Equipment fault diagnosis and predictive maintenance method and system

Abstract

The invention discloses an equipment fault diagnosis and predictive maintenance method and system, and relates to the technical field of equipment detection.The equipment fault diagnosis and predictive maintenance method comprises the steps that equipment torque signal data and vibration signal data are input into a fault detection and action prediction model to output a fault diagnosis result and an action prediction result; the fault detection and action prediction model is constructed and trained as follows: torque signal data and vibration signal data of equipment are collected and preprocessed, and the preprocessed torque signal data and vibration signal data of the equipment are obtained; feature extraction and SOM clustering are carried out on the preprocessed equipment torque signal data, a fault diagnosis model is constructed to diagnose whether the equipment breaks down or not, and a fault diagnosis result of the faulty equipment is output; training a prediction model of a long and short term neural network based on quantile regression based on the preprocessed vibration signal data, and when the equipment does not fail, outputting an equipment action prediction result based on the trained prediction model; according to the equipment fault diagnosis and predictive maintenance method and system, the stability and economic benefits of the whole production line are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment detection, and in particular to a method and system for equipment fault diagnosis and predictive maintenance. Background Art

[0002] In manufacturing enterprises, mechanical equipment such as CNC machine tools, industrial robots, and automated production lines are an indispensable part of the production process. These equipment may have various faults during long-term operation due to wear, aging, or improper operation, affecting production efficiency and product quality. Therefore, real-time monitoring and fault diagnosis of these equipment is particularly important.

[0003] In practical applications, the raw data collected from the equipment often contains a lot of noise and irrelevant information. If these "dirty" data are used directly for analysis without being processed, the diagnostic results will be inaccurate. In addition, the torque signal of the servo motor of the equipment contains rich time-frequency domain information. But how to extract effective features from these complex signals is the key to fault diagnosis. However, traditional fault diagnosis methods often rely on expert experience and prior knowledge, and are difficult to adapt to complex and changeable actual working conditions.

[0004] CN201911077243.4 discloses a predictive maintenance method and predictive maintenance equipment, the method comprising: receiving first data sent by a data transmission device; classifying the first data, wherein the types of the first data include normal data, abnormal data and fault data; and establishing a prediction model using a variance algorithm and the first data. Since the establishment of a prediction model requires normal data, abnormal data and fault data, and the collection of abnormal data and fault data is usually difficult, this method is not conducive to engineering applications. In addition, in the process of establishing a prediction model using a variance algorithm and the first data, multi-dimensional feature extraction of the data is not performed, and the features used are relatively single, which cannot fully characterize the data and data characteristics, and some key data features and details may be lost.

[0005] CN202010784496.1 discloses a predictive maintenance method, device and storage medium, wherein the predictive maintenance method includes: obtaining first time series data and second time series data; performing a time series stationarity test on the second time series data and determining a predictive maintenance training model; using the predictive maintenance training model to train the first time series data and output a predicted value; and determining the point to be maintained based on the predicted value. On the one hand, this method focuses on using time series data for analysis and diagnosis, does not extract multi-dimensional features from the data, and only uses a single feature, which cannot fully reflect the data and its characteristics. On the other hand, this method does not clean the data, and cannot determine whether the equipment has failed. Determining the point to be maintained based on the predicted results may require multiple repetitive confirmations and judgments, resulting in high labor costs. Summary of the invention

[0006] Based on the technical problems existing in the background technology, the present invention proposes a method and system for equipment fault diagnosis and predictive maintenance, aiming to improve the reliability and production efficiency of equipment, reduce the downtime caused by sudden failures, extend the service life of equipment, optimize maintenance strategies, reduce maintenance costs, and improve the stability and economic benefits of the entire production line.

[0007] The present invention proposes a method for equipment fault diagnosis and predictive maintenance, which pre-processes equipment torque signal data and vibration signal data and inputs them into a fault detection and action prediction model to output fault diagnosis results and action prediction results; based on system operation and maintenance experience, equipment manufacturer maintenance strategy, fault diagnosis results and action prediction results, equipment maintenance suggestions are obtained;

[0008] The fault detection and action prediction model is constructed and trained as follows:

[0009] Step 1: collecting and preprocessing the torque signal data and the vibration signal data of the equipment to obtain the preprocessed torque signal data and the preprocessed vibration signal data of the equipment;

[0010] Step 2: Extract features from the preprocessed equipment torque signal data and perform SOM clustering to build a fault diagnosis model to diagnose whether the equipment is faulty, output the fault diagnosis results of the faulty equipment, and provide a maintenance strategy based on the fault diagnosis results;

[0011] Step 3: Based on the preprocessed vibration signal data, a prediction model of the long-term and short-term neural network based on quantile regression is trained. When the equipment does not fail, the equipment action prediction result is output based on the trained prediction model, and a maintenance strategy is given based on the prediction result.

