Motor rotor defect collaborative analysis system and method based on knowledge-driven clustering

Abstract

The invention discloses a motor rotor defect collaborative analysis system and method based on knowledge-driven clustering, and relates to the technical field of motor rotor defect diagnosis. Comprising the following steps: S1, data acquisition: acquiring motor operation data in real time; s2, feature extraction: carrying out preprocessing and feature extraction on the motor operation data, and generating feature vectors related to motor rotor defects; and S3, a domain knowledge base is constructed, a structured domain knowledge base is established, and the domain knowledge base comprises expert experience rules, physical model constraints and historical defect data labels. According to the invention, multi-dimensional operation data of motor vibration, current, temperature and voltage are collected in real time, so that the problem of incomplete information of traditional single signal collection is avoided; accurate feature vectors are extracted through data cleaning, noise reduction and other preprocessing and time domain, frequency domain and time-frequency domain analysis, and the noise interference problem of original data is solved.

Description

Technical Field

[0001] This invention relates to the field of motor rotor defect diagnosis technology, specifically to a collaborative analysis system and method for motor rotor defects based on knowledge-driven clustering. Background Technology

[0002] As a key power device in industrial production, the operating status of electric motors directly affects production efficiency and safety. The motor rotor is a crucial component; defects such as broken bars, eccentricity, and partial short circuits can lead to decreased motor performance and even shutdown accidents. Therefore, accurate and timely diagnosis of motor rotor defects is essential. Traditional methods for diagnosing motor rotor defects typically rely on human experience, physical models, or signal processing techniques. However, these methods often lack sufficient accuracy and efficiency when dealing with complex operating conditions, multiple faults, or subtle defects. In recent years, with the development of big data and machine learning technologies, data-driven fault diagnosis methods, especially cluster analysis, have begun to be applied to motor defect diagnosis.

[0003] However, existing clustering methods are mostly general-purpose algorithms that fail to fully utilize the unique knowledge in the field of motor fault diagnosis, such as the physical laws governing motor operation, expert diagnostic experience, and historical fault patterns. This results in clustering results that may lack physical meaning, have poor correlation with actual defect types, and be sensitive to noise and anomalies, making it difficult to achieve accurate and reliable automated diagnosis and hindering their widespread application in real-world industrial scenarios. Therefore, how to integrate rich domain knowledge into clustering analysis to overcome the limitations of existing methods is a pressing technical problem to be solved in the field of motor rotor defect diagnosis. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a collaborative analysis system and method for motor rotor defects based on knowledge-driven clustering. This solves the problems that, compared with existing technologies, traditional methods lack accuracy and efficiency in diagnosing complex working conditions, and existing clustering methods do not fully utilize domain knowledge, resulting in clustering results lacking physical meaning and having poor correlation with actual defects.

[0005] To achieve the above objectives, this invention provides the following technical solution: a collaborative analysis method for motor rotor defects based on knowledge-driven clustering, comprising: S1. Data acquisition: Real-time acquisition of motor operation data; S2. Feature extraction: Preprocess and extract features from the motor operation data to generate feature vectors related to motor rotor defects. S3. Construct a domain knowledge base and establish a structured domain knowledge base, which includes expert experience rules, physical model constraints, and historical defect data labels; S4. Knowledge-driven clustering: Design and implement a knowledge-driven clustering algorithm that utilizes information from the domain knowledge base for guidance and optimization during the clustering process. S5. Clustering result evaluation and interpretation: Evaluate the clustering results output by the knowledge-driven clustering algorithm, and associate and interpret the clusters with known defect types in the domain knowledge base.

[0006] Furthermore, the data acquisition includes: The system collects vibration, current, temperature, and voltage signals from the motor in real time. The vibration signals include radial and axial vibrations; the current signals include three-phase and zero-sequence currents; the temperature signals include stator winding and bearing temperatures; and the voltage signals include three-phase and phase-to-phase voltages.

[0007] Furthermore, the feature extraction includes: Perform data cleaning, noise reduction, and normalization preprocessing on the motor operation data; Time-domain analysis, frequency-domain analysis, or time-frequency-domain analysis are performed on the preprocessed motor operation data to obtain time series characteristics, spectral characteristics, and wavelet packet energy characteristics. Based on time series features, spectral features, and wavelet packet energy features, feature vectors characterizing the defect state of the motor rotor are generated for knowledge-driven clustering analysis.

[0008] Furthermore, the construction of the domain knowledge base includes: Establish an expert experience rule base, which describes the relationship between specific feature value ranges, feature combinations, or feature change trends and specific defect types. Establish a physical model constraint library, which is based on the laws of electromagnetism, mechanical vibration or thermodynamics, and imposes restrictions on eigenvalues, eigenre relationships or defect evolution processes. Establish a historical defect data tag library, which records the verified types of motor rotor defects and their corresponding feature patterns in the feature space.

