Sample-extensible water pump fault diagnosis method and system
The sample-scalable water pump fault diagnosis method addresses the challenge of insufficient fault samples by using sound features and a K-nearest neighbor algorithm to identify faults, improving precision and recall rates without retraining, thus enhancing the accuracy and scalability of fault diagnosis.
Patent Information
- Application Number
- CN202411302307.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-18
- Publication Date
- 2024-12-17
AI Technical Summary
Conventional water pump fault diagnosis methods face challenges with insufficient fault samples, especially in early stages or for new equipment, leading to limited model generalization and accurate fault identification, and the difficulty in programmatically screening discriminative features.
A sample-scalable water pump fault diagnosis method that involves obtaining sound clips, forming an initial seed sample set, screening target sound features, extracting feature data, and using a K-nearest neighbor algorithm to identify faults, with the ability to expand the sample set by adding new fault samples and re-screening features as needed.
The method effectively identifies water pump faults with improved precision and recall rates, especially for fault types with fewer samples, without requiring retraining, thus enhancing the accuracy and scalability of fault diagnosis.
Smart Images
Figure CN119150194A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of water pumps and fault diagnosis, and in particular to a water pump fault diagnosis method and system with expandable samples. Background Art
[0002] Water pumps are widely used in industries such as industry, agriculture, and construction. As the use time increases and the working environment changes, water pumps may have various faults, such as impeller wear, bearing damage, seal failure, etc. These faults will not only lead to equipment performance degradation, but may also cause serious economic losses and safety issues. Therefore, timely and accurate diagnosis of water pump faults is crucial to ensure the normal operation of the system.
[0003] At present, the main methods for water pump fault diagnosis are based on water pump detection data diagnosis methods. This method uses the pressure, flow, temperature, etc. of the water pump operation process to judge the working status of the water pump by monitoring the changes in these data; the noise-based fault diagnosis method judges its status by analyzing the sound signals generated during the operation of the water pump; the vibration-based fault diagnosis method detects faults by measuring and analyzing the vibration signals generated during the operation of the water pump. The vibration signal of the water pump can reflect the working status of the equipment, and different fault types will show different characteristics in the vibration spectrum. The above methods collect the operating status data of the water pump from different dimensions, and then extract features and train diagnostic models through machine learning, big data and other methods to realize the fault diagnosis of water pump equipment.
[0004] Conventional automatic monitoring data can hardly detect hidden dangers of equipment in a timely manner, and the fault diagnosis scheme based on vibration sensors and vibration analysis technology is too expensive. In addition, the above methods often face the problem of insufficient number of samples in practical applications. In water pump fault diagnosis, it is often very difficult to obtain fault samples, especially in the early stages or in new equipment, where fault samples are even scarcer. This small number of samples makes it difficult for traditional methods based on big data and deep learning to effectively diagnose faults, because these methods usually rely on a large amount of labeled data to train models. Without sufficient data, the generalization ability of the model will be limited, and it will be difficult to accurately identify unknown faults. In addition, there are many features of audio, but the quality of the features is difficult to confirm. How to programmatically screen more discriminative features is also a major difficulty. Summary of the invention
[0005] In view of the above problems, the present invention provides a sample-expandable water pump fault diagnosis method and system, aiming to solve the problems in the prior art such as the small number of fault samples and the difficulty in supporting the training of fault diagnosis models.
[0006] A sample scalable water pump fault diagnosis method, comprising:
[0007] Step A1, obtaining sound clips of multiple water pumps as samples, and forming an initial seed sample set from the samples and corresponding fault types;
[0008] Step A2, screening the target sound feature type according to the seed sample set, and extracting the sound feature data in the seed sample set according to the screened target sound feature type as the first feature data;
[0009] Step A3, extracting sound feature data from the sound segment to be diagnosed according to the screened target sound feature type as second feature data;
[0010] Step A4, comparing the second characteristic data with the first characteristic data to identify the fault type of the water pump:
[0011] If it is identified that the water pump has no fault, execute step A5;
[0012] If the water pump is identified as faulty, execute step A6;
[0013] Step A5, taking the next sound segment of the water pump that is continuously collected as the sound segment to be diagnosed, and continuing to execute step A3;
[0014] Step A6, add the sound clip corresponding to the second characteristic data as a sample and the identified fault type to the seed sample set to form a new seed sample set, and use the next sound clip of the water pump that is continuously collected as the sound clip to be diagnosed, and continue to execute step A2.
