Gantry crane motor fault diagnosis system and fault diagnosis method

Abstract

The application provides a quay crane motor fault diagnosis system and a fault diagnosis method. In the fault diagnosis method, the collected vibration acceleration data and motor speed data are preprocessed to obtain standardized data, then the standardized data are trained based on a self-encoder, subsequently a fault direction vector is constructed, then a probability vector containing a health state and four types of faults is constructed by combining a feature subspace mapping and the fault direction vector, a normalization process is performed to obtain a probability distribution, a multi-level confidence assessment is performed on the motor fault to obtain a comprehensive confidence, and finally whether the motor needs to be shut down for maintenance is determined according to the probability distribution and the comprehensive confidence. The self-encoder is used to construct an anomaly detection model based on only healthy samples, and a large number of labeled fault data are not required, so that effective and timely monitoring and diagnosis of motor faults in harsh environments such as high salt fog, high humidity and strong wind can be realized.

Description

Technical Field

[0001] This invention relates to the field of motor fault diagnosis technology, and in particular to a fault diagnosis system and method for quay crane motors. Background Technology

[0002] Currently, global port trade volume continues to climb, and port operations are developing towards ultra-large scale, high frequency, and automation. As the core equipment for port loading and unloading, the motors of the hoisting, trolley, and crane mechanisms of quay cranes are the power hub driving the entire terminal logistics chain. Quay crane motors operate under extreme conditions of heavy load, frequent starting and braking, and wide-range speed regulation for extended periods. Once a motor experiences a failure such as bearing wear or rotor misalignment, leading to a shutdown, it will directly paralyze the entire operation line, resulting in huge port demurrage fees. Quay cranes are typically deployed in coastal environments with high salt spray, high humidity, and strong winds. The physical condition of the electrode system is severely corroded by the environment, and manual inspection is not only highly dangerous but also difficult to penetrate deep into the motor to detect hidden problems. Traditional post-maintenance methods can no longer meet the high-efficiency requirements of modern smart ports.

[0003] With the advancement of Industry 4.0, remote monitoring and automated diagnostic technologies have been gradually introduced into quay crane motors. However, in actual operation, existing diagnostic systems still face challenges such as a lack of fault data, which makes fault data-based models ineffective. Traditional signal processing fault identification methods heavily rely on the expert knowledge of on-site personnel. The diverse types of faults and the complex fault mechanisms severely restrict the effectiveness of fault identification. Summary of the Invention

[0004] The purpose of this invention is to provide a fault diagnosis system and method for quay crane motors, in order to solve the problem that existing motor fault diagnosis methods / systems cannot effectively and timely diagnose motor faults in harsh environments such as high salt spray, high humidity and strong winds.

[0005] To address the aforementioned technical problems, this invention provides a method for diagnosing faults in quay crane motors, comprising:

[0006] Step 1: Use the acquisition module to collect at least the following: triaxial vibration acceleration data of the motor bearing housing and motor speed data;

[0007] Step 2: Perform interference removal and optimization processing on the vibration acceleration data and the motor speed data to obtain standardized data for fault feature extraction and identification;

[0008] Step 3: Based on the autoencoder and the standardized data obtained in Step 2, construct a feature subspace for healthy motor operation to perform unsupervised fault detection of the motor.

[0009] Step 4: Based on the physical mechanism of motor faults, construct the fault direction vector to complete the mapping from data features to physical defects. Among them, motor faults are defined as four categories: bearing outer ring fault, bearing inner ring fault, rotor imbalance fault, and rotor misalignment fault.

[0010] Step 5: Combine the feature subspace mapping of motor healthy operation and the fault direction vector to identify motor faults, construct a probability vector containing the healthy state and four types of faults, perform normalization to obtain the probability distribution, and perform multi-level confidence assessment on motor faults to obtain the comprehensive confidence.

[0011] Step 6: Feed back the probability distribution and the overall confidence level to the operation and maintenance module. The operation and maintenance module determines whether the motor needs to be shut down for maintenance based on the probability distribution and the overall confidence level.

[0012] Optionally, in the quay crane motor fault diagnosis method, step 2 includes:

[0013] Step 2.1: Establish a rotor frequency multiplication reference based on motor speed data, and set theoretical fault characteristic frequencies in combination with bearing structural parameters. The theoretical fault characteristic frequencies include at least: inner ring theoretical fault frequency BPFI and outer ring theoretical fault frequency BPFO.

[0014] Step 2.2: Denoise the vibration acceleration data and convert the vibration acceleration data from a time domain signal to a frequency domain signal;

[0015] Step 2.3: Based on the vibration acceleration data converted into frequency domain signals and the rotor frequency harmonics, extract multi-domain features and construct a fault feature vector. The fault feature vector is an 11-dimensional feature vector, and the 11 features are F1-F11, where F1-F4 are the outer ring theoretical fault frequency BPFO fault features, F5-F7 are the inner ring theoretical fault frequency BPFI fault features, and F8-F11 are the rotor fault features.

[0016] Step 2.4: Perform Z-score standardization on the fault feature vector to obtain the mean μ and standard deviation σ.

[0017] Optionally, in the fault diagnosis method for the quay crane motor, in the 11-dimensional feature vector, F1 is the sum of amplitude values ​​from 1st to 3rd harmonics, F2 is the ratio of F1 to the total energy of the envelope spectrum, F3 is the ratio of the fault sideband to the energy of the envelope spectrum, F4 is the kurtosis of the envelope spectrum, F5 is the sum of amplitude values ​​from 1st to 3rd harmonics, F6 is the ratio of F5 to the total energy of the envelope spectrum, F7 is the ratio of the fault sideband to the energy of the envelope spectrum, F8 is the ratio of 1st rotational frequency to local energy, F9 is the ratio of 2nd rotational frequency to local energy, F10 is the ratio of 2nd rotational frequency to 1st rotational frequency, and F11 is the low-frequency energy of the rotational frequency.

[0018] Optionally, in the quay crane motor fault diagnosis method, step 3 includes:

[0019] Step 3.1: Collect feature data under normal motor operation conditions and construct a health training set in the form of an M×N matrix, where M is the number of samples and N is the dimension of the feature vector;

[0020] Step 3.2: Calculate the mean vector μ_std and standard deviation vector σ_std of the healthy training set through Z-score standardization to establish a standardized benchmark;

[0021] Step 3.3: Construct a three-layer autoencoder network structure for autoencoder training, wherein the input layer is N-dimensional, the hidden layer is M-dimensional, and the output layer is N-dimensional;

[0022] Step 3.4: Reconstruct the healthy samples in the health training set using the trained autoencoder model to obtain reconstructed samples. Obtain the reconstruction residual matrix based on the reconstructed samples. Calculate the reconstruction error norm for each sample to obtain the error distribution of the healthy samples. The mean μ_res and standard deviation σ_res of the reconstruction error of the healthy samples are used as the benchmark for motor fault judgment.

