each category in the data set are evenly distributed, avoiding the impact of category imbalance on the results during model training.

[0104] Count each category C kThe number of samples N k , calculate the proportion of categories in the total samples P k , the formula is (19).

[0105]

[0106] Where N is the total number of samples in the dataset.

[0107] The class imbalance coefficient is calculated to evaluate the balance of the dataset based on the variance of the number of samples, as shown in formula (20).

[0108]

[0109] in is the average number of samples in each category.

[0110] The balance index BI is defined to measure the balance of the distribution of category samples, and the formula is (21).

[0111]

[0112] Among them, N min and N max are the number of samples of the category with the least and the most samples respectively.

[0113] Step 4.2.4: Comprehensive data quality assessment.

[0114] Through the above steps, the overall quality of the dataset is comprehensively evaluated by combining inter-class differences, intra-class consistency and data balance. Finally, a comprehensive indicator Q can be used to characterize the quality of the dataset, as shown in formula (22):

[0115]

[0116] Among them, α and β are weight coefficients, S B / S W Reflecting the ratio between inter-class differences and intra-class consistency, BI evaluates the balance of a data set. The larger the comprehensive index Q, the higher the quality of the data set.

[0117] The present invention can monitor the status of the submersible pump in real time, promptly discover the failure of the submersible pump and issue an early warning, which greatly improves the safe operation capability of the water pump and reduces potential equipment damage and economic losses.

[0118] In summary, the method for water pump voiceprint feature fusion and data set construction based on autoencoder provided by the present invention includes the following steps: Step 1: Collect water pump voiceprint data to obtain features of N samples and corresponding fault types, wherein the features include image features and numerical features; Step 2: Design an autoencoder model, wherein the autoencoder model includes an image feature encoder and a numerical feature encoder, and reconstructs image features and numerical features; Step 3: Use samples to train the autoencoder model, and use a loss function to evaluate the sum of reconstruction errors of image features and numerical features; Step 4: Use the trained autoencoder model to fuse image features and numerical features to obtain a feature data set and evaluate the feature data set. By integrating multiple features (such as time domain features, frequency domain features and image features), the present invention can comprehensively capture the operating status of the submersible pump, thereby improving the accuracy of fault diagnosis; compared with a single feature, the use of comprehensive features significantly improves the model's ability to recognize different fault types; at the same time, the present invention utilizes the characteristics of the autoencoder model to effectively process voiceprint data collected in a strong noise environment, and through feature reconstruction and noise suppression, ensures the stability and reliability of the model in practical applications.

[0119] Obviously, those skilled in the art can make various changes and modifications to the invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the invention fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A method for fusion of water pump voiceprint features and data set construction based on autoencoder, characterized in that: The steps include: Step 1: Collect water pump voiceprint data to obtain features of N samples and corresponding fault types, where the features include image features and numerical features; Step 2: Design an autoencoder model, wherein the autoencoder model includes an image feature encoder and a numerical feature encoder, and reconstructs image features and numerical features; Step 3: Using the samples to train the autoencoder model, and using the loss function to evaluate the sum of the reconstruction errors of the image features and the numerical features; Step 4: Use the trained autoencoder model to fuse the image features and numerical features, obtain the feature dataset and evaluate the feature dataset.

2. The method for fusion of water pump voiceprint features and construction of data set based on autoencoder according to claim 1, characterized in that: Step 1 also includes preprocessing the collected water pump voiceprint data.

3. The method for fusion of water pump voiceprint features and construction of data set based on autoencoder according to claim 1, characterized in that: The characteristics of the sample include at least time domain characteristics, frequency domain characteristics, harmonic characteristics and resonance peak characteristics of the water pump.

4. The method for fusion of water pump voiceprint features and construction of data set based on autoencoder according to claim 1, characterized in that: The fault types include at least a normal state, a cavitation fault, a foreign body impact fault, a low speed fault and an impeller damage fault.

5. The method for fusion of water pump voiceprint features and construction of data set based on autoencoder according to claim 1, characterized in that: The image-like feature encoder includes a convolution layer, an activation function and a pooling layer in sequence, and maps the image-like features to a hidden space.

6. The method for fusion of water pump voiceprint features and construction of data set based on autoencoder according to claim 5, characterized in that: The numerical feature encoder includes a fully connected layer and an activation function, which maps the numerical features to a hidden space with the same dimension as the image features.

7. The method for fusion of water pump voiceprint features and construction of data set based on autoencoder according to claim 6, characterized in that: The implicit representation of image features and the implicit representation of numerical features are weighted and fused in the hidden layer to form a comprehensive feature vector.

8. The method for fusion of water pump voiceprint features and construction of data set based on autoencoder according to claim 7, characterized in that: The comprehensive feature vector is decoded back to the high-dimensional space of the original feature to generate a reconstructed feature.

9. The method for fusion of water pump voiceprint features and construction of data set based on autoencoder according to claim 1, characterized in that: The quality of the feature dataset is evaluated based on inter-class difference evaluation, intra-class consistency evaluation and data balance evaluation.

10. The method for fusion of water pump voiceprint features and construction of data set based on autoencoder according to claim 9, characterized in that: The quality of the feature dataset is characterized by the comprehensive indicator Q. Among them, α and β are weight coefficients, S B is the inter-class scatter matrix, S W is the intra-class scatter matrix, S B / S W It reflects the ratio between inter-class differences and intra-class consistency, and BI is the balance index.

Images