[0009] Furthermore, the knowledge-driven clustering includes: The feature weighting mechanism dynamically adjusts the weight of each feature in similarity calculation based on expert experience rules or historical defect data tags stored in the domain knowledge base. This adjustment is made according to the sensitivity of different features to specific defects or their ability to distinguish known defect types, thereby highlighting features that are strongly related to motor rotor defects.

[0010] Furthermore, the knowledge-driven clustering also includes: The clustering constraint mechanism uses physical model constraints in the domain knowledge base to apply soft or hard constraints to the clustering process. Soft constraints guide the clustering direction through penalty terms, while hard constraints force the clustering results to meet preset physical boundary conditions to ensure that the formation of clusters conforms to the physical laws of motor operation.

[0011] Furthermore, the knowledge-driven clustering also includes: The cluster center initialization guidance mechanism uses known defect patterns or expert-defined typical defect feature value ranges from historical defect data tags in the domain knowledge base to initialize the cluster centers of the clustering algorithm, thereby accelerating the convergence of the clustering algorithm and improving the quality and stability of the clustering results.

[0012] Furthermore, the evaluation and interpretation of the clustering results include: The clusters output by the knowledge-driven clustering algorithm are evaluated internally or externally. The internal evaluation uses the silhouette coefficient or Davies-Bouldin index to measure compactness and separation, while the external evaluation is compared and verified with pre-labeled test data. The evaluated clusters are associated and mapped with known defect types in the domain knowledge base, and a motor rotor defect diagnosis report is generated, providing an explanation of the defect type, severity, and possible location.

[0013] Furthermore, it also includes: Defect diagnosis feedback and knowledge base iterative updates: Based on the diagnosis report, the diagnosis results are confirmed through expert review or further experimental verification. Based on the confirmed diagnostic results, the historical defect data tags in the domain knowledge base are dynamically updated, and new defect patterns and feature descriptions are supplemented or corrected. Based on the newly added domain knowledge, optimize expert experience rules or adjust physical model constraints to improve the diagnostic accuracy and adaptability of knowledge-driven clustering algorithms.

[0014] This invention also provides a knowledge-driven clustering-based collaborative analysis system for motor rotor defects, applied to any of the knowledge-driven clustering-based collaborative analysis methods for motor rotor defects described above, comprising: The data acquisition module is used to collect motor operating data in real time; The feature extraction module is used to preprocess and extract features from the motor operation data to generate feature vectors; The knowledge base construction module is used to build a structured domain knowledge base, which includes expert experience rules, physical model constraints, and historical defect data tags. The clustering analysis module is used to design and implement a knowledge-driven clustering algorithm that utilizes information from the domain knowledge base for guidance and optimization during the clustering process. The evaluation and interpretation module is used to evaluate the clustering results output by the clustering analysis module and to associate and interpret the clusters with known defect types in the domain knowledge base.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention avoids the incomplete information problem of traditional single-signal acquisition by real-time acquisition of multi-dimensional operating data such as motor vibration, current, temperature, and voltage. Preprocessing, including data cleaning and noise reduction, along with time-domain, frequency-domain, and time-frequency-domain analysis, extracts precise feature vectors to solve the problem of noise interference in the raw data. Simultaneously, it constructs a domain knowledge base containing expert experience rules, physical model constraints, and historical defect data labels. Combined with knowledge-driven clustering's feature weighting, clustering constraints, and cluster center initialization mechanisms, it overcomes the problems of existing clustering methods that do not fully utilize domain knowledge, lack physical meaning in the results, and have poor correlation with actual defects. Furthermore, it improves diagnostic accuracy and adaptability through cluster evaluation interpretation and iterative updates to the knowledge base, achieving automated collaborative analysis of motor rotor defects, reducing manual intervention, and solving the problems of insufficient accuracy and efficiency in diagnosing complex operating conditions using traditional methods. This makes it suitable for practical industrial scenarios. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system structure diagram of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1-2 This invention provides a collaborative analysis method for motor rotor defects based on knowledge-driven clustering, including: S1. Data acquisition: Real-time acquisition of motor operation data; S2. Feature extraction: Preprocess and extract features from the motor operation data to generate feature vectors related to motor rotor defects. S3. Construct a domain knowledge base. Establish a structured domain knowledge base, which includes expert experience rules, physical model constraints, and historical defect data labels. S4. Knowledge-driven clustering: Design and implement a knowledge-driven clustering algorithm that uses information from the domain knowledge base for guidance and optimization during the clustering process. S5. Clustering Result Evaluation and Interpretation: Evaluate the clustering results output by the knowledge-driven clustering algorithm, and associate and interpret the clusters with known defect types in the domain knowledge base.