[0015] Further, step A2 includes:
[0016] Step A21, extracting sound feature data of a seed sample set according to the candidate multiple sound feature types as third feature data;
[0017] Step A22, calculating the degree of distinction of each candidate sound feature type according to the third feature data;
[0018] Step A23, selecting a sound feature type with the greatest degree of distinction from the candidate sound feature types as a target sound feature type;
[0019] Step A24, extracting sound feature data in the seed sample set according to the screened target sound feature type as first feature data.
[0020] Furthermore, the seed sample set involves multiple fault types, and each fault type corresponds to multiple samples;
[0021] Step A22 includes:
[0022] Step A221, calculating the average intra-class similarity and the average extra-class similarity of each candidate sound feature type;
[0023] Step A222: For each candidate sound feature type, the ratio of the average intra-class similarity to the average extra-class similarity is used as the discrimination degree of the candidate sound feature type.
[0024] Further, step A4 includes:
[0025] Step A41, calculating the distance between the first feature data and the second feature data of each sample;
[0026] Step A42, selecting a predetermined number of samples as neighbor samples in ascending order of distance;
[0027] Step A43: The fault type with the largest number of neighbor samples is taken as the fault type of the water pump.
[0028] Further, in step A21, the multiple candidate sound feature types include spectrum centroid, spectrum bandwidth, Mel spectrum and Mel frequency cepstral coefficients and chrominance.
[0029] A sample expandable water pump fault diagnosis system, used to execute the aforementioned sample expandable water pump fault diagnosis method, comprising:
[0030] A sample forming module, used to obtain a plurality of water pump sound clips as samples, and form an initial seed sample set from the samples and corresponding fault types;
[0031] A sound collection module, used for continuously collecting sound clips of the water pump during operation to form sound clips to be diagnosed;
[0032] A first feature extraction module, connected to the sample formation module, is used to screen the target sound feature type according to the seed sample set, and extract the sound feature data in the seed sample set according to the screened target sound feature type as the first feature data;
[0033] A second feature extraction module, connected to the sound collection module and the first feature extraction module, respectively, is used to extract sound feature data from the sound segment to be diagnosed according to the screened target sound feature type as the second feature data;
[0034] A fault identification module, connected to the first feature extraction module and the second feature extraction module respectively, for comparing the second feature data with the most recently formed first feature data to identify the fault type of the water pump and obtain a fault identification result;
[0035] The sample forming module is connected to the fault identification module and is further used to: when the fault identification result is that the water pump is faulty, use the sound segment to be diagnosed corresponding to the fault identification result as a new sample, and add the new sample and the identified fault type to the seed sample set to form a new seed sample set;
[0036] The first feature extraction module is connected to the fault identification module and is also used for:
[0037] When the fault identification result is that the water pump is faulty, the target sound feature type is screened again from the new seed sample set, and the sound feature data of the seed sample set is extracted as the first feature data according to the target sound feature type screened again;
[0038] When the fault identification result is that the water pump does not have a fault, the most recently screened target sound feature type and the corresponding first feature data are retained.
[0039] Furthermore, the first feature extraction module includes:
[0040] A first extraction unit is used to extract sound feature data in a seed sample set according to multiple candidate sound feature types as third feature data;
[0041] a discrimination degree calculation unit, connected to the first extraction unit, for calculating the discrimination degree of each candidate sound feature type according to the third feature data;
[0042] The type screening unit is connected to the discrimination calculation unit and is used to screen out the sound feature type with the greatest discrimination from the candidate sound feature types as the target sound feature type.
[0043] Furthermore, the seed sample set involves multiple fault types, and each fault type corresponds to multiple samples;
[0044] The discrimination calculation unit includes:
[0045] A similarity calculation subunit, used to calculate the average intra-class similarity and the average extra-class similarity of each candidate sound feature type;
[0046] The discrimination calculation subunit is connected to the similarity calculation subunit and is used to use the ratio of the average intra-class similarity to the average extra-class similarity as the discrimination of the candidate sound feature type for each candidate sound feature type.