[0023] Step 3.5: Obtain the squared prediction error of healthy samples based on the reconstruction error norm of each sample;

[0024] Step 3.6: Based on the squared prediction error distribution of healthy samples, the Welch approximation method is used to obtain the upper limit of the squared prediction error control.

[0025] Optionally, in the fault diagnosis method for quay crane motors, in step 3.3, the following training parameters are used to optimize the autoencoder network structure: the maximum number of training rounds is 200 rounds; the L2 weight regularization coefficient is 0.004 to prevent overfitting; the sparsity regularization coefficient is 4; and the sparsity ratio is 0.15, so that the hidden layer learns key health patterns. The decoder uses a linear transfer function to ensure reconstruction accuracy.

[0026] Optionally, in the quay crane motor fault diagnosis method, step 4 includes:

[0027] Step 4.1: Assign sensitivity weights to each type of fault in the 11-dimensional feature space and construct a 4×11 fault direction vector matrix, where 4 represents the four types of faults and 11 represents the 11 types of features;

[0028] Step 4.2: Perform L2 normalization on each fault direction vector to ensure the comparability of weights for each type of fault and eliminate the influence of dimensions.

[0029] Optionally, in the quay crane motor fault diagnosis method, step 5 includes:

[0030] Step 5.1: Obtain the 1×11 feature vector data_test of the real-time collected test sample, and normalize the feature vector data_test based on the standardized parameters of the healthy sample.

[0031] Step 5.2: Input the standardized test sample into the autoencoder model trained in step S3 to obtain the reconstructed sample data_reconstructed_test;

[0032] Step 5.3: Based on the normalized feature vector and the reconstructed sample data_reconstructed_test, obtain the reconstructed residual vector r, where the reconstructed residual vector r represents the direction and degree of the test sample deviating from the feature subspace of the healthy sample;

[0033] Step 5.4: Based on the reconstructed residual vector r, obtain the residual norm res_norm_test, and based on the residual norm res_norm_test, obtain the squared prediction error SPE_test; where the residual norm res_norm_test =||r||2; and the squared prediction error SPE_test = res_norm_test².

[0034] Step 5.4: Based on the Gaussian distribution assumption, calculate the confidence level that the test sample belongs to a healthy state. The confidence level that the test sample belongs to a healthy state reflects the similarity between the test sample and the healthy pattern. The closer the confidence level is to 1, the healthier the sample is.

[0035] Step 5.5: Calculate the cosine similarity between the reconstructed residual vector r and each fault direction vector FDV, extract the positively correlated part of the cosine similarity as the fault score, and quantitatively evaluate the matching degree of each type of motor fault. The closer the cosine similarity is to 1, the more consistent the reconstructed residual vector r is with the corresponding fault direction vector, and the higher the probability of the fault.

[0036] Step 5.6: Construct a probability vector containing the health status and four types of faults, and perform normalization to obtain the probability distribution;

[0037] Step 5.7: Set up a multi-level confidence assessment system and weight and fuse the multi-level confidence scores, and judge the credibility of the diagnostic results based on the weighted and fused multi-level confidence scores;

[0038] The confidence assessment system includes at least the following levels:

[0039] Level 1: Confidence level of the squared prediction error statistic. The significance of anomalies is determined based on the upper control limit of the squared prediction error. Level 1 reflects the statistical significance of the test sample deviating from the feature subspace of the healthy sample.

[0040] Level 2: Classification entropy confidence score, which calculates the probability distribution entropy. The smaller the entropy, the more certain the classification result and the higher the confidence score.

[0041] Level 3: Competition gap confidence. The probability vectors are sorted in descending order. The gap between the highest probability and the second highest probability is calculated and normalized to obtain the competition gap confidence. The larger the gap, the higher the differentiation between the diagnostic result and the competing category, and the lower the risk of misjudgment.

[0042] The formula for weighted fusion multi-level confidence is: overall_conf = a1 * squared prediction error statistic confidence + a2 * classification entropy confidence + a3 * competition interval confidence, where a1, a2, and a3 are weight coefficients, and a1 + a2 + a3 = 1.

[0043] When the overall confidence level (overall_conf) is greater than or equal to 0.85, the diagnostic results are considered highly reliable and can be directly adopted.

[0044] When 0.65 ≤ overall confidence level < 0.85, the diagnostic result is considered moderately reliable, and manual review is recommended.

[0045] When the overall confidence level (overall_conf) is less than 0.65, the diagnostic result is considered to be of low confidence and requires further examination and / or resampling.

[0046] Optionally, in the above method for diagnosing faults in quay crane motors, the formula for normalizing the feature vector data_test based on the standardized parameters of healthy samples is as follows:

[0047] data_test_norm = (data_test-μ_std) / σ_std;

[0048] Wherein, data_test_norm is the normalized feature vector, data_test is the 1×11 feature vector of the real-time collected test sample, μ_std is the mean vector of the healthy training set, and σ_std is the standard deviation vector of the healthy training set.

[0049] Optionally, in the quay crane motor fault diagnosis method, the formula for calculating the reconstruction residual vector r based on the normalized feature vector and the reconstructed sample data_reconstructed_test is as follows:

[0050] r = data_test_norm - data_reconstructed_test;

[0051] Where r is the reconstructed residual vector, data_test_norm is the normalized feature vector, and data_reconstructed_test is the reconstructed sample output by the autoencoder model.

[0052] Optionally, in the aforementioned quay crane motor fault diagnosis method, based on the Gaussian distribution assumption, the formula for calculating the confidence level that a test sample belongs to a healthy state is as follows:

[0053] conf_health=exp(-(res_norm_test-μ_res)² / (2σ_res²));

[0054] Where conf_healt is the confidence level that the test sample belongs to the healthy state, res_norm_test is the residual norm, μ_res is the mean of the healthy state residuals, and σ_res is the standard deviation of the healthy state residuals.