[0019] Specifically, a multi-channel data acquisition system is employed to collect multi-dimensional signals during motor operation in real time, avoiding the incomplete information problems caused by traditional single-signal acquisition. The acquired signals include vibration, current, temperature, and voltage signals: vibration signals require simultaneous acquisition of radial and axial vibrations, with acquisition locations such as the front and rear bearing housings of the motor, using piezoelectric accelerometers at a sampling frequency of, for example, 20kHz; current signals acquire three-phase current and zero-sequence current, with acquisition points set at the stator winding output terminals, using Hall effect current sensors, and the sampling frequency is consistent with the vibration signal; temperature signals acquire stator winding temperature and bearing temperature, with stator winding temperature acquired using embedded PT100 platinum resistance sensors and bearing temperature acquired using surface-mounted temperature sensors, with a sampling interval of, for example, 1 second; voltage signals acquire three-phase voltage and phase-to-phase voltage, with acquisition points at the same locations as the current signal acquisition points, using voltage transformers, and the sampling frequency is consistent with the current signal. During acquisition, an NI9234 data acquisition card converts analog signals into digital signals and stores them in an industrial database, providing comprehensive, high-quality raw data for subsequent analysis.

[0020] First, the collected motor operation data is preprocessed to address the impact of noise and outliers on feature quality: Data cleaning employs the 3σ criterion; when a data point deviates from the mean of a signal dimension by more than three times the standard deviation, it is identified as an outlier and removed. Missing data is then supplemented using linear interpolation. Data denoising uses wavelet thresholding, selecting a db4 wavelet basis for three-level signal decomposition. Thresholds are determined using heuristic rules, and high-frequency coefficients are thresholded before signal reconstruction. Data normalization uses the min-max normalization method to ensure that features of different dimensions can be subsequently calculated. The normalization formula is: ; in, These are the original eigenvalues. This is the minimum value of this feature dimension. This represents the maximum value of this feature dimension. These are the normalized eigenvalues. All normalized eigenvalues ​​are in the interval [0,1] and have the same dimensions.

[0021] After preprocessing, time-domain analysis, frequency-domain analysis, and time-frequency-domain analysis are performed to extract features: Time-domain analysis calculates statistical features of the signal such as mean, variance, peak value, peak factor, and kurtosis. For example, the peak factor of a vibration signal can reflect impact defects. Frequency-domain analysis converts the time-domain signal into a frequency-domain signal using Fast Fourier Transform (FFT) to calculate features such as power spectral density and characteristic frequency amplitude. For example, the characteristic frequency of the stator current corresponding to a broken rotor bar can be extracted through frequency-domain analysis. Time-frequency-domain analysis uses wavelet packet decomposition to perform four-level wavelet packet decomposition on the signal, calculating the energy proportion of each frequency band as a feature. The wavelet packet energy calculation formula is as follows: ; ; in, The energy of the k-th wavelet packet frequency band. Let be the value of the i-th sampling point in the k-th frequency band, and N be the number of sampling points in that frequency band. Let M be the energy percentage of the k-th frequency band, and M be the total number of frequency bands after wavelet packet decomposition.

[0022] Finally, key features are selected from all extracted features. For example, redundant features are eliminated by mutual information method, and a feature vector containing 20 key features is finally generated. This vector can accurately characterize the defect state of the motor rotor.

[0023] A structured domain knowledge base is established to transform expert experience, physical laws, and historical data in the field of motor fault diagnosis into knowledge that can be used to guide clustering, thus addressing the problem of existing clustering methods lacking domain knowledge support. The knowledge base consists of three parts: Expert Experience Rule Base: Five experts with over 10 years of experience in diagnosing motor rotor defects were invited to develop rules using the Delphi method. The rules relate specific characteristic value ranges, characteristic combinations, or characteristic change trends to specific defect types. For example, "When the radial vibration peak factor is greater than 5 and the stator winding temperature exceeds 120℃, it is likely a rotor bar breakage defect." "When the zero-sequence current is greater than 5A and the effective value of axial vibration is greater than 0.1mm, it may be a rotor eccentricity defect." All rules are stored in a structured language for easy retrieval later.

[0024] Physical model constraint library: Based on the physical laws of motor electromagnetics, mechanical vibration, and thermodynamics, it imposes constraints on eigenvalues, characteristic relationships, or defect evolution processes. For example, based on electromagnetic laws, the characteristic frequency of the stator current when the rotor bar breaks is... ,in The frequency is the power grid frequency, usually 50Hz, and s is the motor slip, which is generally between 0.01 and 0.05. Therefore, the characteristic current amplitude near this characteristic frequency must satisfy "0.2V ≤ characteristic amplitude ≤ 2V". Based on the mechanical vibration law, the characteristic frequency of radial vibration when the rotor is eccentric includes twice the rotational frequency. The vibration amplitude corresponding to this frequency must satisfy "0.05mm ≤ amplitude ≤ 0.5mm". All constraints are stored in the form of mathematical expressions.

[0025] Historical Defect Data Tag Library: Collects motor rotor defect data from various industries such as thermal power, chemical industry, and manufacturing over the past 5 years. It covers common defect types such as rotor broken bars, rotor eccentricity, and rotor partial short circuit. 100 sets of verified data are collected for each defect type. Each set of data includes the preprocessed raw signal, extracted feature vectors, and defect type tags that have been reviewed by experts and verified by experiments. For example, the tag for a set of data is "rotor broken bar - moderate - near stator winding A phase". The data is stored in the database in tabular form, supporting subsequent clustering initialization and feature weighting.