[0047] Furthermore, the fault identification module includes:
[0048] A distance calculation unit, used to calculate the distance between the first feature data and the second feature data of each sample;
[0049] A neighbor selection unit, connected to the distance calculation unit, is used to select a predetermined number of samples in ascending order of distance as neighbor samples;
[0050] The fault determination unit is connected to the nearest neighbor selection unit and is used to select the fault type with the largest number of nearest neighbor samples as the fault type of the water pump.
[0051] Furthermore, the multiple candidate sound feature types include spectral centroid, spectral bandwidth, Mel-spectrogram and Mel-frequency cepstral coefficients, and chrominance.
[0052] The beneficial technical effect of the present invention is that the present invention screens the sound clips of water pump operation through a small number of samples, programmatically selects the most distinguishing sound features, and can scalably collect water pump fault noise samples without assuming the distribution of data or requiring a training process. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1-4 A flow chart of steps of a sample expandable water pump fault diagnosis method of the present invention;
[0054] Figure 5-8 A schematic diagram of a module of a sample expandable water pump fault diagnosis system of the present invention;
[0055] Fig. 9 Mel spectrum feature map of loose blade imbalance based on pump fault type;
[0056] Fig.10 This is the Mel spectrum feature diagram when the pump fault type is bearing fault and blade imbalance;
[0057] Fig.11 This is the Mel spectrum feature diagram when the water pump is operating normally without fault. DETAILED DESCRIPTION
[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0059] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.
[0060] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.
[0061] See also Figure 1 The present invention provides a sample-expandable water pump fault diagnosis method, comprising:
[0062] Step A1, obtaining sound clips of multiple water pumps as samples, and forming an initial seed sample set from the samples and corresponding fault types;
[0063] Step A2, screening the target sound feature type according to the seed sample set, and extracting the sound feature data in the seed sample set according to the screened target sound feature type as the first feature data;
[0064] Step A3, extracting sound feature data from the sound segment to be diagnosed according to the screened target sound feature type as second feature data;
[0065] Step A4, comparing the second characteristic data with the first characteristic data to identify the fault type of the water pump:
[0066] If it is identified that the water pump has no fault, execute step A5;
[0067] If the water pump is identified as faulty, execute step A6;
[0068] Step A5, taking the next sound segment of the water pump that is continuously collected as the sound segment to be diagnosed, and continuing to execute step A3;
[0069] Step A6, add the sound clip corresponding to the second characteristic data as a sample and the identified fault type to the seed sample set to form a new seed sample set, and use the next sound clip of the water pump that is continuously collected as the sound clip to be diagnosed, and continue to execute step A2.
[0070] In step A1, if the sound clip of the water pump has no fault, the fault type is none or empty, and if it has a fault, the fault type is the corresponding fault type. That is, the seed sample set includes several sound clips without faults and sound clips with faults.
[0071] The sound clips automatically added to the seed sample set can also be manually reviewed for fault types, and the seed sample set can be modified if there is an identification error. For example, if the fault type is incorrect, the fault type is changed; if there is no fault, the fault type is changed to no fault or the sample is removed.
[0072] The seed sample set includes sound clips of the water pump running without fault as samples, and the fault type of the samples without fault type is none. In this way, if the largest number of neighboring samples is samples of the water pump without fault type, it is considered that the water pump has no fault.
[0073] The present invention screens the sound clips of water pump operation through a small number of samples, programmatically selects the most distinguishing sound features, and can collect water pump fault noise samples in an extensible manner without assuming the distribution of data and requiring no training process.
[0074] See also Figure 2 , further, step A2 comprises:
[0075] Step A21, extracting sound feature data of a seed sample set according to the candidate multiple sound feature types as third feature data;
[0076] Step A22, calculating the degree of distinction of each candidate sound feature type according to the third feature data;
[0077] Step A23, selecting a sound feature type with the greatest degree of distinction from the candidate sound feature types as a target sound feature type;
[0078] Step A24, extracting sound feature data in the seed sample set according to the screened target sound feature type as first feature data.
[0079] Further, in step A21, the multiple candidate sound feature types include spectrum centroid, spectrum bandwidth, Mel spectrum and Mel frequency cepstral coefficients and chrominance.
