Induction motor fault detection model construction method, fault detection method and equipment

By constructing an initial neural network and performing clustering processing, an induction motor fault detection model is built, which solves the problem of low fault identification efficiency in existing technologies and achieves fast and accurate fault detection.

CN118916776APending Publication Date: 2024-11-08YANGJIANG NUCLEAR POWER +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410888696.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-03
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively construct fault detection models for induction motors, resulting in low fault identification efficiency.

Method used

A fault detection model for induction motors is constructed by building an initial neural network to obtain induction data of fault categories, building an initial discriminator, performing clustering processing, and finally constructing the fault detection model.

Benefits of technology

It enables rapid identification of induction motor faults, improving the accuracy and efficiency of fault detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916776A_ABST
    Figure CN118916776A_ABST
Patent Text Reader

Abstract

The invention relates to an induction motor fault detection model construction method, a fault detection method and equipment. The method comprises the following steps: S11, constructing an initial neural network, and obtaining corresponding induction data as first sample data based on the fault category of the induction motor; s12, constructing an initial discriminator corresponding to the fault category in the initial neural network according to the first sample data, and obtaining a node address corresponding to the initial discriminator and node data corresponding to the node address; s13, clustering is carried out on the plurality of initial discriminators corresponding to the fault type, a clustering discriminator corresponding to the fault type is obtained according to a clustering result, and the clustering discriminator comprises clustered node addresses and node data; and S14, according to the clustering discriminator and the initial neural network, constructing a final fault detection model. According to the invention, the fault of the motor can be rapidly identified.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of motor detection technology, and more specifically, to an induction motor fault detection model construction method, a fault detection method and a device. Background Art

[0002] Industrial motors are subject to various faults during operation, which are caused by complete or partial damage to system components. The fault behavior of the machine is usually manifested through sensor data, such as current, vibration, and acoustic signals. The first step in developing a fault diagnosis system is to obtain prior knowledge of system performance and fault behavior. Model-based methods represent system knowledge and fault behavior through explicit mathematical models, while data-driven methods implicitly imply this knowledge information. How to detect and identify faults in industrial motors has become a relatively important topic. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide an induction motor fault detection model construction method, a fault detection method and a device.

[0004] The technical solution adopted by the present invention to solve the technical problem is: constructing a method for building an induction motor fault detection model, the method comprising the following steps:

[0005] S11, constructing an initial neural network, and acquiring corresponding induction data as first sample data based on the fault type of the induction motor;

[0006] S12, constructing an initial discriminator corresponding to the fault category in the initial neural network according to the first sample data, and obtaining a node address corresponding to the initial discriminator and node data corresponding to the node address;

[0007] S13, clustering a number of initial discriminators corresponding to the fault category, and obtaining a cluster discriminator corresponding to the fault category according to the clustering result, wherein the cluster discriminator includes a clustered node address and node data;

[0008] S14. Constructing a final fault detection model based on the cluster discriminator and the initial neural network.

[0009] Preferably, in the induction motor fault detection model construction method of the present invention, in the step S12, the initial discriminator corresponding to the fault category in the initial neural network is constructed according to the first sample data, and a node address corresponding to the initial discriminator and node data corresponding to the node address are obtained; including:

[0010] S121, preprocessing and feature extraction of the first sample data to obtain a multi-dimensional real-valued vector corresponding to the first sample data;

[0011] S122, normalizing the multidimensional real-valued vector within an integer range, and performing binary conversion to obtain a corresponding multidimensional binary array;

[0012] S123, making the length of the multi-dimensional binary array a preset length by bit filling, and splitting the multi-dimensional binary array based on the number of nodes to obtain a plurality of data strings, wherein the length of the data string corresponds to the node address of the discriminator;

[0013] S124. Obtain a node address corresponding to the identifier and node data corresponding to the node address according to the content of the data string.

[0014] Preferably, in the induction motor fault detection model construction method described in the present invention, in the step S123, the multidimensional binary array is split based on the number of nodes to obtain a number of data strings, including: updating the node data to update the length of the data string, so as to train the node address corresponding to the discriminator and the node data corresponding to the node address to obtain the final node address and node data corresponding to the discriminator.