[0026] An improved K-means clustering algorithm was designed, incorporating information from a domain knowledge base to address the problems of traditional clustering results lacking physical meaning and having poor correlation with defect types. The algorithm comprises three core mechanisms: Feature weighting mechanism: Based on expert experience rules and historical defect data tags in the domain knowledge base, the weight of each feature in the similarity calculation is dynamically adjusted to highlight features strongly associated with the defect. The weight calculation uses a weighted fusion method, with the following formula: ; ; in, Let be the initial weight of the i-th feature. This is the weighting coefficient, with a value range of 0 < α < 1, for example, 0.6. The importance score given by experts for the i-th feature is calculated, with expert scores ranging from 1 to 10, and then normalized to [0,1] using the min-max method. Let be the discrimination index of the i-th feature. The correlation between the feature and the defect type is calculated using mutual information and normalized to [0,1]. To ensure the final weights after normalization, the sum of all feature weights is 1. For example, for rotor bar breakage defects, the current characteristic... , Calculated After unification The weight is much higher than that of temperature features, for example, 0.2, to highlight key features.

[0027] Clustering constraint mechanism: Using constraints from the physical model constraint library, soft and hard constraints are applied to the clustering process to ensure that the clustering results conform to the physical laws of motor operation. Hard constraints directly restrict the cluster affiliation of samples. For example, if the stator winding temperature characteristic value of a sample exceeds 155℃, which is the maximum allowable temperature for a Class A insulated motor, it is forcibly classified as an "abnormal defect cluster." Soft constraints guide the clustering direction through penalty terms, improving the K-means objective function to: ; in, The objective function value, For the number of clusters, For the k-th cluster, For sample feature vectors, The center of the k-th cluster, This is a penalty coefficient, for example, 0.5. Let x be the constraint function, which is defined when sample x violates physical constraints. ,otherwise By using this objective function, the clustering process will prioritize grouping samples that meet the physical constraints into the same cluster, thereby improving the physical rationality of the clustering results.

[0028] Cluster center initialization guidance mechanism: This mechanism utilizes known defect patterns from a historical defect data tag library to initialize the cluster centers of the clustering algorithm, accelerating algorithm convergence and improving clustering quality. Specifically, it calculates the mean of the feature vectors for each defect type in the historical data and uses this mean vector as the initial cluster center for the corresponding defect type, such as the initial cluster center for rotor bar breakage defects. ,in The mean of the i-th feature of the broken bar defect sample, and the initial cluster center of the rotor eccentricity defect. If there is insufficient historical data for a certain type of defect, such as fewer than 30 sets, then the characteristic value range of typical defects defined by experts is adopted, and the midpoint of the range is taken as the initial cluster center. For example, if the effective value range of axial vibration of eccentric defects is defined by experts as 0.08-0.12mm, then 0.1mm is taken as the initial cluster center component of this feature.

[0029] First, the clustering results are evaluated from multiple dimensions to ensure clustering quality: the internal evaluation uses the silhouette coefficient and the Davies-Bouldin index. The silhouette coefficient is calculated using the following formula: ; in, Let be the contour coefficient of sample i. Let i be the average distance from sample i to all other samples in the same cluster. Let i be the average distance from sample i to all samples in the nearest heterogeneous cluster. The closer the index is to 1, the more reasonable the cluster assignment of the samples, and the better the cluster compactness and separation. The Davies-Bouldin index calculates the ratio of the mean intra-cluster distance to the inter-cluster distance; the smaller the index, the better the clustering effect. External evaluation compares the clustering results with pre-labeled test data, such as 50 groups of samples with known defect types, and calculates the accuracy rate, which is the proportion of correctly associated samples out of the total number of test samples, to verify the correlation between the clustering results and the actual defect types.

[0030] After the evaluation, the clusters are associated with known defect types in the domain knowledge base: by calculating the similarity between the center of each cluster and the typical feature vectors of various defects in the historical defect data label library, using Euclidean distance, the defect type with the highest similarity is taken as the corresponding defect type of the cluster. For example, if the distance between the center of a cluster and the typical vector of a broken bar defect is 0.1, and the distance with the eccentric defect is 0.8, then the cluster is determined to correspond to a rotor broken bar defect. Simultaneously, the severity and location of the defect are interpreted by combining expert experience rules and physical model constraints. For example, according to the rule that "the radial vibration peak factor is positively correlated with the severity of broken bars," if the mean peak factor of the cluster samples is 6, then the defect severity is determined to be "moderate"; based on the phase of the characteristic frequency of the current signal, the defect location is determined to be "near phase B of the stator winding." Finally, a motor rotor defect diagnosis report is generated, which includes the defect type, severity, possible location, and judgment basis.

[0031] In this embodiment, data acquisition includes: The system collects vibration, current, temperature, and voltage signals from the motor in real time. The vibration signals include radial and axial vibrations; the current signals include three-phase and zero-sequence currents; the temperature signals include stator winding and bearing temperatures; and the voltage signals include three-phase and phase-to-phase voltages.