[0080] Spectral Centroid: Also known as the "spectral centroid", it is the "center of gravity" position of the sound signal spectrum. It represents the average frequency position of the sound energy and can reflect the brightness of the sound.
[0081] Spectral Bandwidth: It is a way to measure the distribution of spectral energy, which describes the degree of energy dispersion of the sound. It is usually defined as the standard deviation around the centroid of the spectrum, reflecting the spectral width of the sound.
[0082] Mel Spectrogram: It converts the sound signal into a spectrum on the Mel frequency scale. The Mel scale is a nonlinear frequency scale that simulates the human ear's perception of sounds of different frequencies.
[0083] Mel Frequency Cepstral Coefficients (MFCCs): MFCCs is a technology used for sound signal processing based on the nonlinear frequency scale of the human ear. MFCC maps the frequency of the sound signal to the Mel scale and performs a linear cosine transform on it to obtain the Mel frequency cepstral coefficients. These coefficients can better reflect the human ear's perception of sound frequency, so they are widely used in speech recognition, audio compression and other fields.
[0084] Chromagram or Chroma Features: Chroma features are extracted from music signals and represent the periodic pattern of pitch.
[0085] These types of sound features are widely used in audio signal processing, speech recognition, music information retrieval, etc. By extracting these features, audio data can be better understood and analyzed.
[0086] See also Figure 3 ,Furthermore, the seed sample set involves multiple fault types, and each fault type ,corresponds to multiple samples;
[0087] Step A22 includes:
[0088] Step A221, calculating the average intra-class similarity and the average extra-class similarity of each candidate sound feature type;
[0089] Step A222: For each candidate sound feature type, the ratio of the average intra-class similarity to the average extra-class similarity is used as the discrimination degree of the candidate sound feature type.
[0090] Average Within-Class Similarity refers to the average similarity between all samples of the same fault type. A high intra-class similarity means that samples belonging to the same fault type are very similar to each other.
[0091] Average Between-Class Similarity refers to the average of the similarities between samples of different fault types. A lower between-class similarity means that samples of different fault types should be different from each other.
[0092] The higher the discriminability of the noise feature, the lower the performance requirement and dependence on the subsequent diagnostic model. The present invention proposes to use the intra / inter-class similarity ratio to quantitatively evaluate the discriminability of the feature. Assume that the given initial seed sample set is expressed as follows:
[0093] S={(x i ,y i )|y i ∈R M}
[0094] Among them, x i represents the fault sound segment, that is, the sample, y i The corresponding fault type.
[0095] For each candidate feature type m∈M, firstly, a suitable similarity metric is selected, such as Euclidean distance, cosine similarity, Manhattan distance, etc. As a preferred embodiment, the present invention adopts cosine similarity as the similarity metric. For two sound clips, the cosine similarity definition formula is as follows:
[0096]
[0097] In the above formula, f m The first feature data and the second feature data of the present invention are in the form of sound feature vectors. · represents the dot product, and × represents the cross product. First, the similarity within the sound feature type is calculated to measure the closeness between data points within the same candidate sound feature type.
[0098] For each fault type k, assume that there are N k samples, and calculate the average similarity between these sample pairs. The calculation formula is as follows:
[0099]
[0100] The inter-category similarity measures the degree of distinction between data points of different categories. For each pair of different fault types k and l, assuming that there are N k With N l samples, and calculate the average distance between all cross-category sample pairs. The calculation formula is as follows:
[0101]
[0102] The formula for calculating the average intra-class similarity of all fault types is as follows:
[0103]
[0104] The formula for calculating the average out-of-class similarity of all fault types is as follows:
[0105]
[0106] Finally, the discrimination calculation formula of the candidate feature type is as follows:
[0107]
[0108] The larger the discrimination index is, the greater the similarity within the class is than the similarity between classes, and the better the discrimination of the candidate sound feature type is.
[0109] The smaller the discrimination index is, the closer the similarity within the class is to the similarity between the classes, and the less discriminative the candidate sound feature types are.
[0110] For all candidate feature types, find the target feature type m∈M such that:
[0111] m=argmax(distinction m )
[0112] Extract the sound features f of the seed sample set m The first sound feature data is used as a reference for data diagnosis and classification of subsequent K-nearest neighbor algorithm fault identification.