[0015] Preferably, in the induction motor fault detection model construction method of the present invention, in the step S122, the multidimensional real-valued vector is normalized within an integer range and binary converted to obtain a corresponding multidimensional binary array, including:

[0016] The decimal value obtained by normalization is converted into binary characters, and binary reflected Gray code representation is performed to generate a corresponding multi-dimensional binary array.

[0017] Preferably, in the induction motor fault detection model construction method of the present invention, in the step S13, clustering the several initial discriminators corresponding to the fault category and obtaining the cluster discriminator corresponding to the fault category according to the clustering result includes:

[0018] S131, obtaining the node address cluster center of the discriminator corresponding to each fault category, and obtaining the node data cluster center corresponding to the node address cluster center according to the node address cluster center, so as to obtain the clustered node address and node data.

[0019] Preferably, in the method for constructing an induction motor fault detection model according to the present invention, the method further comprises:

[0020] S15, acquiring corresponding induction data based on the fault type of the induction motor as second sample data, wherein the second sample data is not repeated with the first sample data;

[0021] S16. Acquire a cluster identifier corresponding to the fault category based on the second sample data, and re-cluster the node address and node data of the cluster identifier to obtain an updated cluster identifier.

[0022] Preferably, in the method for constructing an induction motor fault detection model according to the present invention, in the step S11, the fault category of the induction motor includes: one or more of broken rotor bars, bearing faults, unbalanced loads and broken cage rings; and / or

[0023] In the step S11, the sensing data includes current data and vibration signal data.

[0024] The present invention also constructs an induction motor fault detection model construction device, the device comprising:

[0025] A first processing module is used to construct an initial neural network and obtain corresponding induction data as first sample data based on the fault category of the induction motor;

[0026] A second processing module is used to construct an initial discriminator corresponding to the fault category in the initial neural network according to the first sample data, and obtain a node address corresponding to the initial discriminator and node data corresponding to the node address;

[0027] A third processing module, used for clustering a number of initial discriminators corresponding to the fault category, and obtaining a cluster discriminator corresponding to the fault category according to the clustering result, wherein the cluster discriminator includes a clustered node address and node data;

[0028] The fourth processing module is used to construct a final fault detection model according to the cluster discriminator and the initial neural network.

[0029] The present invention also constructs an induction motor fault detection method, comprising the following steps:

[0030] Acquire induction data of induction motor;

[0031] Acquire the sensing data to acquire a corresponding node address and node data corresponding to the node address;

[0032] Based on a preset fault detection model, a corresponding discriminator is obtained according to the node address and node data corresponding to the sensing data, and a detection result is output through the preset fault detection model;

[0033] Wherein, the preset fault detection model is obtained through the induction motor fault detection model construction method described above.

[0034] The present invention also constructs an induction motor fault detection device, comprising:

[0035] An input module for acquiring induction data of the induction motor;

[0036] A data processing module, used for acquiring the node address corresponding to the sensing data and the node data corresponding to the node address;

[0037] A result generating module, used for obtaining a corresponding discriminator according to the node address and node data corresponding to the sensing data based on a preset fault detection model, and outputting a detection result through the preset fault detection model;

[0038] Wherein, the preset fault detection model is obtained through the induction motor fault detection model construction method described above.

[0039] The implementation of the induction motor fault detection model construction method, fault detection method and device of the present invention has the following beneficial effects: the motor fault can be quickly identified. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0041] Figure 1 It is a program flow chart of an embodiment of a method for constructing an induction motor fault detection model according to the present invention;

[0042] Figure 2 is a program flow chart of another embodiment of a method for constructing an induction motor fault detection model according to the present invention;

[0043] Figure 3 is a program flow chart of another embodiment of a method for constructing an induction motor fault detection model according to the present invention;

[0044] Figure 4 It is a logic block diagram of an embodiment of an induction motor fault detection model building device of the present invention;

[0045] Figure 5 is a program flow chart of an embodiment of an induction motor fault detection method of the present invention;

[0046] Figure 6 It is a logic block diagram of an embodiment of an induction motor fault detection device of the present invention;

[0047] Figure 7 It is a partial logic relationship diagram of an embodiment of a method for constructing an induction motor fault detection model according to the present invention;

[0048] Figure 8 It is a partial logic relationship diagram of another embodiment of a method for constructing an induction motor fault detection model of the present invention;