[0055] Optionally, in the quay crane motor fault diagnosis method, in step 5.5, the formula for calculating the cosine similarity between the reconstructed residual vector r and each fault direction vector is:

[0056] cosine_similarity(f) = (r·FDV(f,:)) / (||r||×||FDV(f,:)||);

[0057] Where cosine_similarity(f) is the cosine similarity between the reconstructed residual vector r and each fault direction vector FDV; r is the reconstructed residual vector; FDV(f,:) is each fault direction vector;

[0058] The formula for extracting the positive correlation component of cosine similarity as the fault score is: fault_score(f) = max(0, cosine_similarity(f)).

[0059] Here, fault_score(f) is the fault score.

[0060] Optionally, in the quay crane motor fault diagnosis method, step 6 includes:

[0061] The probability distribution and the overall confidence level (overall_conf) are fed back to the operation and maintenance module.

[0062] If the overall confidence level (overall_conf) is greater than or equal to 0.85 and the failure probability is greater than 80%, an alarm should be triggered immediately, and a shutdown for maintenance should be arranged to prevent the motor failure from escalating.

[0063] If 0.65 ≤ overall confidence level < 0.85 and the failure probability > 60%, mark the motor as a key monitoring target, shorten the inspection cycle of the corresponding motor, and continuously track the trend;

[0064] If the overall confidence level (overall_conf) is less than 0.65 or the health probability is greater than 70%, the corresponding motor will be kept running normally, and the corresponding motor will be monitored regularly.

[0065] On the other hand, the present invention also provides a quay crane motor fault diagnosis system, comprising:

[0066] The data acquisition module is used to acquire at least the following: triaxial vibration acceleration data of the motor bearing housing and motor speed data;

[0067] The data preprocessing module is used to remove interference and optimize the vibration acceleration data and the motor speed data to obtain standardized data for fault feature extraction and identification.

[0068] The autoencoder data training module is used to construct a feature subspace for healthy motor operation based on the standardized data, so as to perform unsupervised fault detection of the motor.

[0069] The physical mechanism mapping module constructs a fault direction vector based on the physical mechanism of motor faults, and completes the mapping from data features to physical defects. Motor faults are defined as four types: bearing outer ring fault, bearing inner ring fault, rotor imbalance fault, and rotor misalignment fault.

[0070] The fault identification, processing and evaluation module is used to identify motor faults by combining the feature subspace mapping of motor healthy operation and the fault direction vector, construct a probability vector containing the healthy state and four types of faults, perform normalization processing to obtain the probability distribution, and perform multi-level confidence evaluation of motor faults to obtain the comprehensive confidence.

[0071] The operation and maintenance module is used to receive the probability distribution and the comprehensive confidence level output by the fault identification and evaluation module, and to determine whether the motor needs to be shut down for maintenance based on the probability distribution and the comprehensive confidence level.

[0072] In summary, (1) the fault diagnosis method provided in this application first preprocesses the collected vibration acceleration data and motor speed data to obtain standardized data, then trains the standardized data based on the autoencoder, then constructs a fault direction vector, and then combines the feature subspace mapping and the fault direction vector to construct a probability vector containing the healthy state and four types of faults, performs normalization processing to obtain the probability distribution, and performs multi-level confidence assessment on the motor fault to obtain the comprehensive confidence. Finally, based on the probability distribution and the comprehensive confidence, it is determined whether the motor needs to be shut down for maintenance. This application constructs an anomaly detection model based only on healthy samples through the autoencoder, which can realize effective and timely monitoring and diagnosis of motor faults in harsh environments such as high salt spray, high humidity and strong wind without a large amount of labeled fault data, solving the problems of difficult acquisition and high cost of motor fault samples in the prior art.

[0073] (2) This application constructs an anomaly detection model based on healthy samples through an autoencoder and adopts an unsupervised learning mechanism, which can also identify novel or compound faults that have not been seen before, significantly improving the generalization and engineering usability of the fault diagnosis system.

[0074] (3) This application introduces the Fault Direction Vector (FDV) mechanism to establish a quantitative mapping relationship between abstract data features and specific physical defects (damage to the outer ring of the motor bearing, damage to the inner ring of the motor bearing, rotor imbalance, and rotor misalignment). By analyzing the cosine similarity, the fault type and its confidence level can be clarified. This can get rid of the limitations of black box classification in existing motor fault diagnosis technology. The diagnosis results not only give the fault category, but also point out the specific location, providing a physical basis for precise maintenance, realizing physical mechanism mapping, improving the interpretability of diagnosis, and solving the pain points of existing motor fault diagnosis technology being difficult to interpret and blind in maintenance.

[0075] (4) This application constructs a multi-level confidence system to reduce the false alarm rate. Specifically, this application establishes a three-level evaluation mechanism: confidence of squared prediction error (SPE) statistic, confidence of classification entropy and confidence of competition interval, and outputs a comprehensive confidence by weighted fusion. By quantitatively characterizing the reliability of the diagnostic conclusion, it effectively solves the problem of high false alarm rate caused by single index judgment in existing motor fault diagnosis technology, and ensures that maintenance decision is triggered only when a high confidence result is output, which significantly improves the robustness and reliability of the fault diagnosis system in the complex coastal environment.

[0076] (5) This application constructs a complete closed-loop mechanism of data acquisition-preprocessing-autoencoder-based anomaly detection-mechanism mapping-confidence assessment-feedback verification-model optimization, realizing closed-loop management throughout the entire life cycle, improving the level of intelligence, realizing continuous tracking and adaptive learning of motor health status, and fully considering the complex scenarios of multiple working conditions and multiple fault types of quay crane motors. Through configurable FDV weights, adjustable confidence threshold ranges, and expandable fault categories, it adapts to the needs of different motor models, different operating modes, and different maintenance strategies. The physical mechanism basis of the fault direction vector in this application ensures the portability of the fault diagnosis method between different devices, reduces the cost of remodeling, and solves the problems of poor adaptability and weak generalization ability of existing motor fault diagnosis technologies. Attached Figure Description

[0077] Those skilled in the art will understand that the accompanying drawings are provided to better understand the invention and do not constitute any limitation on the scope of the invention.