[0113] See also Figure 4 , further, step A4 comprises:
[0114] Step A41, calculating the distance between the first feature data and the second feature data of each sample;
[0115] Step A42, selecting a predetermined number of samples as neighbor samples in ascending order of distance;
[0116] Step A43: The fault type with the largest number of neighbor samples is taken as the fault type of the water pump.
[0117] The seed sample set includes sound clips of the water pump running without fault as samples, and the fault type of the samples without fault type is none. In this way, if the largest number of neighboring samples is samples of the water pump without fault type, it is considered that the water pump has no fault.
[0118] The present invention identifies fault types based on the K-nearest neighbor algorithm. K-nearest neighbor (KNN) is a supervised learning classification algorithm that finds K nearest neighbors in the data, and each sample to be detected can be represented by the category with the largest number of its closest K neighbors.
[0119] For the collected sound clips to be diagnosed x u , extract the sound feature vector f corresponding to the target sound feature type m m (x u ) (i.e., the second feature data), the present invention preferably uses the Euclidean distance to calculate the distance between the sound segment to be diagnosed and the seed sample set. That is, in step A41, the Euclidean distance between the first feature data and the second feature data of each sample is calculated. The calculation formula of the Euclidean distance is as follows:
[0120]
[0121] Calculate the distance between it and all samples in the seed sample set, and select K samples with the smallest distance (K is a predetermined number) as its nearest neighbors. The fault type with the largest number of samples among the K nearest neighbors is taken as the fault type of the water pump sound clip to be diagnosed, that is, the fault type of the water pump. The selection formula is as follows:
[0122]
[0123] Among them, δ(y i ,k) is the indicator function, if yi = k, then it is 1, otherwise it is 0; k is the fault type, y u is the predicted fault type. If the sound segment to be diagnosed is detected and identified as a water pump fault, the sound segment is added as a sample and the identified fault type to the seed sample set to gradually expand the samples in the seed sample set.
[0124] In short, the process of fault type diagnosis and sample expansion of the present invention is roughly as follows:
[0125] The collected few samples of pump fault sound clips and their fault types are used as the subsequent initial seed sample set. 1) The most discriminative feature type in the seed sample set is selected from different sound feature types, namely the target sound feature type. Next, the feature of all sound clips in the seed sample set needs to be extracted as the data point of the K nearest neighbor algorithm; 2) In the application, the sound clips of the water pump in operation are collected in real time, and the discriminative features are extracted as the fragments to be predicted. The fault type of the sound clip to be predicted is determined by the KNN algorithm; 3) If it is a fault, the sound clip and the fault type are added to the seed sample set, and the sound clips are continued to be collected. The new seed sample set is used for the fault type identification process, and corresponding measures are taken for the fault type. If it is not a fault, the sound clips of the water pump in operation are continued to be collected, and the previous seed sample set is used for the fault type identification process.
[0126] The application examples of the present invention are as follows:
[0127] We selected noise clips of pump failures of different powers in a pump room, including three types of faults: normal operation noise of pumps, loose blade imbalance of pumps, and blade imbalance of bearing failures. We selected the first clips in chronological order as the initial seed sample set, and the last clips as the test set. The data distribution is as follows:
[0128] Table 1 Pump noise data distribution
[0129] Initial seed sample set Test Set Normal operation 6 20 Loose pump blades and unbalanced 3 10 Bearing failure Blade imbalance 5 8
[0130] The selected audio features include spectrum centroid, spectrum bandwidth, Mel spectrum, Mel frequency cepstral coefficient, and chromaticity as a set of candidate sound feature types. For the initial seed sample set, the discrimination results of different candidate feature types are calculated as shown in the following table:
[0131] Table 2 Discrimination under different features
[0132]
[0133] As shown in Table 2, for different candidate sound feature types, the higher the average intra-class similarity, the lower the out-of-class similarity, indicating that the candidate sound feature type has a higher degree of discrimination. Among them, the Mel spectrum feature has the highest discrimination index, and the intra-class similarity is more than three times the out-of-class similarity. Although the intra-class similarity of the spectrum centroid and the spectrum bandwidth is high, the out-of-class similarity is also high, with no obvious discrimination. Chroma and Mel spectrum cepstral coefficients also have certain discrimination. Although the intra-class similarity of the Mel spectrum feature is low, the higher degree of discrimination makes it the preferred feature of the present invention, that is, it becomes the target sound feature type. The Mel spectrum feature diagram under different fault types is shown in the figure. Figure 9-11 As shown. Figure 9-11 It can be seen that the characteristic Mel spectrum has discrimination for different fault types, but it can only be distinguished from the perspective of image visualization. The water pump noise discriminative feature screening of the present invention can quantitatively evaluate the discrimination of different sound feature types and whether it can meet the needs of automation.