[0032] Specifically, during data acquisition, the radial vibration signal is acquired at locations such as the No. 3 bearing housing at the front end and the No. 5 bearing housing at the rear end of the motor, while the axial acquisition location is at the end face of the motor rotor output end, using a piezoelectric accelerometer with a sensitivity of 100mV / g. Current signal acquisition is achieved using a Hall effect current sensor with a range of 0-50A and an output signal of 0-5V, matching the input range of the data acquisition card. For temperature signals, the stator winding temperature sensor is embedded in the stator core slots near the winding end, and the bearing temperature sensor is attached to the outer ring surface of the bearing. The sampling interval can be adjusted according to the motor's operating conditions; for example, it is set to 0.5 seconds during startup and 1 second during stable operation. Voltage signal acquisition uses a voltage transformer with a range of 0-1000V and an output signal of 0-5V, acquired synchronously with the current signal. This multi-dimensional, high-precision signal acquisition provides comprehensive data support for subsequent feature extraction and defect diagnosis, avoiding diagnostic blind spots caused by traditional single-signal acquisition.

[0033] In this embodiment, feature extraction includes: Perform data cleaning, noise reduction, and normalization preprocessing on the motor operation data; Time-domain analysis, frequency-domain analysis, or time-frequency-domain analysis are performed on the preprocessed motor operation data to obtain time series characteristics, spectral characteristics, and wavelet packet energy characteristics. Based on time series features, spectral features, and wavelet packet energy features, feature vectors characterizing the defect state of the motor rotor are generated for knowledge-driven clustering analysis.

[0034] Specifically, in the preprocessing stage of feature extraction, data cleaning, in addition to the 3σ criterion, employs a sliding window method for smoothing periodic anomalies, such as current surges caused by power grid fluctuations; wavelet threshold selection for data denoising uses the Birgé-Massart strategy, adaptively adjusting the threshold to adapt to different signal noise levels; normalization strictly follows the min-max formula to ensure uniformity of all feature dimensions. In the analysis stage, time-domain features, besides mean and variance, also calculate impulse factors and margin factors, features sensitive to impact defects; frequency-domain analysis reduces spectral leakage by adding a Hanning window and calculates characteristic frequencies, such as those related to rotor bar breakage. The rotor eccentricity is determined by twice the rotational frequency, corresponding to the amplitude and phase. The number of wavelet packet decomposition layers in the time-frequency domain analysis is determined based on the signal frequency range; for example, for a 20kHz vibration signal, decomposition into 4 layers can cover the 1.25-20kHz frequency band. The final generated feature vector is further dimensionality-reduced using principal component analysis (PCA), for example, reducing 20 features to 10 principal components. This reduces computational load while retaining key information, ensuring that the feature vector can accurately and efficiently characterize the motor rotor defect state.

[0035] In this embodiment, constructing a domain knowledge base includes: Establish an expert experience rule base, which describes the relationship between specific feature value ranges, feature combinations, or feature change trends and specific defect types. Establish a physical model constraint library, which is based on the laws of electromagnetism, mechanical vibration or thermodynamics, and imposes restrictions on eigenvalues, eigenre relationships or defect evolution processes. Establish a historical defect data tag library, which records the verified types of motor rotor defects and their corresponding feature patterns in the feature space.

[0036] Specifically, when constructing the domain knowledge base, the rules for the expert experience rule base need to undergo multiple rounds of expert discussion and verification. For example, the initial rule "radial vibration peak factor greater than 5 indicates a broken bar defect" was adjusted to "radial vibration peak factor greater than 4.8 and current characteristic frequency" after verification with 100 sets of historical data. "Amplitudes greater than 0.5V are considered broken bar defects," improving the accuracy of the rules. The physical model constraint library needs to be dynamically adjusted based on motor design parameters. For example, for a motor with a rated power of 100kW, the slip s value range is constrained to 0.02-0.04, and the corresponding current characteristic frequency range is adjusted to 48-49Hz and 51-52Hz. The historical defect data label library needs to regularly filter invalid data, such as removing abnormal data caused by sensor failures, while supplementing new defect data, such as rotor crack defects, to ensure the completeness and timeliness of the knowledge base. Through the construction of a structured knowledge base, reliable domain knowledge support is provided for knowledge-driven clustering, avoiding the blindness of the clustering process.

[0037] In this embodiment, knowledge-driven clustering includes: The feature weighting mechanism dynamically adjusts the weight of each feature in similarity calculation based on expert experience rules or historical defect data tags stored in the domain knowledge base. This adjustment is made according to the sensitivity of different features to specific defects or their ability to distinguish known defect types, thereby highlighting features that are strongly related to motor rotor defects.