[0049] Fig. 9 It is a partial logic relationship diagram of another embodiment of a method for constructing an induction motor fault detection model of the present invention. DETAILED DESCRIPTION

[0050] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0051] like Figure 1 As shown, an embodiment of the method for constructing an induction motor fault detection model of the present invention is shown. Figure 1 In one embodiment of the method for constructing an induction motor fault detection model of the present invention, the method includes: S11, constructing an initial neural network, and obtaining corresponding induction data as first sample data based on the fault category of the induction motor; S12, constructing an initial discriminator corresponding to the fault category in the initial neural network according to the first sample data, and obtaining a node address corresponding to the initial discriminator and node data corresponding to the node address; S13, clustering several initial discriminators corresponding to the fault category, and obtaining a cluster discriminator corresponding to the fault category according to the clustering result, wherein the cluster discriminator includes the node address and node data after clustering; S14, constructing a final fault detection model according to the cluster discriminator and the initial neural network.

[0052] That is, based on the above process, a fault detection model for induction motor fault detection and identification can be obtained. In step S11, an initial neural network is first built. The initial neural network adopts an unweighted neural network, in which a discriminator, also called a classifier, is set, which can handle multi-class identification with real-valued input features. At the same time, the corresponding induction data is obtained according to the fault category of the induction motor, and the obtained induction data is used as the first sample data. In one embodiment, the fault category of the induction motor includes: one or more of broken rotor bars, bearing faults, unbalanced loads, and broken squirrel cage rings. In another embodiment, in step S11, the induction data includes current data and vibration signal data.

[0053] Based on step S12, the initial discriminator is constructed by using the acquired first sample data, mainly by constructing the node address of the initial discriminator and the node data corresponding to the node address to obtain the initial discriminator. The node address and node data of the initial discriminator can establish a relationship according to a preset function, and the node data corresponding to the node address is obtained by updating the preset function parameters through the first sample data. Each discriminator has a corresponding node address and node data. It can be understood that there can be multiple initial discriminators corresponding to each fault category, and each initial discriminator contains multiple nodes, and each node has a corresponding node address and node data. Among them, the node can be understood as a network node, that is, a RAM node. The node address is used to confirm the address to be accessed in the discriminator node. The process of obtaining node data based on the preset function is as follows: Figure 8 For example, every 5 node subaddresses obtained are a substring, and the substring corresponds to the storage content of a RAM node. The length of the RAM node is 6 bits, and the storage content is determined by the content of the substring. It is worth noting that the length of the node subaddress substring (5 bits) is not necessarily consistent with the length of the RAM node (6 bits). The storage content in the RAM is a mapping of the storage content of the node subaddress. The decision rule is determined by setting the number of bits of the substring and the number of bits of the RAM node, and is determined by a preset built-in function. Through continuous adjustment, it is found that the best effect is when the substring length is 5 bits and the RAM node length is 6 bits. During the training process, the RAM node content at each position will be modified according to the specific subaddress.

[0054] Based on step S13, the corresponding initial discriminator is obtained for the fault category and clustered, wherein the clustering process is to cluster the node addresses and node data of multiple initial discriminators, and the final cluster discriminator is composed of the clustered node addresses and node data. That is, each fault category should correspond to only one discriminator, but in the above process, each fault category may have multiple different initial discriminators. Therefore, a final cluster discriminator corresponding to a fault category is obtained by clustering, and here the corresponding aggregate member value AMV and its corresponding threshold are obtained according to the final cluster discriminator. A category label can be established according to the threshold of the aggregate member value, and category identification can be performed according to the category label. In the actual fault category identification process, the judgment result is obtained according to the comparison result of the obtained aggregate member value AMV and the threshold β.

[0055] Based on step S14, the discriminator in the initial neural network is updated according to the discriminator obtained by clustering, and finally a corresponding fault detection model can be obtained. The fault category of the induction motor can be detected through the corresponding discriminator.

[0056] In one embodiment, if Figure 2As shown, in the step S12, the initial discriminator corresponding to the fault category in the initial neural network is constructed according to the first sample data, and the node address corresponding to the initial discriminator and the node data corresponding to the node address are obtained; including: S121, preprocessing and feature extraction of the first sample data to obtain a multidimensional real-valued vector corresponding to the first sample data; S122, normalizing the multidimensional real-valued vector within an integer range, and performing binary conversion to obtain a corresponding multidimensional binary array; S123, bit filling to make the length of the multidimensional binary array a preset length, and splitting the multidimensional binary array based on the number of nodes to obtain a number of data strings, wherein the length of the data string corresponds to the node address of the discriminator; S124, obtaining the node address corresponding to the discriminator and the node data corresponding to the node address according to the content of the data string.