[0078] Figure 1 This is a flowchart of the fault diagnosis method for quay crane motors according to an embodiment of the present invention;

[0079] Figure 2 This is a system architecture diagram of the quay crane motor fault diagnosis method according to an embodiment of the present invention;

[0080] Figure 3 This is a distribution trend diagram of the squared prediction error (SPE) according to an embodiment of the present invention;

[0081] Figure 4 This is a schematic diagram of the feature residual contribution of an embodiment of the present invention;

[0082] Figure 5 This is a schematic diagram of the failure probability distribution according to an embodiment of the present invention;

[0083] Figure 6 This is a schematic diagram illustrating the three levels of confidence assessment in an embodiment of the present invention;

[0084] Figure 7 This is a schematic diagram of the structure of the quay crane motor fault diagnosis system according to an embodiment of the present invention;

[0085] The reference numerals in the attached figures are explained as follows:

[0086] 10 - Acquisition module, 20 - Data preprocessing module, 30 - Autoencoder, 40 - Physical mechanism mapping module, 50 - Fault identification, processing and evaluation module, 60 - Operation and maintenance module. Detailed Implementation

[0087] To make the objectives, advantages, and features of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the drawings are all in a very simplified form and are not drawn to scale, and are only used to facilitate and clearly illustrate the objectives of the embodiments of the present invention. Furthermore, the structures shown in the drawings are often part of the actual structures. In particular, different figures may emphasize different aspects and sometimes use different scales. It should also be understood that, unless specifically stated or indicated, the terms "first," "second," "third," etc., in the specification are only used to distinguish the various components, elements, steps, etc., in the specification, and are not used to indicate the logical or sequential relationships between the various components, elements, steps, etc.

[0088] This invention provides a method for diagnosing faults in quay crane motors, with reference to... Figure 1 and Figure 2 , Figure 1 This is a flowchart of a fault diagnosis method for quay crane motors according to an embodiment of the present invention. Figure 2 This is a system architecture diagram of a quay crane motor fault diagnosis method according to an embodiment of the present invention. The quay crane motor fault diagnosis method includes:

[0089] First, perform step 1 (S1): Use the acquisition module to acquire at least the following: triaxial vibration acceleration data of the motor bearing housing and motor speed data.

[0090] Then, step 2 (S2) is performed: the vibration acceleration data and the motor speed data are subjected to interference removal and optimization processing to obtain standardized data for fault feature extraction and identification.

[0091] Preferably, step 2 may specifically include:

[0092] Step 2.1: Establish a rotor frequency multiplication reference based on motor speed data, and set theoretical fault characteristic frequencies in combination with bearing structural parameters. The theoretical fault characteristic frequencies include at least: inner ring theoretical fault frequency BPFI and outer ring theoretical fault frequency BPFO.

[0093] Step 2.2: Denoise the vibration acceleration data and convert the vibration acceleration data from a time domain signal to a frequency domain signal;

[0094] Step 2.3: Based on the vibration acceleration data converted into frequency domain signals and the rotor frequency harmonics, extract multi-domain features and construct a fault feature vector. The fault feature vector is an 11-dimensional feature vector, and the 11 features are F1-F11, where F1-F4 are the outer ring theoretical fault frequency BPFO fault features, F5-F7 are the inner ring theoretical fault frequency BPFI fault features, and F8-F11 are the rotor fault features.

[0095] Step 2.4: Perform Z-score standardization on the fault feature vector to obtain the mean μ and standard deviation σ.

[0096] In step 2.3, among the 11-dimensional feature vectors, F1 is the sum of amplitudes from 1st to 3rd harmonics, F2 is the ratio of F1 to the total energy of the envelope spectrum, F3 is the ratio of the fault sideband to the energy of the envelope spectrum, F4 is the kurtosis of the envelope spectrum, F5 is the sum of amplitudes from 1st to 3rd harmonics, F6 is the ratio of F5 to the total energy of the envelope spectrum, F7 is the ratio of the fault sideband to the energy of the envelope spectrum, F8 is the ratio of 1st harmonic frequency to local energy, F9 is the ratio of 2nd harmonic frequency to local energy, F10 is the ratio of 2nd harmonic frequency to 1st harmonic frequency, and F11 is the low-frequency energy of the harmonic frequency.

[0097] In this embodiment, 100 sets of 11-dimensional feature data of the motor under normal operating conditions can be extracted from the health sample data.

[0098] Next, step 3 (S3) is executed: based on the autoencoder and the standardized data obtained in step 2, a feature subspace for motor healthy operation is constructed to perform unsupervised fault detection of the motor.

[0099] Preferably, step 3 may specifically include:

[0100] Step 3.1: Collect feature data under normal motor operation conditions and construct a health training set in the form of an M×N matrix, where M is the number of samples and N is the dimension of the feature vector;

[0101] Step 3.2: Calculate the mean vector μ_std and standard deviation vector σ_std of the healthy training set through Z-score standardization to establish a standardized benchmark;

[0102] Step 3.3: Construct a three-layer autoencoder network structure for autoencoder training, wherein the input layer is N-dimensional, the hidden layer is M-dimensional, and the output layer is N-dimensional;

[0103] Step 3.4: Reconstruct the healthy samples in the healthy training set using the trained autoencoder model to obtain reconstructed samples. Obtain the reconstruction residual matrix based on the reconstructed samples. Calculate the reconstruction error norm for each sample to obtain the error distribution of the healthy samples. The mean μ_res and standard deviation σ_res of the reconstruction error of the healthy samples are used as the benchmark for motor fault judgment. The formula for calculating the reconstruction error norm for each sample is: Reconstruction error norm res_norm = ||R||2; where R is the reconstruction residual matrix.

[0104] Step 3.5: Obtain the squared prediction error of healthy samples based on the reconstruction error norm of each sample;

[0105] Step 3.6: Based on the squared prediction error distribution of healthy samples, the upper control limit (UCL) of the squared prediction error (SPE) is obtained using the Welch approximation method.

[0106] In this embodiment, the Welch approximation method is used to fit the squared prediction error (SPE) distribution of healthy samples, and the control upper limit (UCL) of the squared prediction error (SPE) at 95% and 99% confidence levels is calculated.

[0107] In step 3.3, the following training parameters are used to optimize the autoencoder network structure: the maximum number of training rounds is 200; the L2 weight regularization coefficient is 0.004 to prevent overfitting; the sparsity regularization coefficient is 4; and the sparsity ratio is 0.15 so that the hidden layer learns key health patterns. The decoder uses a linear transfer function to ensure reconstruction accuracy.

[0108] In this application, an anomaly detection model is constructed based on healthy samples using an autoencoder. An unsupervised learning mechanism is adopted, which also enables the identification of novel or complex faults that have not been seen before, significantly improving the generalization and engineering usability of the fault diagnosis system.

[0109] Further, step 4 (S4) is executed: based on the physical mechanism of motor faults, a fault direction vector is constructed to complete the mapping from data features to physical defects. Among them, motor faults are defined as four types: bearing outer ring faults, bearing inner ring faults, rotor imbalance faults, and rotor misalignment faults.

