Conveying line motor residual life prediction method and system based on health indexes

Abstract

The invention belongs to the technical field of equipment management, and particularly discloses a conveyor line motor residual life prediction method and system based on health indicators, and the method comprises the following steps: obtaining a low-frequency vibration signal of a motor; calculating a time domain feature sequence and a probability density function feature sequence; dividing a motor by using a K-means clustering algorithm to obtain a plurality of equipment groups; establishing a health index sub-model for the motors in the equipment group; performing mutual testing on the health index sub-models in the equipment group, performing smoothing processing and trend screening on health index testing results of all motors, and performing similarity integration to obtain a health index sequence; and predicting the residual life of the motor by using the health index model. According to the method, a composite feature extraction system and a dynamic integrated modeling framework oriented to a low-frequency sampling scene are constructed, the feature representation capability of field data can be improved, and the technical problem of high equipment group degradation isomerism is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment management, and particularly relates to a method and system for predicting the remaining life of a conveyor line motor based on health indicators. Background Art

[0002] Driven by the development of industry and the construction of smart grids, equipment health management technology has become a key technology for ensuring production safety and reducing operation and maintenance costs. As a core link of equipment health management technology, remaining life prediction equipment realizes pre-fault warning by analyzing equipment degradation data, and promotes the transformation of maintenance strategies from "repair after failure" to "predictive maintenance". Currently, remaining life prediction technology is widely used in fields such as aerospace, manufacturing, and power systems.

[0003] From the perspective of the prediction path, remaining life prediction can be divided into two categories: direct prediction and indirect prediction. The direct prediction method adopts an end-to-end architecture, and realizes the mapping from raw sensor data to the remaining service life through complex models such as deep neural networks. Although this method can effectively process high-dimensional non-linear features and has the advantage of adaptive learning, its performance highly depends on a large amount of training data, and there are inherent limitations such as insufficient model interpretability and high sensitivity to data noise. In contrast, the indirect prediction method follows a two-stage modeling path from "degradation characterization" to "trend prediction": first, a health indicator reflecting the degradation state of the equipment is constructed, and then a degradation trend prediction model is established based on the degradation law of the health indicator to indirectly predict the remaining life.

[0004] Existing health indicator modeling methods usually target industrial equipment with a full life cycle. The conveyor line drive motors in actual industrial sites are usually in continuous operation. Due to the lack of prior information such as missing equipment service duration records and insufficient historical working condition data, the equipment group shows significant individual degradation characterization differences. Traditional methods cannot effectively adapt to equipment groups with highly heterogeneous characteristics, seriously affecting the reliability of predictive maintenance decisions. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for predicting the remaining life of a conveyor line motor based on health indicators, which can improve the feature representation ability of on-site data, ensure the adaptability of the model to equipment individual differences, and improve the reliability of health indicator prediction.

[0006] To achieve the above purpose, the present invention is implemented by adopting the following technical solutions: According to one aspect of the present invention, a method for predicting the remaining life of a conveyor line motor based on health indicators is provided, including the following steps: Obtain the low-frequency vibration signal of the motor; Calculate the time-domain feature sequence based on the low-frequency vibration signal; determine the amplitude information according to the low-frequency vibration signal, and obtain the probability density function feature sequence of the amplitude information by using the kernel density estimation method; and determine the monotonicity and trend results of the time-domain feature sequence and the probability density function feature sequence; Use the K-means clustering algorithm to divide the motors according to the monotonicity and trend results, and obtain several equipment groups; Use a multi-layer perceptron model to establish a health index sub-model for each motor in the equipment group according to the time-domain feature sequence and the probability density function feature sequence; The health index sub-models within the equipment group are tested with each other to determine the trend of each health index test result. According to the trend, the health index test results that meet the preset conditions are screened out, and the similarity integration is performed on the screened health index test results to obtain the health index sequence; Use the pre-trained ARIMA model (Autoregressive Integrated Moving Average model) to predict the remaining life of the motor according to the health index sequence.