[0057] Based on step S121, in the process of generating the initial discriminator, the first sample data is preprocessed and feature extracted. The discriminant unit can be preliminarily determined based on the prior knowledge of the fault type and data label, and then the useful signal information is converted into fault features.

[0058] Taking the detection and identification of different faults in single-phase and three-phase induction motors as an example, these faults are loaded into the system in the form of broken rotor bars, bearing faults, unbalanced loads and broken cage rings, that is, the fault categories are obtained. The causes of these faults are partial mechanical component failures caused by mechanical damage to the above mechanical components. Sensor data of normal and faulty motors are obtained, and fault features are extracted from the acquired sensor data. Three real states of the system are analyzed: healthy state, bearing fault (ball damage) and load fault (load unbalanced component). The raw data includes current and vibration signals.

[0059] Taking a single-phase induction motor as an example, current and vibration signal samples in three states are collected respectively. For each sample, feature extraction is performed, and four features are extracted, including the mean and standard deviation of the current and the mean and standard deviation of the vibration (μI, σI, μV, σV). The result of feature extraction is 180 d-dimensional real-valued input vectors (d = 4). Two-thirds of the data set (i.e., 120 feature vectors) are used for classifier training. The WNN discriminator is trained according to different motor operating speeds (i.e., three levels) and motor states (healthy and two faults). For a specific operating speed and motor state, 40 training samples are obtained, totaling 120 samples. The remaining one-third of the real-valued input vectors (i.e., 60 feature vectors) are used for testing and performance evaluation.

[0060] The three-phase induction motor setup was followed by the single-phase induction motor setup. The experimental setup included a 0.55kW, 220V three-phase induction motor with the following operating parameters: 50Hz frequency, 1400rpm rotor speed, and 2.52 / 1.47A (Δ / Y) current. Four states of the three-phase induction motor were analyzed: healthy three-phase induction motor, three-phase induction motor with damaged squirrel cage rings, three-phase induction motor with broken rotor bars, and three-phase induction motor with two broken rotor bars. For acoustic-based fault diagnosis, a digital recording device (range: 20Hz-20kHz) was used. Other parameters include format: WAVE, sampling frequency 44100Hz, 16-bit, mono, omnidirectional microphone. The sound was split into smaller 1-second audio files, normalized in the range [-1,1], and then feature extraction was performed. In dataset 2, 440 signal samples (i.e., 110 samples for each motor state) were obtained. For each sample, feature extraction was performed and four features were extracted, including the mean, variance, kurtosis and spectral mean (μ, σ2, k, μS1) of the acoustic signal. The result of feature extraction was 440 d-dimensional real-valued input vectors (d = 4). A total of 400 feature vectors were used for classifier training, and the remaining 40 feature vectors were used for testing and performance evaluation.

[0061] Based on step S122, after obtaining the multidimensional real-valued vector corresponding to the sample data, the multidimensional real-valued vector is normalized, wherein the multidimensional real-valued vector can be understood as a real-valued vector corresponding to each feature. Figure 8 As shown in the figure, the real-valued vectors obtained based on the raw data of the sensor include 1.45, 0.02, 0.41 and 0.98, and the normalized integers are 198, 2, 56 and 107. And the decimal values ​​obtained are converted into binary data values ​​according to the preset rules. For example, the multidimensional binary arrays obtained are 10100101, 00000011, 00100100 and 010101110. Among them, the d-dimensional real-valued vector is converted into a binary data value in the integer range of 0 to 2. n -1, where "n" defines the level of detail of the decimal-to-binary conversion, and each integer value in the range corresponds to a specific n-bit binary pattern. In one embodiment, the level of detail n can be modified to obtain a different binary array.

[0062] In a specific embodiment, in step S122, the multidimensional real-valued vector is normalized within an integer range and converted into a binary format to obtain a corresponding multidimensional binary array, including: converting the decimal value obtained by the normalization process into a binary character, and performing a binary reflected Gray code representation to generate a corresponding multidimensional binary array. In some cases, a simple real-value to binary string encoding method may cause the discriminator function to fail, and the possibility of failure can be reduced by using a binary reflected Gray code.