[0038] Specifically, in the feature weighting mechanism, experts assign a 10-point score to the importance of features. For example, if experts consider the current feature to be 9 points and the temperature feature to be 5 points for the importance of the broken bar defect, after min-max normalization, the current feature... Temperature characteristics Feature discrimination index Mutual information is calculated, and a higher mutual information value indicates a stronger correlation between the feature and the defect type. For example, the mutual information value for the current feature is 0.8, which, after normalization... The weighting coefficient α is adjusted according to the diagnostic scenario. For example, in high-noise conditions, α is set to 0.7 to increase the weight of expert experience and reduce noise interference; in stable conditions, α is set to 0.5 to balance expert experience and historical data. Through dynamic feature weighting, the clustering process prioritizes features strongly associated with defects, improving the matching degree between clustering results and actual defect types.

[0039] In this embodiment, knowledge-driven clustering further includes: The clustering constraint mechanism uses physical model constraints in the domain knowledge base to apply soft or hard constraints to the clustering process. Soft constraints guide the clustering direction through penalty terms, while hard constraints force the clustering results to meet preset physical boundary conditions to ensure that the formation of clusters conforms to the physical laws of motor operation.

[0040] Specifically, in the clustering constraint mechanism, the physical boundary conditions of the hard constraints need to be determined according to the motor standard. For example, according to GB755-2019 "Ratings and Performance of Rotating Electrical Machines", the maximum allowable temperature of the stator winding of a Class A insulated motor is 155℃. Therefore, the hard constraint threshold for temperature characteristics is set to 155℃. The penalty coefficient λ for the soft constraints is determined through cross-validation. For example, the clustering accuracy is tested on 5 different datasets with λ=0.3, 0.5, and 0.7, and the highest accuracy λ=0.5 is selected as the final value. When a sample violates the soft constraints, such as when the current characteristic frequency exceeds... Within the range, the objective function will add a penalty term to prompt the clustering algorithm to classify the sample into a cluster that better fits the constraints, ensuring that the clustering results conform to the physical laws of motor operation and avoiding the formation of meaningless clusters.

[0041] In this embodiment, knowledge-driven clustering further includes: The cluster center initialization guidance mechanism uses known defect patterns or expert-defined typical defect feature value ranges from historical defect data tags in the domain knowledge base to initialize the cluster centers of the clustering algorithm, thereby accelerating the convergence of the clustering algorithm and improving the quality and stability of the clustering results.

[0042] Specifically, in the cluster center initialization guidance mechanism, if the sample size of a certain type of defect in the historical defect data label library is sufficient, such as exceeding 50 groups, the mean of all feature vectors of that type of sample is calculated as the initial cluster center. For example, in the 50 samples of rotor local short circuit defects, the mean of the effective value of axial vibration is 0.15 mm, and the mean of the zero-sequence current component is 3 A. The corresponding components of the initial cluster center are 0.15 mm and 3 A, respectively. If the sample size is insufficient, such as 20 groups, experts are invited to define the range of typical defect feature values. For example, if experts define the voltage imbalance range of local short circuit defects as 2%-5%, 3.5% is taken as the voltage imbalance component of the initial cluster center. Compared with random initialization, this mechanism improves the convergence speed of the clustering algorithm by about 30%, while reducing the probability of clustering getting trapped in local optima, thus improving the quality and stability of the clustering results.

[0043] In this embodiment, the evaluation and interpretation of clustering results include: The clusters output by the knowledge-driven clustering algorithm are evaluated internally or externally. The internal evaluation uses the silhouette coefficient or Davies-Bouldin index to measure compactness and separation, while the external evaluation is compared and verified with pre-labeled test data. The evaluated clusters are associated and mapped with known defect types in the domain knowledge base, and a motor rotor defect diagnosis report is generated, providing an explanation of the defect type, severity, and possible location.

[0044] Specifically, in the clustering result evaluation and interpretation phase, the silhouette coefficient calculation for internal evaluation must cover all samples. For example, for the clustering results of 200 samples, the silhouette coefficient of each sample is calculated and the average is taken. If the average is greater than 0.6, it indicates good cluster compactness and separation. If the Davies-Bouldin index is less than 1.0, it indicates that the clustering result is reasonable. The test data for external evaluation must cover defect samples under different operating conditions, such as samples under high noise and load fluctuation conditions. If the accuracy exceeds 85%, the reliability of the clustering results is verified. When associating defect types, if a cluster has a high similarity to multiple defect types, such as a distance of 0.12 to a broken bar defect and a distance of 0.15 to an eccentric defect, further judgment is made based on physical model constraints. For example, if the current characteristic frequency of the samples in this cluster meets the constraints of a broken bar defect, it is therefore determined to be a broken bar defect. The diagnostic report adopts a visual format, such as displaying the trend of characteristic value changes through line graphs and listing the defect judgment criteria through tables, making the diagnostic results easier to understand and facilitating maintenance measures by on-site engineers.

[0045] In this embodiment, it also includes: Defect diagnosis feedback and knowledge base iterative updates: Based on the diagnosis report, the diagnosis results are confirmed through expert review or further experimental verification. Based on the confirmed diagnostic results, the historical defect data tags in the domain knowledge base are dynamically updated, and new defect patterns and feature descriptions are supplemented or corrected. Based on the newly added domain knowledge, optimize expert experience rules or adjust physical model constraints to improve the diagnostic accuracy and adaptability of knowledge-driven clustering algorithms.