[0007] Adopt the above technical solution to extract the features of the low-frequency vibration signal of the industrial conveyor line motor from two aspects: one is based on the traditional time-domain statistical features, and the other is the probability density function distribution feature of the amplitude information obtained by the kernel density estimation method. In this way, a composite feature extraction system and a dynamic integration modeling framework for the low-frequency sampling scenario are constructed, aiming at the core pain point of the limited data acquisition frequency in the industrial field, solving the problem of insufficient feature representation ability under the low-frequency sampling conditions in the industrial field, and significantly improving the feature representation ability of the low-frequency vibration signal.

[0008] Moreover, a multi-layer perceptron sub-model is constructed within the group of similar devices, effectively avoiding the inherent defects of the traditional single model in cross-device generalization ability. More prominently, by introducing the sub-model mutual test mechanism and the dynamic smoothing technology, combined with the weighted integration strategy, the fusion of multi-source prediction results is realized, ensuring the adaptability of the model to the individual differences of the devices, and significantly improving the reliability of the health index prediction through the group wisdom effect, which has outstanding advantages in data noise suppression and model generalization ability improvement.

[0009] A further improvement of the present invention is that the time-domain feature sequence is calculated according to the original vibration signal information, including: peak value, mean value, variance, kurtosis, skewness and root mean square; The probability density function feature sequence includes: information entropy, information entropy increment, peak position, full width at half maximum, 25% quantile feature, 50% quantile feature, 75% quantile feature, cumulative distribution offset, mean of the probability density function change rate, variance of the probability density function change rate, maximum value of the probability density function change rate, and minimum value of the probability density function change rate.

[0010] Through the dual-path feature extraction strategy that fuses time-domain statistical features and probability density function distribution features, it breaks through the dependence of traditional time-frequency analysis methods on high-frequency signals and improves the feature representation ability of low-frequency vibration signals. Especially in the stage of probability density function feature extraction, by designing a feature engineering system including 12 multi-dimensional statistics such as information entropy increment and cumulative distribution function offset, it effectively captures the microscopic change law of the amplitude distribution during the equipment degradation process.

[0011] A further improvement of the present invention is that in the step of calculating the time-domain feature sequence from the original data and obtaining the probability density function feature sequence of the amplitude information by using the kernel density estimation method, it further includes: Determine the monotonicity and trend of each feature in the time-domain feature sequence and the probability density function feature sequence; The monotonicity of the time-domain feature sequence and the probability density function feature sequence is calculated according to the following formula:

[0012] The trend of the time-domain feature sequence and the probability density function feature sequence is calculated according to the following formula:

[0013] In the formula, represents the feature index sequence composed of the time-domain feature sequence and the probability density function feature sequence, represents the feature index sequence is the first-order difference sequence of, and N is the length of the feature index sequence.

[0014] Screen out the time-series features and probability density function features that meet the preset conditions according to the monotonicity and trend, and the screened time-series features and probability density function features are merged to obtain an excellent feature sequence; Perform dimensionality reduction processing on the excellent feature sequence to obtain a feature index sequence.

[0015] Further, the step of performing dimensionality reduction processing on the excellent feature sequence to obtain a feature index sequence includes the following steps: Calculate the covariance matrix of the excellent feature sequence; Perform eigenvalue decomposition on the covariance matrix to obtain an eigenvalue diagonal matrix and the corresponding orthogonal eigenvector matrix. Each orthogonal eigenvector matrix defines a principal component direction; Construct the principal components according to the principal component directions defined by the orthogonal eigenvector matrix to obtain the feature matrix after dimensionality reduction, which is the feature index sequence.

[0016] Combined with the dimensionality reduction technology of principal component analysis, while retaining the core features related to degradation, eliminate the interference of redundant noise, and provide feature inputs with high robustness for subsequent modeling.

[0017] A further improvement of the present invention is that the health index sub-model includes: an input layer, a plurality of fully connected hidden layers, and an output layer; The dimension of the input layer is equal to the input dimension of the feature index sequence; The output layer is a single-neuron linear output; The plurality of fully connected hidden layers introduce the ReLU (Rectified Linear Unit) activation function.