[0063] Based on step S123, because the lengths of the obtained multidimensional binary arrays may be different, bit padding is used to make the lengths of each binary array the same. The multidimensional binary array is split according to the number of nodes to obtain several data strings. When splitting the multidimensional binary array, the multidimensional binary array can be sequentially composed into a one-dimensional binary array, and the length of the data string is determined according to the number of nodes required, wherein the more nodes there are, the shorter the length of the data string. In the process of obtaining the data string, it is necessary to satisfy #nodes×node_size=d×n+Δ; wherein #nodeS1 is the number of nodes corresponding to the discriminator, node_S1ize is the node data size of the discriminator, d is the dimension of the multidimensional real-valued vector, n is the level of detail from decimal to binary during binary conversion; Δ is the number of padding bits. The node address of the discriminator is denoted by the length of the data string. For example, after inputting a feature vector with a dimension of 4, the binary string is concatenated into a string with a total length of 4×8 bits through reflective Gray code conversion and padding to an 8-bit binary string. Assume that the data size of each RAM node is node_S1ize, and the number of RAM nodes is nodes, which is equal to the number of node sub-address strings. At this time, the size of the node sub-address string is 32 / nodes.

[0064] Optionally, in the step S123, the multidimensional binary array is split based on the number of nodes to obtain a number of data strings, including: updating the node data to update the length of the data string, so as to train the node address corresponding to the discriminator and the node data corresponding to the node address to obtain the final node address and node data corresponding to the discriminator. Considering the sensitivity and influence on the model, different experiments are carried out by changing the design parameters in the process of obtaining the node address and node data. Based on the single-phase induction motor, referring to Table 1, for each group of experiments, the training and testing process is repeated 20 times, and the average fault recognition accuracy and misclassification error rate are obtained. It is worth noting that by increasing the number of nodes, the generalization ability of the classifier can be improved, so that the average classification recognition rate is improved, but at the same time the misclassification error will also increase, and vice versa. Through the trial and error method, a reasonable balance can be found between accurate classification recognition and misclassification errors. In addition, the optimal value of the number of bits of binary representation and the node size depends on the distribution of the input data. If the data distribution between categories and within categories has a large deviation and a small variance, a smaller "n" and a smaller node size may be effective. Otherwise, if there is overlap between categories, an appropriate value of “n” should be chosen to capture subtle differences in the data distribution.

[0065] After obtaining the optimal values ​​for the number of nodes, node size, and “n”, further increasing or decreasing these values ​​does not improve the classification performance. Similarly, in training phase II, higher values ​​of the threshold “α” for the expansion of the recognition region will lead to overfitting, while smaller values ​​may lead to underfitting of the classifier. The value of the threshold “β” depends on the number of nodes in the discriminator and ranges between (0, #nodes). Smaller values ​​of “β” may lead to better class identification accuracy, but at the same time, the misclassification error will increase. For each set of experiments, the choice of threshold and scaling factor was determined by trial and error. The role of the scaling factor is to adjust the weight of different features in the address mapping.

[0066] Table 1 Single-phase induction motor fault identification results

[0067]

[0068] Based on a three-phase induction motor, four features were extracted from the collected acoustic signals: mean, variance, kurtosis, and spectral mean (μ, σ2, k, μS1) of the acoustic signals. In order to investigate the fault identification performance of the developed method, different experiments were conducted to determine the design parameters, which are d=4, m=4, and n=4. The training and testing process was repeated 20 times, and the average fault identification accuracy and misclassification error rate were obtained. The scaling factor γ and thresholds α and β were 1.5, 2, and 3, respectively.

[0069] Table 2. Three-phase induction motor fault identification results

[0070]

[0071] Based on step S124, after obtaining the node address of the discriminator, the node data corresponding to the node address is obtained according to the content of the data string and the node address, and a node of the discriminator is formed by the node address and the node data. The node here can be understood as a network node in a neural network.