[0134] In order to verify the effectiveness of the sample scalability method adopted by the present invention, the following comparison schemes are set up: Scheme 1: Always keep the initial seed sample set to screen the target sound feature type, and extract the sound feature data of the target sound feature type in the initial sample set as the data sample in the K nearest neighbor in the subsequent test process. Scheme 2: In the initial state, the initial seed sample set is also used to screen features, and the sound features of the target sound feature type in the initial sample set are extracted as the initial K nearest neighbor data samples. However, if a fault sound segment is detected during the test, the fault sound segment is added to the seed sample set as a fault sample, and the target sound feature type is re-screened, and the sound features of the target sound feature type in the seed sample set are extracted as new K nearest neighbor data samples. The test set water pump fault diagnosis evaluation results under different schemes are as follows:
[0135] Table 3 Water pump fault diagnosis and evaluation results
[0136]
[0137] As shown in Table 3, the pump fault detection method using the sample scalable method has improved the precision p, recall r, and F1 value in different categories, especially for fault types with fewer fault samples, the recall rate has increased by more than 10%, indicating that the scalable solution can timely increase the number of samples in the K nearest neighbor method, making the diagnosis results more accurate and greatly improving the precision and recall rates. In addition, the analysis of the results of each round shows that the more samples are expanded, the more accurate the diagnosis results of the scalable solution. And after the expansion, the solution does not need to be retrained, meeting the real-time requirements.
[0138] The present invention proposes a noise-based, few-sample, scalable water pump fault diagnosis method. Aiming at the problems that the quality of water pump noise features is difficult to evaluate and the features are difficult to screen, the method proposes an intra-class and inter-class similarity ratio to quantitatively calculate the discrimination of different candidate feature types. Aiming at the problems that the water pump fault noise samples are insufficient and the model is difficult to train, a fault diagnosis method using K nearest neighbors is proposed. The method is simple and effective, and meets the needs of sample scalability. It does not require retraining and saves resources. The results show that compared with the non-scalable method, the scalable method can effectively improve the accuracy of fault detection.
[0139] See also Figure 5 The present invention also provides a sample-expandable water pump fault diagnosis system, which is used to execute the aforementioned sample-expandable water pump fault diagnosis method, comprising:
[0140] A sample forming module (1) is used to obtain a plurality of water pump sound clips as samples, and form an initial seed sample set from the samples and corresponding fault types;
[0141] A sound collection module (2) is used to continuously collect sound segments of the water pump during operation to form sound segments to be diagnosed;
[0142] A first feature extraction module (3), connected to the sample formation module (1), is used to screen the target sound feature type according to the seed sample set, and extract sound feature data in the seed sample set according to the screened target sound feature type as the first feature data;
[0143] A second feature extraction module (4) is connected to the sound collection module (2) and the first feature extraction module (3) respectively, and is used to extract sound feature data from the sound segment to be diagnosed according to the screened target sound feature type, as second feature data;
[0144] A fault identification module (5) is connected to the first feature extraction module (3) and the second feature extraction module (4) respectively, and is used to compare the second feature data with the most recently formed first feature data to identify the fault type of the water pump and obtain a fault identification result;
[0145] The sample forming module (1) is connected to the fault identification module (5) and is further used to: when the fault identification result is that the water pump has a fault, use the sound segment to be diagnosed corresponding to the fault identification result as a new sample, and add the new sample and the identified fault type to the seed sample set to form a new seed sample set;
[0146] The first feature extraction module (3), connected to the fault identification module (5), is further used for:
[0147] When the fault identification result is that the water pump is faulty, the target sound feature type is screened again from the new seed sample set, and the sound feature data of the seed sample set is extracted as the first feature data according to the target sound feature type screened again;
[0148] When the fault identification result is that the water pump does not have a fault, the most recently screened target sound feature type and the corresponding first feature data are retained.