[0018] A further improvement of the present invention is that the health index sub-models within the device group perform mutual testing to determine the trend of each health index test result, screen out the health index test results that meet the preset conditions according to the trend, and perform similarity integration on the screened health index test results to obtain the health index sequence. The steps include the following steps: Use the health index sub-model of each motor in the device group to predict the health indexes of the remaining motors in the current device group according to the feature index sequences of each motor in the device group to obtain the health index test results; Through Kalman filtering, perform dynamic smoothing processing on the test results of the health indexes; Calculate the trend of the health index test results and screen the health indexes according to the trend; Use the reciprocal of the Euclidean distance between feature vectors as the similarity quantization index, and integrate the health index test results that meet the trend threshold conditions according to the similarity to obtain the health index sequence.

[0019] Among them, the reciprocal of the Euclidean distance between feature vectors is calculated according to the following formula:

[0020] The integration weight coefficient is calculated according to the following formula:

[0021] In the formula, represents the Euclidean distance between the feature vectors of motor i and motor j.

[0022] The sub-model mutual testing mechanism and the Kalman filter dynamic smoothing technology are introduced, combined with the weighted integration strategy based on the Euclidean distance similarity, to achieve the fusion of multi-source prediction results and ensure the adaptability of the model to the individual differences of motors.

[0023] A further improvement of the present invention lies in that, in the step of predicting the remaining life of the motor according to the health index sequence by using the pre-trained ARIMA model, First, model order determination: Use the health index sequences of each motor in the equipment group to train the corresponding ARIMA model, and select the optimal ARIMA model according to the Akaike information criterion; The ARIMA model consists of autoregression, differencing, and moving average, corresponding to the parameters (p, d, q) respectively. The ADF test is used to test whether the obtained health index sequence is stationary.

[0024] Then, model prediction: Use the optimal ARIMA model to obtain the health index prediction curve of the motor according to the health index sequence of the corresponding motor in the equipment group, and predict the remaining life of the motor through a preset failure threshold. Extend the health index prediction curve downward to the moment of the preset health index failure threshold to obtain the failure moment (End of Life, EoL) of the motor. Then, the remaining life at the observation end time t of the motor is the difference between EoL and the current time: 。

[0025] According to one aspect of the present invention, a device for predicting the remaining life of a conveyor line motor based on health indicators is provided, including: A data acquisition module for acquiring the low-frequency vibration signal of the motor; A feature extraction module for calculating the time-domain feature sequence according to the low-frequency vibration signal; determining the amplitude information according to the low-frequency vibration signal, and obtaining the probability density function feature sequence of the amplitude information by using the kernel density estimation method; and determining the monotonicity and trend results of the time-domain feature sequence and the probability density function feature sequence; A clustering and grouping module for using the K-means clustering algorithm to divide the motors according to the monotonicity and trend results to obtain several equipment groups; A health index construction module for using a multi-layer perceptron model to establish a health index sub-model for each motor in the equipment group according to the time-domain feature sequence and the probability density function feature sequence; mutually test the health index sub-models in the equipment group, determine the trend of each health index test result, screen out the health index test results that meet the preset conditions according to the trend, and perform similarity integration on the screened health index test results to obtain a health index sequence; A remaining life prediction module for predicting the remaining life of the motor according to the health index sequence by using the pre-trained ARIMA model.

[0026] According to one aspect of the present invention, there is provided a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, any one of the above-mentioned conveyor line motor remaining life prediction methods based on health indicators is implemented.

[0027] According to one aspect of the present invention, there is provided a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, any one of the above-mentioned conveyor line motor remaining life prediction methods based on health indicators is implemented.

[0028] Compared with the prior art, the present invention at least includes the following beneficial effects: 1. The conveyor line motor remaining life prediction method based on health indicators provided by the present invention creatively constructs a composite feature extraction system and a dynamic integration modeling framework for low-frequency sampling scenarios. Aiming at the core pain point of limited data acquisition frequency in industrial fields, through a dual-path feature extraction strategy that combines time-domain statistical features and probability density function distribution features, it breaks through the dependence of traditional time-frequency analysis methods on high-frequency signals and significantly improves the feature representation ability of low-frequency vibration signals. Especially in the probability density function feature extraction stage, by designing a feature engineering system including 12 multi-dimensional statistics such as information entropy increment and cumulative distribution function offset, it effectively captures the microscopic change law of amplitude distribution during the equipment degradation process. Further combined with the principal component analysis dimensionality reduction technology, while retaining the core features related to degradation, it eliminates redundant noise interference and provides high-robustness feature input for subsequent modeling.