[0072] Optionally, in step S13, the several initial discriminators corresponding to the fault category are clustered, and the cluster discriminators corresponding to the fault category are obtained according to the clustering results, including: S131, obtaining the node address cluster center of the discriminator corresponding to each fault category, and obtaining the node data cluster center corresponding to the node address cluster center according to the node address cluster center, so as to obtain the clustered node address and node data. Subsequently, the optimal threshold β is determined through experiments and adjustments. First, an initial threshold is selected (for example, based on the average aggregate membership value on the training set). Then, the performance of the classifier is evaluated on the validation set, the accuracy is calculated, and the threshold is gradually adjusted to observe its impact on the classification performance. Specifically, as Figure 8 and Fig. 9 As shown in the figure, if the input 4-dimensional vectors are 1.45, 0.02, 0.41, and 0.98, first find the node address with a member value of 1 from all RAM nodes as the initial clustering point. The storage content of the first two initial RAM nodes obtained according to the above steps is 111100 and 001110, and the center point is calculated to be (111100+001110) / 2=111110. Then update the content of the RAM node closest to the cluster center, assuming it is 111010, and update its member value from 0 to a higher value, such as 0.8, indicating a high similarity to the input pattern. Further, update the node content according to the Hamming distance, and for other RAM nodes far from the cluster center, update their member values ​​based on their Hamming distance from the center point. Nodes with a small Hamming distance will obtain a higher member value, and nodes with a large distance will obtain a lower member value. By calculating the cluster center point and updating the member value based on the Hamming distance, the recognition area of ​​the node address is expanded and the generalization ability of the input pattern is improved.

[0073] like Figure 7 As shown, a schematic diagram of the working of the fault detection model is shown. Among them, the initial neural network can also be understood as a discriminator network composed of multiple discriminators. In the construction process, a discriminator network containing multiple discriminators is constructed. The multi-dimensional real-valued input (i.e., d-dimensional real-valued input) is obtained by preprocessing and feature extraction of the original data. The input to address mapping is realized through the above process, and the node addresses and node contents of multiple discriminators are constructed.

[0074] like Figure 3 As shown, optionally, in the method for constructing an induction motor fault detection model of the present invention, it also includes: S15, obtaining the corresponding induction data based on the fault category of the induction motor as the second sample data, wherein the second sample data is not repeated with the first sample data; S16, obtaining the cluster discriminator corresponding to the fault category based on the second sample data, and re-clustering the node address and node data of the cluster discriminator to obtain an updated cluster discriminator. The cluster discriminator can be trained, that is, the corresponding second sample data is obtained. In one embodiment, all the sample data can be obtained at one time, and part of the sample data can be extracted as the first sample data, and the remaining sample data can be used as the second sample data.

[0075] For example, for each fault category, the corresponding cluster discriminator is trained using sample data of the fault category. For each real-valued training instance, the content of a specific node position is modified to "1" according to the sub-address obtained by inputting to the address mapping. This process is repeated for all categories of discriminators. During the training process, the information in each RAM node in the discriminator is summarized by the input pattern learning method, and the recognition area of ​​the discriminator node address is expanded. The expansion process starts with the RAM node address with a cluster value of "1". The center point of each cluster is calculated, and the content of the RAM node position closest to the cluster center is modified by storing a new member value. Based on the Hamming distance between the cluster center address and other unit addresses (the storage value is "0"), the content of the RAM node position is updated. The content update formula for each RAM node is as follows.

[0076] for each node i(i=l to#nodes)

[0077] for each node-cluster j(j=l to#cluster)

[0078] ifAddress jthcentroid -Address cell <α&Content cell =0

[0079]

[0080] Where α is the threshold that controls the expansion of the recognition region, and γ is the scaling factor. Address is the RAM node address, content is the stored content, and jthcentroid is the jth cluster center. The expansion of the recognition region is proportional to the similarity measure between the address location and the previously stored input pattern address location. Higher category membership values ​​are assigned to the address locations closest to the cluster center, which show the possibility that the input pattern belongs to the corresponding category. If the input training data is limited, the expansion of the recognition region of the RAM node will lead to better generalization of the input pattern.