[0046] Specifically, during the defect diagnosis feedback and knowledge base iteration update process, after the diagnostic report is generated, it is reviewed by at least two experts. If experts disagree on the diagnostic results—for example, one expert believes it is a broken bar defect while another believes it is an eccentricity defect—then a motor disassembly experiment is conducted for verification. For instance, if disassembling the motor reveals two broken bars on the rotor, the diagnostic result is confirmed as a broken bar defect, and the feature vector and correction label of this sample are added to the historical defect data label library. Based on the newly added historical data, the feature discrimination index is recalculated. For example, after adding 50 sets of broken bar defect data, the current characteristics... The feature weights were adjusted accordingly, increasing the value from 0.8 to 0.85. Simultaneously, expert experience rules were revised; for example, "current characteristic frequency amplitude greater than 0.5V" was changed to "greater than 0.45V," making the rules more closely aligned with actual data. Through continuous iterative updates, the accuracy and adaptability of the knowledge base are constantly improved, thereby gradually increasing the diagnostic accuracy of the knowledge-driven clustering algorithm and better adapting to the needs of diagnosing rotor defects in motors of different models and operating conditions.

[0047] This invention also provides a knowledge-driven clustering-based collaborative analysis system for motor rotor defects, applicable to any of the aforementioned knowledge-driven clustering-based collaborative analysis methods for motor rotor defects, including: The data acquisition module is used to collect motor operating data in real time; The feature extraction module is used to preprocess and extract features from the motor operation data to generate feature vectors; The knowledge base construction module is used to build a structured domain knowledge base, which includes expert experience rules, physical model constraints, and historical defect data labels. The clustering analysis module is used to design and implement a knowledge-driven clustering algorithm that utilizes information from the domain knowledge base for guidance and optimization during the clustering process. The evaluation and interpretation module is used to evaluate the clustering results output by the clustering analysis module and to associate and interpret the clusters with known defect types in the domain knowledge base.

[0048] Specifically, the data acquisition module consists of a piezoelectric accelerometer, a Hall current sensor, a PT100 temperature sensor, a voltage transformer, and a multi-channel data acquisition card. The acquisition card connects to an industrial computer via Ethernet to achieve real-time data transmission and storage. The feature extraction module uses Python to write the algorithm, integrating data cleaning, noise reduction, normalization, and feature calculation functions. The program runs on the industrial computer and can automatically read the raw data in the database and output feature vectors. The knowledge base construction module uses a MySQL relational database and designs expert experience rule tables, physical model constraint tables, and historical defect data tables, supporting data addition, deletion, modification, and query operations. It also develops a knowledge base management interface to facilitate expert knowledge maintenance. The clustering analysis module integrates an improved K-means algorithm into a C++ program. The program calls data from the knowledge base to perform feature weighting, cluster constraints, and cluster center initialization, and outputs the clustering results. The evaluation and interpretation module uses Tableau to develop a visual reporting system, reads the clustering results and knowledge base data, automatically generates defect diagnosis reports, and supports report export and printing. The entire system automates and coordinates the diagnosis of motor rotor defects, reduces manual intervention, improves diagnostic efficiency, and can be widely applied to online monitoring and fault diagnosis scenarios for industrial motors.

[0049] In summary, this invention avoids the incomplete information problem of traditional single-signal acquisition by real-time acquisition of multi-dimensional operating data such as motor vibration, current, temperature, and voltage. Preprocessing such as data cleaning and noise reduction, along with time-domain, frequency-domain, and time-frequency-domain analysis, extracts accurate feature vectors to solve the problem of noise interference in the original data. Simultaneously, it constructs a domain knowledge base containing expert experience rules, physical model constraints, and historical defect data labels. Combined with knowledge-driven clustering's feature weighting, clustering constraints, and cluster center initialization mechanisms, it overcomes the problems of existing clustering methods that do not fully utilize domain knowledge, lack physical meaning in the results, and have poor correlation with actual defects. Furthermore, it improves diagnostic accuracy and adaptability through cluster evaluation interpretation and iterative updates of the knowledge base, achieving automated collaborative analysis of motor rotor defects, reducing manual intervention, and solving the problems of insufficient accuracy and efficiency in diagnosing complex operating conditions using traditional methods. This makes it suitable for practical industrial scenarios.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A collaborative analysis method for motor rotor defects based on knowledge-driven clustering, characterized in that, include: S1. Data acquisition: Real-time acquisition of motor operation data; S2. Feature extraction: Preprocess and extract features from the motor operation data to generate feature vectors related to motor rotor defects. S3. Construct a domain knowledge base and establish a structured domain knowledge base, which includes expert experience rules, physical model constraints, and historical defect data labels; S4. Knowledge-driven clustering: Design and implement a knowledge-driven clustering algorithm that utilizes information from the domain knowledge base for guidance and optimization during the clustering process. S5. Clustering result evaluation and interpretation: Evaluate the clustering results output by the knowledge-driven clustering algorithm, and associate and interpret the clusters with known defect types in the domain knowledge base.