[0110] Preferably, step 4 may specifically include:

[0111] Step 4.1: Assign sensitivity weights to each type of fault in the 11-dimensional feature space and construct a 4×11 fault direction vector matrix, where 4 represents the four types of faults and 11 represents the 11 types of features;

[0112] Step 4.2: Perform L2 normalization on each fault direction vector to ensure the comparability of weights for each type of fault and eliminate the influence of dimensions.

[0113] The calculation formula for L2 normalization of each fault direction vector is: FDV(f,:) = FDV(f,:) / ||FDV(f,:)||2.

[0114] Next, step 5 (S5) is executed: combining the feature subspace mapping of motor healthy operation and the fault direction vector, motor faults are identified, a probability vector containing the healthy state and four types of faults is constructed, the probability distribution is obtained by normalization, and a multi-level confidence assessment of motor faults is performed to obtain a comprehensive confidence level.

[0115] Step 5 includes:

[0116] Step 5.1: Obtain the 1×11 feature vector data_test of the real-time collected test sample of the motor, and normalize the feature vector data_test based on the standardized parameters of the healthy sample.

[0117] The formula for normalizing the feature vector data_test based on the standardized parameters of healthy samples is as follows:

[0118] data_test_norm = (data_test-μ_std) / σ_std;

[0119] Wherein, data_test_norm is the normalized feature vector, data_test is the 1×11 feature vector of the real-time collected test sample, μ_std is the mean vector of the healthy training set, and σ_std is the standard deviation vector of the healthy training set.

[0120] Step 5.2: Input the standardized test samples into the autoencoder model trained in step S3 to obtain the reconstructed samples data_reconstructed_test; wherein, the formula for calculating the reconstructed residual vector r is:

[0121] r = data_test_norm - data_reconstructed_test;

[0122] Where r is the reconstructed residual vector, data_test_norm is the normalized feature vector, and data_reconstructed_test is the reconstructed sample output by the autoencoder model.

[0123] Step 5.3: Based on the normalized feature vector and the reconstructed sample data_reconstructed_test, obtain the reconstructed residual vector r, where the reconstructed residual vector r represents the direction and degree of the test sample deviating from the feature subspace of the healthy sample;

[0124] Step 5.4: Based on the reconstructed residual vector r, obtain the residual norm res_norm_test, and based on the residual norm res_norm_test, obtain the squared prediction error SPE_test; where the residual norm res_norm_test =||r||2; and the squared prediction error SPE_test = res_norm_test².

[0125] Step 5.4: Based on the Gaussian distribution assumption, calculate the confidence level that the test sample belongs to a healthy state. The confidence level reflects the similarity between the test sample and a healthy pattern; a confidence level closer to 1 indicates a healthier state. The formula for calculating the confidence level that the test sample belongs to a healthy state based on the Gaussian distribution assumption is as follows:

[0126] conf_health=exp(-(res_norm_test-μ_res)² / (2σ_res²));

[0127] Where conf_healt is the confidence level that the test sample belongs to the healthy state, res_norm_test is the residual norm, μ_res is the mean of the healthy state residuals, and σ_res is the standard deviation of the healthy state residuals.

[0128] Step 5.5: Calculate the cosine similarity between the reconstructed residual vector r and each fault direction vector FDV, extract the positively correlated part of the cosine similarity as the fault score, and quantitatively evaluate the matching degree of each type of motor fault. The closer the cosine similarity is to 1, the more consistent the reconstructed residual vector r is with the corresponding fault direction vector, and the higher the probability of the fault.

[0129] In step 5.5, the formula for calculating the cosine similarity between the reconstructed residual vector r and each fault direction vector is as follows:

[0130] cosine_similarity(f) = (r·FDV(f,:)) / (||r||×||FDV(f,:)||);

[0131] Where cosine_similarity(f) is the cosine similarity between the reconstructed residual vector r and each fault direction vector FDV; r is the reconstructed residual vector; FDV(f,:) is each fault direction vector;

[0132] The formula for extracting the positive correlation component of cosine similarity as the fault score is: fault_score(f) = max(0, cosine_similarity(f)).

[0133] Here, fault_score(f) is the fault score.

[0134] Step 5.6: Construct a probability vector containing the health status and four types of faults, and perform normalization to obtain the probability distribution, where,

[0135] The probability vector total_scores contains the health status and four types of faults: total_scores = [conf_health;fault_score(1);fault_score(2);fault_score(3);fault_score(4)]; where conf_health is the confidence that the test sample belongs to the health status, and fault_score(1)-fault_score(4) are the fault scores of the four types of faults, respectively.

[0136] Furthermore, the formula for normalizing this probability vector is as follows:

[0137] P(i) = total_scores(i) / ;

[0138] Where P(i) represents the probability of the healthy state and the four types of fault states, respectively, and the integer i takes the value of 1 to 5; total_scores(i) and total_scores(j) correspond to one of the five components / elements in the probability vector of the healthy state and the four types of fault states.

[0139] Step 5.7: Set up a multi-level confidence assessment system and weight and fuse the multi-level confidence scores, and judge the credibility of the diagnostic results based on the weighted and fused multi-level confidence scores;

[0140] The confidence assessment system includes at least the following levels:

[0141] Level 1: Confidence level of the squared prediction error statistic. The significance of anomalies is determined based on the upper control limit of the squared prediction error. Level 1 reflects the statistical significance of the test sample deviating from the feature subspace of the healthy sample.

[0142] Level 2: Classification entropy confidence score, which calculates the probability distribution entropy. The smaller the entropy, the more certain the classification result and the higher the confidence score.

[0143] Level 3: Competition gap confidence. The probability vectors are sorted in descending order. The gap between the highest probability and the second highest probability is calculated and normalized to obtain the competition gap confidence. The larger the gap, the higher the differentiation between the diagnostic result and the competing category, and the lower the risk of misjudgment.

[0144] The formula for weighted fusion multi-level confidence is: overall_conf = a1 * squared prediction error statistic confidence + a2 * classification entropy confidence + a3 * competition interval confidence, where a1, a2, and a3 are weight coefficients, and a1 + a2 + a3 = 1.

[0145] When the overall confidence level (overall_conf) is greater than or equal to 0.85, the diagnostic results are considered highly reliable and can be directly adopted.

[0146] When 0.65 ≤ overall confidence level < 0.85, the diagnostic result is considered moderately reliable, and manual review is recommended.