[0149] The present invention screens the sound clips of water pump operation through a small number of samples, programmatically selects the most distinguishing sound features, and can collect water pump fault noise samples in an extensible manner without assuming the distribution of data and requiring no training process.
[0150] See also Figure 6 , further, the first feature extraction module (3) comprises:
[0151] A first extraction unit (31) is used to extract sound feature data in a seed sample set according to multiple candidate sound feature types as third feature data;
[0152] A distinguishability calculation unit (32), connected to the first extraction unit (31), for calculating the distinguishability of each candidate sound feature type according to the third feature data;
[0153] The type screening unit (33) is connected to the discrimination calculation unit (32) and is used to screen out the sound feature type with the greatest discrimination from the candidate sound feature types as the target sound feature type.
[0154] Furthermore, the multiple candidate sound feature types include spectral centroid, spectral bandwidth, Mel-spectrogram and Mel-frequency cepstral coefficients, and chrominance.
[0155] See also Figure 7 ,Furthermore, the seed sample set involves multiple fault types, and each fault type ,corresponds to multiple samples;
[0156] The discrimination degree calculation unit (32) comprises:
[0157] A similarity calculation subunit (321), used for calculating the average intra-class similarity and the average extra-class similarity of each candidate sound feature type;
[0158] The discrimination calculation subunit (322) is connected to the similarity calculation subunit (321) and is used to use the ratio of the average intra-class similarity to the average extra-class similarity as the discrimination of the candidate sound feature type for each candidate sound feature type.
[0159] Average Within-Class Similarity refers to the average similarity between all samples of the same fault type. A high intra-class similarity means that samples belonging to the same fault type are very similar to each other.
[0160] Average Between-Class Similarity refers to the average of the similarities between samples of different fault types. A lower between-class similarity means that samples of different fault types should be different from each other.
[0161] See also Figure 8 , further, the fault identification module (5) comprises:
[0162] A distance calculation unit (51), used for calculating the distance between the first feature data and the second feature data of each sample;
[0163] A neighbor selection unit (52), connected to the distance calculation unit (51), is used to select a predetermined number of samples in ascending order of distance as neighbor samples;
[0164] The fault determination unit (53) is connected to the nearest neighbor selection unit (52) and is used to select the fault type with the largest number of nearest neighbor samples as the fault type of the water pump.
[0165] The fault diagnosis method based on K nearest neighbors does not require any assumptions about the data distribution or training process. The samples to be diagnosed are classified into the majority class of their closest K neighbors. This intuitive classification mechanism is easy to understand and implement in the case of few samples.
[0166] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.
Claims
1. A sample-scalable water pump fault diagnosis method, characterized in that: include: Step A1, obtaining sound clips of multiple water pumps as samples, and forming an initial seed sample set from the samples and corresponding fault types; Step A2, screening a target sound feature type according to the seed sample set, and extracting sound feature data in the seed sample set according to the screened target sound feature type as first feature data; Step A3, extracting sound feature data from the sound segment to be diagnosed according to the screened target sound feature type as second feature data; Step A4, comparing the second characteristic data with the first characteristic data to identify the fault type of the water pump: If it is identified that the water pump has no fault, execute step A5; If the water pump is identified as faulty, execute step A6; Step A5, taking the next sound segment of the water pump that is continuously collected as the sound segment to be diagnosed, and continuing to execute step A3; Step A6, add the sound clip corresponding to the second characteristic data as a sample and the identified fault type to the seed sample set to form a new seed sample set, and use the next sound clip of the water pump that is continuously collected as the sound clip to be diagnosed, and continue to execute step A2.
2. A sample expandable water pump fault diagnosis method as claimed in claim 1, characterized in that: The step A2 comprises: Step A21, extracting sound feature data of the seed sample set according to the candidate multiple sound feature types as third feature data; Step A22, calculating the degree of distinction of each candidate sound feature type according to the third feature data; Step A23, selecting a sound feature type with the greatest degree of distinction from the candidate sound feature types as the target sound feature type; Step A24: extracting sound feature data from the seed sample set according to the screened target sound feature type as the first feature data.