[0029] 2. The integrated health indicator modeling mechanism based on equipment clustering proposed by the present invention innovatively solves the technical problem of strong heterogeneity in the group degradation of equipment. Through the K-means clustering algorithm, it realizes the adaptive grouping of equipment degradation states, and constructs a multi-layer perceptron sub-model within the same type of equipment group, effectively avoiding the inherent defects of traditional single models in cross-device generalization ability. More prominently, by introducing a sub-model mutual test mechanism and a Kalman filter dynamic smoothing technology, combined with a weighted integration strategy based on Euclidean distance similarity, it realizes the fusion of multi-source prediction results, ensures the adaptability of the model to individual differences of equipment, and significantly improves the reliability of health indicator prediction through the group wisdom effect, having outstanding advantages in data noise suppression and model generalization ability improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The specification drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings: Figure 1This is a flowchart of a method for predicting the remaining useful life of a conveyor line motor based on health indicators according to the present invention. Specific embodiments

[0031] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments. It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0032] The following detailed descriptions are all exemplary descriptions, aiming to provide further detailed descriptions of the present invention. Unless otherwise specified, all technical terms used in the present invention have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs. The terms used in the present invention are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention.

[0033] Embodiment 1 A method for predicting the remaining useful life of a conveyor line motor based on health indicators, as Figure 1 shown, includes the following steps: S1. Feature extraction Extract the features of the low-frequency vibration signal of the industrial conveyor line motor from two aspects, including obtaining the time-domain feature sequence based on the traditional time-domain statistical feature extraction, and obtaining the probability density function feature sequence of the amplitude information based on the kernel density estimation method; Among them, the obtaining of the time-domain feature sequence includes extracting 6 time-domain feature statistics from the original vibration signal, which are: peak value , mean value , variance , kurtosis , skewness , root mean square , where is the i-th data point on a certain working day, . Assume that there are currently M motors, and obtain the time-domain feature sequence m of the -th motor, is the m -th time-domain feature vector of the i -th motor, and the length of the vector is the number of sampling days.

[0034] The probability density function feature sequence is obtained based on the kernel density estimation method. Specifically, the kernel density estimation method is used to estimate the probability density function of the sample data from a finite data sample. By assigning a kernel function to each data point and then adding all the kernel functions together, the probability density of the entire data set is obtained. The expression of kernel density estimation is as follows:

[0035] Among them, is the estimated density value at position x, n is the total number of data points on a certain working day, is the i th data point, h is the bandwidth (smoothing parameter), which determines the width of the kernel function; K represents the kernel function, and the Gaussian kernel is selected here:

[0036] After constructing the probability density function for each working day using the kernel density estimation method, 12 statistical feature quantities of 7 categories that characterize the distribution characteristics of the function are extracted, specifically including: (1) Information entropy, which measures the amount and complexity of the distribution information. The probability density values at N discrete points are obtained using the kernel density estimation method , and the calculation formula for information entropy is:

[0037] where, is the probability density value at the i th discrete point, and Δ x is the spacing between discrete points; (2) Information entropy increment, which represents the entropy difference of the probability density function in different intervals and is used to measure the distribution complexity in different regions. The calculation formula for the information entropy increment is as follows:

[0038] where, is the total number of intervals, is the information entropy boundary index, is the i th interval width; (3) Peak position, corresponding to the mode, that is, the one with the highest frequency in the data, and it is one of the important indicators describing the central tendency of the data. Especially in an asymmetric distribution, it can reflect the typical value of the data better than the mean and median. Its mathematical expression is as follows:

[0039] (4) Full width at half maximum, which is the width of the probability density function at half of the peak height, that is, the horizontal distance from the left half peak to the right half peak, and is used to describe the degree of distribution concentration. The formula is as follows:

[0040] (5) Quantile feature, integrating the probability density function to obtain the cumulative distribution function:

[0041] Extract the 25th percentile feature, 50th percentile feature, and 75th percentile feature to describe the overall distribution of the data. The definition of percentile is as follows:

[0042] Among them, 。

[0043] (6) Cumulative distribution offset. Taking the cumulative distribution function (CDF) on a certain reference date (such as the first day) as a reference, then calculate the root mean square error of the subsequent date's cumulative distribution relative to the reference to quantify the overall offset of the cumulative distribution function curve and reflect the relative stability or drift trend of the distribution;

[0044] Among them, is the CDF on the first day.