[0147] When the overall confidence level (overall_conf) is less than 0.65, the diagnostic result is considered to be of low confidence and requires further examination and / or resampling.

[0148] refer to Figure 3 and Figure 4 , Figure 3 This is a distribution trend diagram of the squared prediction error (SPE) according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the feature residual contribution of an embodiment of the present invention. Figure 3 The SPE trend chart shows that most points are below the upper control limit (UCL) of the squared prediction error (SPE) at a 95% confidence level, demonstrating extremely high reconstruction accuracy. Figure 4 In the characteristic residual contribution diagram, features F1-F4 show significant residual energy, which highly matches the outer ring fault feature fingerprint. Combined with reference... Figure 5 , Figure 5 This is a schematic diagram of the fault probability distribution according to an embodiment of the present invention, showing that the identification probability of faults in the outer ring is over 90%. (Reference) Figure 6 , Figure 6 This is a schematic diagram of the three levels of confidence assessment in an embodiment of the present invention. In the confidence assessment, L1-SPE (squared prediction error statistic confidence), L2-entropy (classification entropy confidence), and L3-distance (competitive distance confidence) are obtained respectively, and the final comprehensive confidence is output by weighted fusion.

[0149] This application introduces a Fault Direction Vector (FDV) mechanism to establish a quantitative mapping relationship between abstract data features and specific physical defects (damage to the outer race of motor bearings, damage to the inner race of motor bearings, rotor imbalance, and rotor misalignment). By using cosine similarity analysis to clarify the fault type and its confidence level, this approach overcomes the limitations of black-box classification in existing motor fault diagnosis technologies. The diagnostic results not only provide the fault category but also pinpoint the specific location, providing a physical basis for precise maintenance, achieving a mapping of physical mechanisms, improving diagnostic interpretability, and addressing the pain points of existing motor fault diagnosis technologies, such as difficulty in interpretation and blind maintenance. Furthermore, this application constructs a multi-level confidence system to reduce the false alarm rate. Specifically, it establishes a three-layer evaluation mechanism: confidence level of the squared prediction error (SPE) statistic, confidence level of the classification entropy, and confidence level of the competition gap. A weighted fusion is used to output a comprehensive confidence level, quantitatively characterizing the reliability of the diagnostic conclusion. This effectively solves the problem of high false alarm rates caused by single-indicator judgment in existing motor fault diagnosis technologies, ensuring that maintenance decisions are triggered only when high-confidence results are output, significantly improving the robustness and reliability of the fault diagnosis system in complex coastal environments.

[0150] Finally, step 6 (S6) is executed: the probability distribution and the overall confidence level are fed back to the operation and maintenance module, which determines whether the motor needs to be shut down for maintenance based on the probability distribution and the overall confidence level.

[0151] Specifically, step 6 may include:

[0152] The probability distribution and the overall confidence level (overall_conf) are fed back to the operation and maintenance module.

[0153] If the overall confidence level (overall_conf) is greater than or equal to 0.85 and the failure probability is greater than 80%, an alarm should be triggered immediately, and a shutdown for maintenance should be arranged to prevent the motor failure from escalating.

[0154] If 0.65 ≤ overall confidence level < 0.85 and the failure probability > 60%, mark the motor as a key monitoring target, shorten the inspection cycle of the corresponding motor, and continuously track the trend;

[0155] If the overall confidence level (overall_conf) is less than 0.65 or the health probability is greater than 70%, the corresponding motor will be kept running normally, and the corresponding motor will be monitored regularly.

[0156] This application constructs a complete closed-loop mechanism: data acquisition, preprocessing, autoencoder-based anomaly detection, mechanism mapping, confidence assessment, feedback verification, and model optimization. This mechanism enables closed-loop management throughout the entire lifecycle, improves intelligence, and achieves continuous tracking and adaptive learning of motor health status. It fully considers the complex scenarios of multiple operating conditions and fault types in quay crane motors, adapting to different motor models, operating modes, and maintenance strategies through configurable FDV weights, adjustable confidence threshold ranges, and expandable fault categories. The physical mechanism of the fault direction vector in this application ensures the transferability of the fault diagnosis method across different devices, reduces remodeling costs, and solves the problems of poor adaptability and weak generalization ability in existing motor fault diagnosis technologies.

[0157] In summary, the fault diagnosis method provided in this application first preprocesses the collected vibration acceleration data and motor speed data to obtain standardized data. Then, it trains the standardized data based on an autoencoder to construct a fault direction vector. Next, it combines the feature subspace mapping and the fault direction vector to construct a probability vector containing healthy states and four types of faults. This vector is then normalized to obtain a probability distribution. Furthermore, a multi-level confidence assessment of the motor fault is performed to obtain a comprehensive confidence score. Finally, based on the probability distribution and the comprehensive confidence score, it is determined whether the motor needs to be shut down for maintenance. This application constructs an anomaly detection model based solely on healthy samples using an autoencoder, enabling effective and timely monitoring and diagnosis of motor faults in harsh environments such as high salt spray, high humidity, and strong winds without requiring a large amount of labeled fault data. This solves the problems of difficult and costly acquisition of motor fault samples in existing technologies.

[0158] Based on the same inventive concept, this invention also provides a quay crane motor fault diagnosis system, see reference. Figure 7 , Figure 7 This is a schematic diagram of the structure of a quay crane motor fault diagnosis system according to an embodiment of the present invention. The quay crane motor fault diagnosis system includes:

[0159] The acquisition module 10 is used to acquire at least: triaxial vibration acceleration data of the motor bearing housing and motor speed data;

[0160] The data preprocessing module 20 is used to perform interference removal and optimization processing on the vibration acceleration data and the motor speed data to obtain standardized data for fault feature extraction and identification.

[0161] The autoencoder data training module 30 is used to construct a feature subspace for healthy motor operation based on the standardized data, so as to perform unsupervised fault detection of the motor.

[0162] The physical mechanism mapping module 40 constructs a fault direction vector based on the physical mechanism of motor faults, and completes the mapping from data features to physical defects. Among them, motor faults are defined as four types: bearing outer ring fault, bearing inner ring fault, rotor imbalance fault and rotor misalignment fault.

[0163] The fault identification, processing and evaluation module 50 is used to identify motor faults by combining the feature subspace mapping of motor healthy operation and the fault direction vector, construct a probability vector containing the healthy state and four types of faults, perform normalization processing to obtain the probability distribution, and perform multi-level confidence evaluation of motor faults to obtain a comprehensive confidence level.