3. A sample expandable water pump fault diagnosis method as claimed in claim 2, characterized in that: The seed sample set involves multiple fault types, and each fault type corresponds to multiple samples; The step A22 comprises: Step A221, calculating the average intra-class similarity and the average extra-class similarity of each candidate sound feature type; Step A222: For each candidate sound feature type, the ratio of the average intra-class similarity to the average extra-class similarity is used as the discrimination degree of the candidate sound feature type.
4. A sample expandable water pump fault diagnosis method as claimed in claim 1, characterized in that: The step A4 comprises: Step A41, calculating the distance between the first feature data and the second feature data of each sample; Step A42, selecting a predetermined number of samples as neighbor samples in ascending order of distance; Step A43: taking the fault type with the largest number of neighbor samples as the fault type of the water pump.
5. The method for diagnosing pump faults with expandable samples as claimed in claim 2, characterized in that: In step A21, the multiple candidate sound feature types include spectrum centroid, spectrum bandwidth, Mel spectrum and Mel frequency cepstral coefficients, and chrominance.
6. A sample-expandable water pump fault diagnosis system, characterized in that: A method for diagnosing a pump fault according to any one of claims 1 to 5, comprising: A sample forming module, used to obtain a plurality of water pump sound clips as samples, and form an initial seed sample set from the samples and corresponding fault types; A sound collection module, used for continuously collecting sound clips of the water pump during operation to form sound clips to be diagnosed; A first feature extraction module, connected to the sample formation module, is used to screen a target sound feature type according to the seed sample set, and extract sound feature data in the seed sample set according to the screened target sound feature type as first feature data; a second feature extraction module, connected to the sound collection module and the first feature extraction module respectively, and configured to extract sound feature data from the sound segment to be diagnosed according to the screened target sound feature type as second feature data; a fault identification module, connected to the first feature extraction module and the second feature extraction module respectively, and used to compare the second feature data with the first feature data formed most recently to identify the fault type of the water pump and obtain a fault identification result; The sample forming module is connected to the fault identification module and is further used to: when the fault identification result is that the water pump is faulty, use the sound segment to be diagnosed corresponding to the fault identification result as a new sample, and add the new sample and the identified fault type to the seed sample set to form a new seed sample set; The first feature extraction module is connected to the fault identification module and is further used to: When the fault identification result is that the water pump is faulty, the target sound feature type is screened again from the new seed sample set, and the sound feature data of the seed sample set is extracted as the first feature data according to the target sound feature type screened again; When the fault identification result is that the water pump does not have a fault, the most recently screened target sound feature type and the corresponding first feature data are retained.
7. A sample expandable water pump fault diagnosis system as claimed in claim 6, characterized in that: The first feature extraction module comprises: A first extraction unit, configured to extract sound feature data in the seed sample set according to a plurality of candidate sound feature types as third feature data; a discrimination degree calculation unit, connected to the first extraction unit, and configured to calculate the discrimination degree of each candidate sound feature type according to the third feature data; A type screening unit is connected to the discrimination calculation unit and is used to screen out a sound feature type with the greatest discrimination from the candidate sound feature types as the target sound feature type.
8. The sample expandable water pump fault diagnosis system according to claim 7, characterized in that: The seed sample set involves multiple fault types, and each fault type corresponds to multiple samples; The discrimination calculation unit comprises: A similarity calculation subunit, used to calculate the average intra-class similarity and the average extra-class similarity of each candidate sound feature type; The discrimination calculation subunit is connected to the similarity calculation subunit and is used for taking the ratio of the average intra-class similarity to the average extra-class similarity as the discrimination of the candidate sound feature type for each candidate sound feature type.
9. The sample expandable water pump fault diagnosis system according to claim 6, characterized in that: The fault identification module comprises: a distance calculation unit, configured to calculate the distance between the first feature data and the second feature data of each sample; A neighbor selection unit, connected to the distance calculation unit, for selecting a predetermined number of samples in ascending order of distance as neighbor samples; A fault determination unit is connected to the neighbor selection unit and is used to take the fault type with the largest number of neighbor samples as the fault type of the water pump.
10. The sample expandable water pump fault diagnosis system according to claim 7, characterized in that: The various candidate sound feature types include spectral centroid, spectral bandwidth, Mel-spectrogram and Mel-frequency cepstral coefficients, and chrominance.