[0012] Furthermore, in step 1, the preprocessing process of the equipment torque signal data Dataset_T0 and the vibration signal data Dataset_V0 is as follows:

[0013] Missing value processing: If the number of consecutive missing points in the samples in Dataset_T0 and Dataset_V0 is less than 5, the missing values ​​are filled with the average of the 5 data before and after the missing points; if the number of consecutive missing points in the samples is greater than or equal to 5, the sample data most similar to the sample to be processed is found according to the similarity measurement, and the data at the corresponding position is used to fill in the missing value processing equipment torque signal data Dataset_T1 and vibration signal data Dataset_V1;

[0014] Noise processing: Using a sliding average filter, apply a sliding window to the continuous data points of the samples in Dataset_T1 and Dataset_V1 to calculate the average value in turn to reduce the impact of random noise, and obtain the preprocessed equipment torque signal data Dataset_T2 and the preprocessed vibration signal data Dataset_V2.

[0015] Furthermore, in step 2, specifically:

[0016] (a1) extracting the time-frequency domain features of the preprocessed equipment torque signal data to obtain the torque feature dataset Dataset_T3 after feature extraction;

[0017] (a2) Determine the size of the SOM, initialize the weights between SOM neurons to random numbers between 0 and 1, and establish the initial neuron field radius and learning rate;

[0018] (a3) After normalizing the feature dataset Dataset_T3, the Euclidean distance is used as the discriminant function to find the input torque sample Y in the SOM neuron i The best matching node is used as the output node, and the weight of the output node satisfies: data = argmin||Y d -w (d) ||, where data is the weight, d is the number of iterations, and Y d is the torque sample of the dth iteration, w (d) is the most matching node of the dth iteration, i∈[1,m], m is the number of samples of torque signal data;

[0019] (a4) Update weights: Use the Gaussian function as the neighborhood function and update the weights of the winning neuron and its neighboring neurons according to the neighborhood function;

[0020] (a5) based on steps (a1) to (a4), iterating all input torque samples until a maximum number of iterations is reached, and then stopping the iteration to obtain a fault diagnosis model;

[0021] (a6) Determine the distance D from all input torque samples to the most matching node in the fault diagnosis model, perform Box-Cox transformation on the distance D to obtain the distance D′, and use the 95% confidence level upper limit of the distance D′ as the threshold r. If the distance between the new torque signal data and the most matching node exceeds the threshold r, it is judged that a fault has occurred and an alarm is issued.

[0022] Furthermore, in step 3, the training and verification process of the prediction model of the target device is as follows:

[0023] Divide the preprocessed vibration signal data Dataset_V2 into a training set and a test set;

[0024] The training set and the test set are divided into two groups, one group is used as the response variable of the vibration sample data, and the other group is used as the explanatory variable of the vibration sample data, so as to construct a training data set including the response variable and the explanatory variable, and a test data set including the response variable and the explanatory variable;

[0025] Construct a prediction model based on long-term and short-term neural network of quantile regression, train the prediction model with training data set, and continuously adjust the prediction model weight and bias until the maximum number of iterations is reached, so as to obtain a trained prediction model;

[0026] The explanatory variables of the vibration sample data in the test data set are input into the trained prediction model to obtain the conditional quantiles of the test data set at the Z quantiles. The conditional quantiles at the Z quantiles are used as the input of the Gaussian kernel function to calculate the probability density prediction results of the vibration data of any day e in the test data set.

[0027] Output the value below the 99th percentile as the predicted value for the next week.

[0028] Furthermore, a prediction model of long-term and short-term neural network based on quantile regression is constructed, specifically:

[0029]

[0030] Among them, Q is the prediction model, represents the zth quantile, and z=1,2,…,Z, where Z is the number of quantiles; N represents the number of groups of vibration sample data in the training set Train; Represents the weight parameter matrix set at the zth quantile; represents the set of connection bias vectors at the zth quantile, N is the number of data instances, Y Train is the response variable of the vibration sample data constructed in the training set Train, X Train is the explanatory variable of the vibration sample data constructed in the training set Train, represents quantile, and f(.) represents LSTM function mapping.

[0031] Furthermore, the probability density prediction results The calculation formula is as follows:

[0032]

[0033] Where h is the window width, K(·) is the Gaussian kernel function, G z Represents the conditional quantile of the test set at the zth quantile, where Z is the number of quantiles.

[0034] An equipment fault diagnosis and predictive maintenance system inputs equipment torque signal data and vibration signal data into a fault detection and action prediction model to output fault diagnosis results and action prediction results; equipment maintenance and repair suggestions are obtained based on system operation and maintenance experience, equipment manufacturer maintenance strategy, fault diagnosis results and action prediction results;

[0035] The construction and training of the fault detection and action prediction model includes an acquisition module, a fault diagnosis module and a prediction module;

[0036] The acquisition module is used to acquire and preprocess the equipment torque signal data and vibration signal data to obtain the preprocessed equipment torque signal data and preprocessed vibration signal data;

[0037] The fault diagnosis module is used to extract features and perform SOM clustering on the preprocessed equipment torque signal data, build a fault diagnosis model to diagnose whether the equipment has a fault, output the fault diagnosis results of the faulty equipment, and provide a maintenance strategy based on the fault diagnosis results;

[0038] The prediction module trains a prediction model of a long-term and short-term neural network based on quantile regression based on the preprocessed vibration signal data, outputs the equipment action prediction results based on the trained prediction model, and provides a maintenance strategy based on the prediction results.