[0045] (7) Probability density function change rate, which reflects the local change trend of the signal amplitude distribution. The global features extracted include, Mean of the probability density change rate:

[0046] Variance of the probability density change rate:

[0047] Maximum value of the probability density change rate:

[0048] Minimum value of the probability density change rate:

[0049] The above time-domain feature sequence and probability density function feature sequence constitute the original feature sequence ; Use monotonicity and trendiness indicators to screen the above original feature sequence, and select the feature subset that is most relevant and representative to the target task , which is the excellent feature sequence. In the process of screening the excellent feature sequence, the monotonicity and trendiness results of the original feature sequence are combined, and the monotonicity and trendiness results of each feature are sorted in descending order, and the first p statistics are selected as the excellent feature sequence. Among them, the determination of p can be adjusted according to the actual situation.

[0050] Specifically, the monotonicity formula of the feature sequence is as follows:

[0051] Among them, represents the original feature sequence , Indicates the characteristic sequence The first-order difference sequence of N is the length of the characteristic sequence; Trendiness is the time correlation of the sequence, which is the Pearson correlation coefficient between the feature vector and time t , and the specific formula is as follows:

[0052] The excellent characteristic sequences screened according to the monotonicity and trendiness indicators , and dimensionality reduction is performed on the excellent characteristic sequences to obtain the characteristic sequence , is the degenerate characteristic index sequence.

[0053] Perform dimensionality reduction on the excellent characteristic sequences using the principal component analysis method (PCA), and the specific operation process is as follows: First, perform data standardization on the original feature matrix ( n is the number of samples, p is the number of features) for standardization, so that the mean of each feature is 0 and the variance is 1, , where, is the mean vector, is the standard deviation vector; Then, calculate the covariance matrix reflecting the linear correlation between features, ; Then perform eigenvalue decomposition , where, is the eigenvalue diagonal matrix, is the corresponding orthogonal eigenvector matrix, and each eigenvector defines a principal component direction; Finally, perform principal component construction, project the data onto the first k principal components, and then obtain the dimensionality-reduced feature matrix Z, , where, is the matrix composed of the first k eigenvectors.

[0054] S2. Health index construction Use the K-means clustering algorithm to divide motors with similar degradation processes and degrees. For each motor within the same equipment group, establish a health index sub-model using a multi-layer perceptron model. The sub-models test each other, calculate the similarity weights of the outputs of each sub-model, and perform weighted integration on the multi-source test results to form a comprehensive health index. The specific steps are as follows: S2-1. Equipment clustering and grouping Use the K-means clustering algorithm to divide motors with similar degradation processes and degrees, and divide the data into K clusters (i.e.,K a device group), and minimize the sum of the squared distances from the samples within the cluster to the centroid:

[0055] where is the feature vector representing a certain motor degradation feature, and this vector is composed of the monotonicity and trend results obtained from the original feature sequence based on formulas ( ) and ( ). The specific implementation process is as follows: First, initialize the centroid by randomly selecting K data points as the centroid , where represents the centroid of the k th cluster, and d is the data dimension; Then, for each sample , calculate the squared Euclidean distance from it to all centroids, and assign it to the cluster with the closest distance, , ; where represents the set of samples in the kth cluster; Finally, recalculate the centroid of each cluster, that is, the mean value of all samples within the cluster , where is the number of samples in cluster .

[0056] Repeat the above process continuously until the centroid no longer changes or reaches the preset maximum number of iterations.

[0057] S2-2. Health Index Modeling For each motor in the device group, a health index sub-model is established using a multi-layer perceptron model. The health index sub-model consists of an input layer, multiple fully connected hidden layers, and an output layer. The health index sub-model of a certain motor in the device group can be expressed as:

[0058] where is the input feature vector, is the target health index sequence, and represent the weight matrix and bias term of the i-th layer respectively, is the non-linear activation function.