[0164] The operation and maintenance module 60 is used to receive the probability distribution and the comprehensive confidence level output by the fault identification and evaluation module, and to determine whether the motor needs to be shut down for maintenance based on the probability distribution and the comprehensive confidence level.

[0165] Furthermore, it should be understood that although the present invention has been disclosed above with reference to preferred embodiments, these embodiments are not intended to limit the present invention. For any person skilled in the art, many possible variations and modifications can be made to the technical solutions of the present invention based on the disclosed technical content, or equivalent embodiments can be modified accordingly, without departing from the scope of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention, without departing from the content of the present invention, shall still fall within the scope of protection of the present invention.

Claims

1. A method for diagnosing faults in quay crane motors, characterized in that, include: Step 1: Use the acquisition module to collect at least the following: triaxial vibration acceleration data of the motor bearing housing and motor speed data; Step 2: Perform interference removal and optimization processing on the vibration acceleration data and the motor speed data to obtain standardized data for fault feature extraction and identification; Step 3: Based on the autoencoder and the standardized data obtained in Step 2, construct a feature subspace for healthy motor operation to perform unsupervised fault detection of the motor. Step 4: Based on the physical mechanism of motor faults, construct the fault direction vector to complete the mapping from data features to physical defects. Among them, motor faults are defined as four categories: bearing outer ring fault, bearing inner ring fault, rotor imbalance fault, and rotor misalignment fault. Step 5: Combine the feature subspace mapping of motor healthy operation and the fault direction vector to identify motor faults, construct a probability vector containing the healthy state and four types of faults, perform normalization to obtain the probability distribution, and perform multi-level confidence assessment on motor faults to obtain the comprehensive confidence. Step 6: Feed back the probability distribution and the overall confidence level to the operation and maintenance module. The operation and maintenance module determines whether the motor needs to be shut down for maintenance based on the probability distribution and the overall confidence level.

2. The method for diagnosing faults in quay crane motors according to claim 1, characterized in that, Step 2 includes: Step 2.1: Establish a rotor frequency multiplication reference based on motor speed data, and set theoretical fault characteristic frequencies in combination with bearing structural parameters. The theoretical fault characteristic frequencies include at least: inner ring theoretical fault frequency BPFI and outer ring theoretical fault frequency BPFO. Step 2.2: Denoise the vibration acceleration data and convert the vibration acceleration data from a time domain signal to a frequency domain signal; Step 2.3: Based on the vibration acceleration data converted into frequency domain signals and the rotor frequency harmonics, extract multi-domain features and construct a fault feature vector. The fault feature vector is an 11-dimensional feature vector, and the 11 features are F1-F11, where F1-F4 are the outer ring theoretical fault frequency BPFO fault features, F5-F7 are the inner ring theoretical fault frequency BPFI fault features, and F8-F11 are the rotor fault features. Step 2.4: Perform Z-score standardization on the fault feature vector to obtain the mean μ and standard deviation σ.

3. The method for diagnosing faults in quay crane motors according to claim 2, characterized in that, In the 11-dimensional eigenvector, F1 is the sum of amplitudes from 1st to 3rd harmonics, F2 is the ratio of F1 to the total energy of the envelope spectrum, F3 is the ratio of the fault sideband to the energy of the envelope spectrum, F4 is the kurtosis of the envelope spectrum, F5 is the sum of amplitudes from 1st to 3rd harmonics, F6 is the ratio of F5 to the total energy of the envelope spectrum, F7 is the ratio of the fault sideband to the energy of the envelope spectrum, F8 is the ratio of 1st harmonic frequency to local energy, F9 is the ratio of 2nd harmonic frequency to local energy, F10 is the ratio of 2nd harmonic frequency to 1st harmonic frequency, and F11 is the low-frequency energy of the harmonic frequency.

4. The method for diagnosing faults in quay crane motors according to claim 2, characterized in that, Step 3 includes: Step 3.1: Collect feature data under normal motor operation conditions and construct a health training set in the form of an M×N matrix, where M is the number of samples and N is the dimension of the feature vector; Step 3.2: Calculate the mean vector μ_std and standard deviation vector σ_std of the healthy training set through Z-score standardization to establish a standardized benchmark; Step 3.3: Construct a three-layer autoencoder network structure for autoencoder training, wherein the input layer is N-dimensional, the hidden layer is M-dimensional, and the output layer is N-dimensional; Step 3.4: Reconstruct the healthy samples in the health training set using the trained autoencoder model to obtain reconstructed samples. Obtain the reconstruction residual matrix based on the reconstructed samples. Calculate the reconstruction error norm for each sample to obtain the error distribution of the healthy samples. The mean μ_res and standard deviation σ_res of the reconstruction error of the healthy samples are used as the benchmark for motor fault judgment. Step 3.5: Obtain the squared prediction error of healthy samples based on the reconstruction error norm of each sample; Step 3.6: Based on the squared prediction error distribution of healthy samples, the Welch approximation method is used to obtain the upper limit of the squared prediction error control.

5. The method for diagnosing faults in quay crane motors according to claim 4, characterized in that, In step 3.3, the following training parameters are used to optimize the autoencoder network structure: the maximum number of training epochs is 200; the L2 weight regularization coefficient is 0.004 to prevent overfitting. The sparsity regularization coefficient is 4; The sparsity ratio is 0.15 to enable the hidden layer to learn key health patterns, wherein the decoder uses a linear transfer function to ensure reconstruction accuracy.

6. The method for diagnosing faults in quay crane motors according to claim 4, characterized in that, Step 4 includes: Step 4.1: Assign sensitivity weights to each type of fault in the 11-dimensional feature space and construct a 4×11 fault direction vector matrix, where 4 represents the four types of faults and 11 represents the 11 types of features; Step 4.2: Perform L2 normalization on each fault direction vector to ensure the comparability of weights for each type of fault and eliminate the influence of dimensions.