[0081] like Figure 4 As shown, an induction motor fault detection model construction device of the present invention is used to construct an induction motor fault detection model, and the device includes:

[0082] A first processing module 110 is used to construct an initial neural network and obtain corresponding induction data as first sample data based on the fault type of the induction motor;

[0083] A second processing module 120 is used to construct an initial discriminator corresponding to the fault category in the initial neural network according to the first sample data, and obtain a node address corresponding to the initial discriminator and node data corresponding to the node address;

[0084] The third processing module 130 is used to cluster the initial discriminators corresponding to the fault category, and obtain the cluster discriminator corresponding to the fault category according to the clustering result, wherein the cluster discriminator includes the node address and node data after clustering;

[0085] The fourth processing module 140 is used to construct a final fault detection model according to the cluster discriminator and the initial neural network.

[0086] That is, the induction motor fault detection model building device has the function of implementing the corresponding steps performed in the above method. Each function can be implemented by hardware, or by hardware executing the corresponding software. The corresponding hardware or software includes one or more modules corresponding to the above functions. That is, the steps in the above method are respectively executed by one or more modules. The specific coordination operations between the modules can refer to the specific process of the above induction motor fault detection model building method, which will not be repeated here.

[0087] In addition, such as Figure 5As shown, an induction motor fault detection method of the present invention is used to realize fault detection of an induction motor. The method comprises the following steps: S21, obtaining the induction data of the induction motor; S22, obtaining the corresponding node address and the node data corresponding to the node address from the induction data; S23, based on a preset fault detection model, obtaining the corresponding discriminator according to the node address and node data corresponding to the induction data, and outputting the detection result through the preset fault detection model; wherein the preset fault detection model is obtained by the above-mentioned induction motor fault detection model construction method. For example, assuming that the input real-valued feature vector of a group of faulty induction motors is: [5.5, 2.3, 8.9, 3.4], the data is normalized to [140, 64, 227, 87], and then the decimal to binary conversion (using BRGC) is performed and the address string is obtained by filling and splicing to obtain 10101100110000001110001101010111. At this time, if the node sub-address length and RAM node size are both 8 bits. The corresponding RAM nodes are queried and the member values ​​obtained are 0.8, 0.6, 0.9, and 0.7 respectively. The aggregate member value is then calculated, AMV = 0.8 + 0.6 + 0.9 + 0.7 = 3.0. The aggregate member value AMV is compared with the threshold β. Assuming β = 2.5, then 3.0>2.5, so it is determined that the input pattern belongs to this category.

[0088] In addition, such as Figure 6 As shown, an induction motor fault detection device of the present invention is used for fault detection and identification of an induction motor, and the device includes: an input module 210, used to obtain induction data of the induction motor; a data processing module 220, used to obtain the node address corresponding to the induction data and the node data corresponding to the node address; a result generation module 230, used to obtain the corresponding discriminator according to the node address and node data corresponding to the induction data based on a preset fault detection model, and output the detection result through the preset fault detection model; wherein the preset fault detection model is obtained by the above induction motor fault detection model construction method.

[0089] That is, the induction motor fault detection device has the function of implementing the corresponding steps executed in the above-mentioned induction motor fault detection method. Each function can be implemented by hardware, or by hardware executing corresponding software. The corresponding hardware or software includes one or more modules corresponding to the above-mentioned functions. That is, the steps in the above-mentioned method are respectively executed by one or more modules. The specific coordination operations between the modules can refer to the specific process of the above-mentioned induction motor fault detection method, which will not be repeated here.

[0090] It can be understood that the above embodiments only express the preferred implementation modes of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the patent scope of the present invention. It should be pointed out that, for ordinary technicians in this field, the above technical features can be freely combined without departing from the concept of the present invention, and several deformations and improvements can be made, which all belong to the protection scope of the present invention. Therefore, all equivalent changes and modifications made to the scope of the claims of the present invention should belong to the coverage of the claims of the present invention.

Claims

1. A method for constructing an induction motor fault detection model, characterized in that: The method comprises the following steps: S11, constructing an initial neural network, and acquiring corresponding induction data as first sample data based on the fault type of the induction motor; S12, constructing an initial discriminator corresponding to the fault category in the initial neural network according to the first sample data, and obtaining a node address corresponding to the initial discriminator and node data corresponding to the node address; S13, clustering a number of initial discriminators corresponding to the fault category, and obtaining a cluster discriminator corresponding to the fault category according to the clustering result, wherein the cluster discriminator includes a clustered node address and node data; S14. Constructing a final fault detection model based on the cluster discriminator and the initial neural network.