[0059] The specific training and modeling process of the multi-layer perceptron sub-model is as follows: First, perform data preparation by collecting the feature index sequences of each motor in each device group after PCA dimensionality reduction. In this process, training labels for the health indicators are set according to the operating health status of the motors. In this embodiment, the health indicator is set as a one-dimensional sequence that linearly decreases over time, and a fault occurs when the health indicator reaches the set threshold condition. The preset threshold condition can be determined according to the actual situation.

[0060] Then, perform dataset partitioning by randomly partitioning the feature index sequences of each motor in the above device group into a training set and a validation set according to a ratio of 7:3. The training set is used for training the multi-layer perceptron model, and the validation set is used for model hyperparameter tuning and model selection. Continue with model architecture setting. The model architecture of the multi-layer perceptron includes an input layer, an output layer, and hidden layers. The dimension of the input layer d is equal to the input dimension of the feature index sequence . The output layer is a single-neuron linear output corresponding to the target value . Four hidden layers are set and the ReLU activation function is introduced. The ReLU function introduces piecewise linear characteristics, enabling the multi-layer perceptron model to approximate any continuous function.

[0061] Finally, minimize the loss function through model training. The root mean square error is used as the loss function: , where, represents the training label (true value) of the health indicator, represents the predicted value of the health indicator. During the model training process, when the loss function of the validation set reaches the global minimum, the current network weights are saved to prevent model overfitting.

[0062] S2-3. Test Output This process uses the health indicator sub-models established from the data of each motor to predict the health indicators of the remaining motors in the current device group. Assume that there are k motors in the N th device group. The test output results of the n th motor contain a total of ones.

[0063] S2-4. Smoothing Processing Through Kalman filtering, dynamic smoothing of the test results of the health indicators is achieved. This method includes two stages: recursive prediction and update: In the prediction stage, based on the previous state estimate , state transition matrix and control input , predict the current state :

[0064] Based on the covariance matrix at the previous moment and the state transition matrix to predict the current covariance matrix , where

[0065] is the process noise covariance matrix: In the update stage, first calculate the Kalman gain which balances the weights of prediction and observation, where

[0066] is the observation matrix, is the measurement noise covariance: :

[0067] Update the covariance matrix to provide the basis for the next iteration:

[0068] S2-5. Test result screening Use formula (15) to set the time correlation threshold condition to screen the test results of health indicators. The preset time correlation threshold condition can be determined according to the actual situation to screen out the test results of the health indicator sub-model that can reflect the current degradation state of the motor and has strong time trend. The test result of the health indicator sub-model is a number from -1 to 1, and the threshold is a negative number; generally, if the test result of a certain health indicator of the motor is greater than the preset threshold, the corresponding health indicator is excluded; if the test result of a certain health indicator of the motor is less than or equal to the preset threshold, the corresponding health indicator is retained.

[0069] S2-6. Based on similarity integration, integrate the test results of the health indicators that meet the trend threshold condition after smoothing. Use the reciprocal of the Euclidean distance between feature vectors as the similarity quantization index:

[0070] where represents the Euclidean distance between the feature vectors of motor i and motor j . To ensure the normalization constraint of the weight coefficient, calculate the HI integrated weight coefficient in the following way:

[0071] Assume that there are N motors in the k-th equipment group, and the integrated result of the health indicators of the n -th motor is: , where is the data test result of the multi-layer perceptron sub-model of the i -th motor for the n -th motor, and is the corresponding weight coefficient.

[0072] S3. Remaining life prediction For the characteristic index sequence and the comprehensive health index, use the pre-trained ARIMA model to predict the integrated health index sequence, and reduce the health index to the set failure threshold to obtain the remaining life prediction result. Specifically, it includes the following steps: S3-1. Model order determination Use the information criterion method to realize the ARIMA model order determination. First, for each motor in each equipment group, use the health index sequence obtained by the corresponding health index sub-model to train the ARIMA model of the motor. The trained ARIMA model outputs the health index prediction value for the motor based on the health index sequence of the motor.

[0073] Specifically, the ARIMA model consists of autoregression, differencing, and moving average, corresponding to the parameters ( p , d , q ). The ADF test is used to test whether the health index sequence of the motor is stationary. The basic form of the ADF regression model is:

[0074] where , is the constant term, is the time trend term, is the unit root coefficient, and the null hypothesis (there is a unit root and the sequence is non-stationary), k is the lag order, and is the lag difference term coefficient.