7. The method for diagnosing faults in quay crane motors according to claim 6, characterized in that, Step 5 includes: Step 5.1: Obtain the 1×11 feature vector data_test of the real-time collected test sample, and normalize the feature vector data_test based on the standardized parameters of the healthy sample. Step 5.2: Input the standardized test sample into the autoencoder model trained in step S3 to obtain the reconstructed sample data_reconstructed_test; Step 5.3: Based on the normalized feature vector and the reconstructed sample data_reconstructed_test, obtain the reconstructed residual vector r, where the reconstructed residual vector r represents the direction and degree of the test sample deviating from the feature subspace of the healthy sample; Step 5.4: Based on the reconstructed residual vector r, obtain the residual norm res_norm_test, and based on the residual norm res_norm_test, obtain the squared prediction error SPE_test; where the residual norm res_norm_test =||r||2; and the squared prediction error SPE_test = res_norm_test². Step 5.4: Based on the Gaussian distribution assumption, calculate the confidence level that the test sample belongs to a healthy state. The confidence level that the test sample belongs to a healthy state reflects the similarity between the test sample and the healthy pattern. The closer the confidence level is to 1, the healthier the sample is. Step 5.5: Calculate the cosine similarity between the reconstructed residual vector r and each fault direction vector FDV, extract the positively correlated part of the cosine similarity as the fault score, and quantitatively evaluate the matching degree of each type of motor fault. The closer the cosine similarity is to 1, the more consistent the reconstructed residual vector r is with the corresponding fault direction vector, and the higher the probability of the fault. Step 5.6: Construct a probability vector containing the health status and four types of faults, and perform normalization to obtain the probability distribution; Step 5.7: Set up a multi-level confidence assessment system and weight and fuse the multi-level confidence scores, and judge the credibility of the diagnostic results based on the weighted and fused multi-level confidence scores; The confidence assessment system includes at least the following levels: Level 1: Confidence level of the squared prediction error statistic. The significance of anomalies is determined based on the upper control limit of the squared prediction error. Level 1 reflects the statistical significance of the test sample deviating from the feature subspace of the healthy sample. Level 2: Classification entropy confidence score, which calculates the probability distribution entropy. The smaller the entropy, the more certain the classification result and the higher the confidence score. Level 3: Competition gap confidence. The probability vectors are sorted in descending order. The gap between the highest probability and the second highest probability is calculated and normalized to obtain the competition gap confidence. The larger the gap, the higher the differentiation between the diagnostic result and the competing category, and the lower the risk of misjudgment. The formula for weighted fusion multi-level confidence is: overall_conf = a1 * squared prediction error statistic confidence + a2 * classification entropy confidence + a3 * competition interval confidence, where a1, a2, and a3 are weight coefficients, and a1 + a2 + a3 = 1. When the overall confidence level (overall_conf) is greater than or equal to 0.85, the diagnostic results are considered highly reliable and can be directly adopted. When 0.65 ≤ overall confidence level < 0.85, the diagnostic result is considered moderately reliable, and manual review is recommended. When the overall confidence level (overall_conf) is less than 0.65, the diagnostic result is considered to be of low confidence and requires further examination and / or resampling.

8. The method for diagnosing faults in quay crane motors according to claim 7, characterized in that, The formula for normalizing the feature vector data_test based on the standardized parameters of healthy samples is as follows: data_test_norm = (data_test-μ_std) / σ_std; Wherein, data_test_norm is the normalized feature vector, data_test is the 1×11 feature vector of the real-time collected test sample, μ_std is the mean vector of the healthy training set, and σ_std is the standard deviation vector of the healthy training set.

9. The method for diagnosing faults in quay crane motors according to claim 7, characterized in that, Based on the normalized feature vector and the reconstructed sample data_reconstructed_test, the formula for calculating the reconstructed residual vector r is as follows: r = data_test_norm - data_reconstructed_test; Where r is the reconstructed residual vector, data_test_norm is the normalized feature vector, and data_reconstructed_test is the reconstructed sample output by the autoencoder model.

10. The method for diagnosing faults in quay crane motors according to claim 7, characterized in that, Based on the Gaussian distribution assumption, the formula for calculating the confidence level that a test sample belongs to a healthy state is as follows: conf_health=exp(-(res_norm_test-μ_res)² / (2σ_res²)); Where conf_health is the confidence level that the test sample belongs to the healthy state, res_norm_test is the residual norm, μ_res is the mean of the healthy state residuals, and σ_res is the standard deviation of the healthy state residuals.

11. The method for diagnosing faults in quay crane motors according to claim 7, characterized in that, In step 5.5, the formula for calculating the cosine similarity between the reconstructed residual vector r and each fault direction vector is as follows: cosine_similarity(f) = (r·FDV(f,:)) / (||r||×||FDV(f,:)||); Where cosine_similarity(f) is the cosine similarity between the reconstructed residual vector r and each fault direction vector FDV; r is the reconstructed residual vector; FDV(f,:) is each fault direction vector; The formula for extracting the positive correlation component of cosine similarity as the fault score is: fault_score(f) = max(0, cosine_similarity(f)). Here, fault_score(f) is the fault score.

12. The method for diagnosing faults in quay crane motors according to claim 7, characterized in that, Step 6 includes: The probability distribution and the overall confidence level (overall_conf) are fed back to the operation and maintenance module. If the overall confidence level (overall_conf) is greater than or equal to 0.85 and the failure probability is greater than 80%, an alarm should be triggered immediately, and a shutdown for maintenance should be arranged to prevent the motor failure from escalating. If 0.65 ≤ overall confidence level < 0.85 and the failure probability > 60%, mark the motor as a key monitoring target, shorten the inspection cycle of the corresponding motor, and continuously track the trend; If the overall confidence level (overall_conf) is less than 0.65 or the health probability is greater than 70%, the corresponding motor will be kept running normally, and the corresponding motor will be monitored regularly.

13. A fault diagnosis system for quay crane motors, characterized in that, include: The data acquisition module is used to acquire at least the following: triaxial vibration acceleration data of the motor bearing housing and motor speed data; The data preprocessing module is used to remove interference and optimize the vibration acceleration data and the motor speed data to obtain standardized data for fault feature extraction and identification. The autoencoder data training module is used to construct a feature subspace for healthy motor operation based on the standardized data, so as to perform unsupervised fault detection of the motor. The physical mechanism mapping module constructs a fault direction vector based on the physical mechanism of motor faults, and completes the mapping from data features to physical defects. Motor faults are defined as four types: bearing outer ring fault, bearing inner ring fault, rotor imbalance fault, and rotor misalignment fault. The fault identification, processing and evaluation module is used to identify motor faults by combining the feature subspace mapping of motor healthy operation and the fault direction vector, construct a probability vector containing the healthy state and four types of faults, perform normalization processing to obtain the probability distribution, and perform multi-level confidence evaluation of motor faults to obtain the comprehensive confidence. The operation and maintenance module is used to receive the probability distribution and the comprehensive confidence level output by the fault identification and evaluation module, and to determine whether the motor needs to be shut down for maintenance based on the probability distribution and the comprehensive confidence level.

Images