2. The method for constructing an induction motor fault detection model according to claim 1, characterized in that: In the step S12, the initial discriminator corresponding to the fault category in the initial neural network is constructed according to the first sample data, and a node address corresponding to the initial discriminator and node data corresponding to the node address are obtained; including: S121, preprocessing and feature extraction of the first sample data to obtain a multi-dimensional real-valued vector corresponding to the first sample data; S122, normalizing the multidimensional real-valued vector within an integer range, and performing binary conversion to obtain a corresponding multidimensional binary array; S123, making the length of the multi-dimensional binary array a preset length by bit filling, and splitting the multi-dimensional binary array based on the number of nodes to obtain a plurality of data strings, wherein the length of the data string corresponds to the node address of the discriminator; S124. Obtain a node address corresponding to the identifier and node data corresponding to the node address according to the content of the data string.

3. The method for constructing an induction motor fault detection model according to claim 2, characterized in that: In the step S123, the multidimensional binary array is split based on the number of nodes to obtain a number of data strings, including: updating the node data to update the length of the data string, training the node address corresponding to the discriminator and the node data corresponding to the node address to obtain the final node address and node data corresponding to the discriminator.

4. The method for constructing an induction motor fault detection model according to claim 1, characterized in that: In step S122, the multidimensional real-valued vector is normalized within an integer range and binary converted to obtain a corresponding multidimensional binary array, including: The decimal value obtained by normalization is converted into binary characters, and binary reflected Gray code representation is performed to generate a corresponding multi-dimensional binary array.

5. The method for constructing an induction motor fault detection model according to claim 1, characterized in that: In the step S13, clustering the several initial discriminators corresponding to the fault category and obtaining the cluster discriminator corresponding to the fault category according to the clustering result includes: S131, obtaining the node address cluster center of the discriminator corresponding to each fault category, and obtaining the node data cluster center corresponding to the node address cluster center according to the node address cluster center, so as to obtain the clustered node address and node data.

6. The method for constructing an induction motor fault detection model according to claim 1, characterized in that: The method further comprises: S15, acquiring corresponding induction data based on the fault type of the induction motor as second sample data, wherein the second sample data is not repeated with the first sample data; S16. Acquire a cluster identifier corresponding to the fault category based on the second sample data, and re-cluster the node address and node data of the cluster identifier to obtain an updated cluster identifier.

7. The method for constructing an induction motor fault detection model according to claim 1, characterized in that: In step S11, the fault category of the induction motor includes: one or more of broken rotor bars, bearing failure, unbalanced load and broken cage ring; and / or In the step S11, the sensing data includes current data and vibration signal data.

8. An induction motor fault detection model construction device, characterized in that: The device comprises: A first processing module is used to construct an initial neural network and obtain corresponding induction data as first sample data based on the fault category of the induction motor; A second processing module is used to construct an initial discriminator corresponding to the fault category in the initial neural network according to the first sample data, and obtain a node address corresponding to the initial discriminator and node data corresponding to the node address; A third processing module, used for clustering a number of initial discriminators corresponding to the fault category, and obtaining a cluster discriminator corresponding to the fault category according to the clustering result, wherein the cluster discriminator includes a clustered node address and node data; The fourth processing module is used to construct a final fault detection model according to the cluster discriminator and the initial neural network.

9. A method for detecting an induction motor fault, characterized in that: The following steps are involved: Acquire induction data of induction motor; Acquire the sensing data to acquire a corresponding node address and node data corresponding to the node address; Based on a preset fault detection model, a corresponding discriminator is obtained according to the node address and node data corresponding to the sensing data, and a detection result is output through the preset fault detection model; Wherein, the preset fault detection model is obtained by the induction motor fault detection model construction method according to any one of claims 1 to 7.

10. An induction motor fault detection device, characterized in that: include: An input module for acquiring induction data of the induction motor; A data processing module, used for acquiring the node address corresponding to the sensing data and the node data corresponding to the node address; A result generating module, for obtaining a corresponding discriminator according to the node address and node data corresponding to the sensing data based on a preset fault detection model, and outputting a detection result through the preset fault detection model; Wherein, the preset fault detection model is obtained by the induction motor fault detection model construction method according to any one of claims 1 to 7.

Citation Information

Patent Citations

  • Three-phase motor fault detection method and device

    CN106872894A

  • Construction method and device of rotary mechanical equipment fault diagnosis model

    CN116881832A