[0075] Perform the ADF test on the health index sequence. If is rejected (p-value < 0.05), then the sequence is stationary; if is accepted, perform the first-order differencing and then repeat the ADF test until the null hypothesis is rejected. At this time, the number of differencing times is the parameter d ; on the differenced stationary sequence, traverse the parameter p / q The combination is used as a candidate model, and the selection is based on the Akaike information criterion:

[0076] Among them, k is the number of model parameters, L is the maximum likelihood estimate of the model. measures the goodness of fit of the model to the data, measures the penalty for the model complexity, and the ARIMA model corresponding to the minimum Akaike information is considered the optimal one; S3-2. Model prediction Based on the health index sequence of a certain motor in the equipment group, the integrated health index predicted by the optimal ARIMA model is used to obtain a prediction curve. The curve is extended and decreased to the moment of the preset health index failure threshold, and the failure moment (End of Life, EoL) of the motor is obtained. Then the remaining life at the end moment t of the motor observation is the difference between EoL and the current time: .

[0077] Example 2 A device for predicting the remaining life of a conveyor line motor based on health indicators, comprising: A data acquisition module for acquiring the low-frequency vibration signal of the motor; A feature extraction module for calculating the time-domain feature sequence according to the low-frequency vibration signal; determining the amplitude information according to the low-frequency vibration signal, and obtaining the probability density function feature sequence of the amplitude information by using the kernel density estimation method; and determining the monotonicity and trend results of the time-domain feature sequence and the probability density function feature sequence; A clustering and grouping module for dividing the motors by using the K-means clustering algorithm according to the monotonicity and trend results to obtain several equipment groups; A health index construction module for establishing a health index sub-model for each motor in the equipment group by using a multi-layer perceptron model according to the time-domain feature sequence and the probability density function feature sequence; mutually testing the health index sub-models in the equipment group, determining the trend of each health index test result, screening out the health index test results that meet the preset conditions according to the trend, and integrating the similarity of the screened health index test results to obtain a health index sequence; A remaining life prediction module for predicting the remaining life of the motor according to the health index sequence by using a pre-trained ARIMA model.

[0078] Example 3 A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, a prediction method for the remaining life of a conveyor line motor based on health indicators is implemented.

[0079] Embodiment 4 A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, a prediction method for the remaining life of a conveyor line motor based on health indicators is implemented.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent replacements can still be made to the specific implementation manners of the present invention, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered by the protection scope of the claims of the present invention.

Claims

1. A method for predicting the remaining useful life of a conveyor line motor based on health indicators, characterized in that, It includes the following steps: Obtain the low-frequency vibration signal of the motor; Calculate the time-domain feature sequence according to the low-frequency vibration signal; determine the amplitude information according to the low-frequency vibration signal, and obtain the probability density function feature sequence of the amplitude information by using the kernel density estimation method; and determine the monotonicity and trend results of the time-domain feature sequence and the probability density function feature sequence; Use the K-means clustering algorithm to divide the motors according to the monotonicity and trend results, and obtain several equipment groups; Use a multi-layer perceptron model to establish a health index sub-model for each motor in the equipment group according to the time-domain feature sequence and the probability density function feature sequence; The health index sub-models within the equipment group are mutually tested to determine the trend of each health index test result. According to the trend, the health index test results that meet the preset conditions are screened out, and the similarity integration is performed on the screened health index test results to obtain a health index sequence; Use the pre-trained ARIMA model to predict the remaining life of the motor according to the health index sequence.

2. The method for predicting the remaining life of a conveyor line motor based on health indicators according to claim 1, wherein The time-domain feature sequence is calculated according to the original vibration signal information, including: peak value, mean value, variance, kurtosis, skewness and root mean square; The probability density function feature sequence includes: information entropy, information entropy increment, peak position, full width at half maximum, 25% quantile feature, 50% quantile feature, 75% quantile feature, cumulative distribution offset, mean value of the probability density function change rate, variance of the probability density function change rate, maximum value of the probability density function change rate, and minimum value of the probability density function change rate.