2. The method for collaborative analysis of motor rotor defects based on knowledge-driven clustering according to claim 1, characterized in that, The data collection includes: The system collects vibration, current, temperature, and voltage signals from the motor in real time. The vibration signals include radial and axial vibrations; the current signals include three-phase and zero-sequence currents; the temperature signals include stator winding and bearing temperatures; and the voltage signals include three-phase and phase-to-phase voltages.

3. The method for collaborative analysis of motor rotor defects based on knowledge-driven clustering according to claim 1, characterized in that, The feature extraction includes: Perform data cleaning, noise reduction, and normalization preprocessing on the motor operation data; Time-domain analysis, frequency-domain analysis, or time-frequency-domain analysis are performed on the preprocessed motor operation data to obtain time series characteristics, spectral characteristics, and wavelet packet energy characteristics. Based on time series features, spectral features, and wavelet packet energy features, feature vectors characterizing the defect state of the motor rotor are generated for knowledge-driven clustering analysis.

4. The method for collaborative analysis of motor rotor defects based on knowledge-driven clustering according to claim 1, characterized in that, The construction of the domain knowledge base includes: Establish an expert experience rule base, which describes the relationship between specific feature value ranges, feature combinations, or feature change trends and specific defect types. Establish a physical model constraint library, which is based on the laws of electromagnetism, mechanical vibration or thermodynamics, and imposes restrictions on eigenvalues, eigenre relationships or defect evolution processes. Establish a historical defect data tag library, which records the verified types of motor rotor defects and their corresponding feature patterns in the feature space.

5. The method for collaborative analysis of motor rotor defects based on knowledge-driven clustering according to claim 1, characterized in that, The knowledge-driven clustering includes: The feature weighting mechanism dynamically adjusts the weight of each feature in similarity calculation based on expert experience rules or historical defect data tags stored in the domain knowledge base. This adjustment is made according to the sensitivity of different features to specific defects or their ability to distinguish known defect types, thereby highlighting features that are strongly related to motor rotor defects.

6. The method for collaborative analysis of motor rotor defects based on knowledge-driven clustering according to claim 1, characterized in that, The knowledge-driven clustering also includes: The clustering constraint mechanism uses physical model constraints in the domain knowledge base to apply soft or hard constraints to the clustering process. Soft constraints guide the clustering direction through penalty terms, while hard constraints force the clustering results to meet preset physical boundary conditions to ensure that the formation of clusters conforms to the physical laws of motor operation.

7. The method for collaborative analysis of motor rotor defects based on knowledge-driven clustering according to claim 1, characterized in that, The knowledge-driven clustering also includes: The cluster center initialization guidance mechanism uses known defect patterns or expert-defined typical defect feature value ranges from historical defect data tags in the domain knowledge base to initialize the cluster centers of the clustering algorithm, thereby accelerating the convergence of the clustering algorithm and improving the quality and stability of the clustering results.

8. The method for collaborative analysis of motor rotor defects based on knowledge-driven clustering according to claim 1, characterized in that, The evaluation and interpretation of the clustering results include: The clusters output by the knowledge-driven clustering algorithm are evaluated internally or externally. The internal evaluation uses the silhouette coefficient or Davies-Bouldin index to measure compactness and separation, while the external evaluation is compared and verified with pre-labeled test data. The evaluated clusters are associated and mapped with known defect types in the domain knowledge base, and a motor rotor defect diagnosis report is generated, providing an explanation of the defect type, severity, and possible location.

9. The method for collaborative analysis of motor rotor defects based on knowledge-driven clustering according to claim 1, characterized in that, Also includes: Defect diagnosis feedback and knowledge base iterative updates: Based on the diagnosis report, the diagnosis results are confirmed through expert review or further experimental verification. Based on the confirmed diagnostic results, the historical defect data tags in the domain knowledge base are dynamically updated, and new defect patterns and feature descriptions are supplemented or corrected. Based on the newly added domain knowledge, optimize expert experience rules or adjust physical model constraints to improve the diagnostic accuracy and adaptability of knowledge-driven clustering algorithms.

10. A knowledge-driven clustering-based collaborative analysis system for motor rotor defects, applied to the knowledge-driven clustering-based collaborative analysis method for motor rotor defects as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect motor operating data in real time; The feature extraction module is used to preprocess and extract features from the motor operation data to generate feature vectors; The knowledge base construction module is used to build a structured domain knowledge base, which includes expert experience rules, physical model constraints, and historical defect data tags. The clustering analysis module is used to design and implement a knowledge-driven clustering algorithm that utilizes information from the domain knowledge base for guidance and optimization during the clustering process. The evaluation and interpretation module is used to evaluate the clustering results output by the clustering analysis module and to associate and interpret the clusters with known defect types in the domain knowledge base.

Images