3. The method for predicting the remaining life of a conveyor line motor based on health indicators according to claim 1, wherein In the step of calculating the time-domain feature sequence according to the original data and obtaining the probability density function feature sequence of the amplitude information by using the kernel density estimation method, it further includes: Determine the monotonicity and trend of each feature in the time-domain feature sequence and the probability density function feature sequence; Screen out the time series features and probability density function features that meet the preset conditions according to the monotonicity and trend, and the screened time series features and probability density function features are merged to obtain an excellent feature sequence; Perform dimensionality reduction processing on the excellent feature sequence to obtain a feature index sequence.

4. The method for predicting the remaining life of a conveyor line motor based on health indicators according to claim 3, wherein The step of performing dimensionality reduction processing on the excellent feature sequence to obtain a feature index sequence includes: Calculate the covariance matrix of the excellent feature sequence; Perform eigenvalue decomposition on the covariance matrix to obtain an eigenvalue diagonal matrix and a corresponding orthogonal eigenvector matrix. Each orthogonal eigenvector matrix defines a principal component direction; Construct principal components according to the principal component directions defined by the orthogonal eigenvector matrix to obtain a reduced-dimensional feature matrix.

5. The method for predicting the remaining life of a conveyor line motor based on health indicators according to claim 3, wherein The health index sub-model includes: an input layer, multiple fully connected hidden layers and an output layer; The dimension of the input layer is equal to the input dimension of the feature index sequence; The output layer is a single-neuron linear output; The multiple fully connected hidden layers introduce the ReLU activation function.

6. The method for predicting the remaining life of a conveyor line motor based on health indicators according to claim 3, wherein The health indicator sub-models within the equipment group perform mutual tests to determine the trend of each health indicator test result, screen out the health indicator test results that meet the preset conditions according to the trend, and perform similarity integration on the screened health indicator test results to obtain the health indicator sequence. The steps include the following: Using the health indicator sub-model of each motor in the equipment group, predicting the health indicators of the remaining motors in the current equipment group according to the feature index sequences of the motors in the equipment group to obtain the health indicator test results; Performing dynamic smoothing processing on the health indicator test results through Kalman filtering; Calculating the trend of the health indicator test results and screening the health indicators according to the trend; Using the reciprocal of the Euclidean distance between feature vectors as the similarity quantization index, integrating the health indicator test results that meet the trend threshold conditions according to similarity to obtain the health indicator sequence.

7. The method for predicting the remaining life of a conveyor line motor based on health indicators according to claim 1, wherein In the step of predicting the remaining life of the motor according to the health indicator sequence by using the pre-trained ARIMA model, Training the corresponding ARIMA model by using the health indicator sequences of the motors in the equipment group, and screening the optimal ARIMA model according to the Akaike information criterion; Using the optimal ARIMA model, obtaining the health indicator prediction curve of the motor according to the health indicator sequence of the corresponding motor in the equipment group, and predicting the remaining life of the motor through a preset failure threshold.

8. A prediction device for the remaining life of a conveyor line motor based on health indicators, characterized in that Including: A data acquisition module for acquiring the low-frequency vibration signal of the motor; A feature extraction module for calculating the time-domain feature sequence according to the low-frequency vibration signal; determining the amplitude information according to the low-frequency vibration signal, and obtaining the probability density function feature sequence of the amplitude information by using the kernel density estimation method; and determining the monotonicity and trend results of the time-domain feature sequence and the probability density function feature sequence; A clustering and grouping module for dividing the motors by using the K-means clustering algorithm according to the monotonicity and trend results to obtain several equipment groups; A health indicator construction module for using a multi-layer perceptron model to establish a health indicator sub-model for each motor in the equipment group according to the time-domain feature sequence and the probability density function feature sequence; the health indicator sub-models within the equipment group perform mutual tests to determine the trend of each health indicator test result, screen out the health indicator test results that meet the preset conditions according to the trend, and perform similarity integration on the screened health indicator test results to obtain the health indicator sequence; A remaining life prediction module for predicting the remaining life of the motor according to the health indicator sequence by using the pre-trained ARIMA model.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for predicting the remaining life of a conveyor line motor based on health indicators according to any one of claims 1-7 is implemented.

10. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the method for predicting the remaining life of a conveyor line motor based on health indicators according to any one of claims 1-7 is implemented.

Images