[0039] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned equipment fault diagnosis and predictive maintenance method when executing the computer program.

[0040] A computer-readable storage medium stores a plurality of classification programs, which are used to be called by a processor and execute the equipment fault diagnosis and predictive maintenance method as described above.

[0041] The advantages of the equipment fault diagnosis and predictive maintenance method and system provided by the present invention are: the equipment fault diagnosis and predictive maintenance method and system provided in the structure of the present invention, first, pre-process the collected raw data through data cleaning technology to remove noise and irrelevant information to ensure the accuracy of the data. Then, the time-frequency domain features are extracted using the torque signal of the equipment, and the features that can fully reflect the status of the equipment are obtained through a multidimensional feature analysis method. Then, the self-organizing map (SOM) clustering algorithm is used to perform unsupervised learning on the extracted features to determine whether the equipment is faulty. For equipment diagnosed as fault-free, the quantile regression long short-term neural network (QRLSTM) is further used to perform time series analysis on the vibration data to predict the future movement trend of the equipment. This method not only improves the accuracy and reliability of fault diagnosis, but also optimizes resource allocation, reduces false positives and false negatives, ensures that production equipment can continue to operate in the best state, and extends the service life of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the structure flow of the present invention;

[0043] Figure 2 Develop a flow chart for the system;

[0044] Figure 3 It is a schematic diagram of the original data of the torque signal data;

[0045] Figure 4 This is the mapping diagram of the data after SOM clustering on f1 and f10;

[0046] Figure 5 For prediction of vibration signal data Figure 1 ;

[0047] Figure 6 For prediction of vibration signal data Figure 2 ;

[0048] Figure 7 This is the system interface diagram. DETAILED DESCRIPTION

[0049] Below, the technical solution of the present invention is described in detail through specific embodiments. Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific implementation disclosed below.

[0050] like Figures 1 to 7 As shown, the present invention proposes a method for equipment fault diagnosis and predictive maintenance, which inputs the preprocessed equipment torque signal data and vibration signal data into the fault detection and action prediction model to output the fault diagnosis result and the action prediction result; and obtains equipment maintenance and repair suggestions based on the system operation and maintenance experience, the equipment manufacturer's maintenance strategy, the fault diagnosis result and the action prediction result.

[0051] In this embodiment, advanced data processing technology and machine learning algorithms are used to solve the problems of real-time monitoring, fault diagnosis and future action prediction of mechanical equipment in the manufacturing industry, thereby improving the reliability and production efficiency of the equipment.

[0052] In order to improve the accuracy and robustness of monitoring, this embodiment adopts corresponding strategies to identify and process these noise data, identify the impact of abnormal data on the health of the equipment, and use prediction models to predict and warn the actions of the machine in advance.

[0053] This embodiment has the following improvements:

[0054] 1. Multi-source data fusion:

[0055] Improvements: On the basis of the original, more types of sensor data are introduced. For example, this embodiment uses the torque data of the servo motor and the vibration data of the vibration sensor. Temperature, pressure and other data can also be used to perform multi-source data fusion.

[0056] Method: Data fusion algorithm is used to comprehensively process data from different sources: data cleaning and preprocessing (such as noise removal, missing value filling and normalization), data alignment (different data may be collected at different time points, and these data need to be aligned to a common time axis), feature transformation (converting the original single feature into other features of more dimensions), data fusion (comprehensive processing of data from different sources) to extract more comprehensive features.

[0057] Objective: To further improve the accuracy and reliability of fault diagnosis and achieve more comprehensive equipment status monitoring.

[0058] 3. Deep learning model optimization:

[0059] Improvement: In the fault diagnosis and action prediction stages, more complex deep learning models are introduced, such as the Quantile Regression Long Short-Term Neural Network (QRLSTM).

[0060] Method: Combining the advantages of probability density prediction and long-term short-term neural network, probability density prediction not only gives a point estimate (such as the mean), but also provides the entire conditional probability distribution. This enables the model to capture the uncertainty of the data, such as the confidence interval of the predicted value. Long-term short-term neural network is specially designed for processing and predicting time series data. It can effectively capture long-term dependencies in the data, that is, it can "remember" information over a long time span. This combination enables the model to extract features from multiple angles and levels, build a hybrid model, and improve the accuracy of fault diagnosis and action prediction.

[0061] Objective: To enhance the intelligence level of the system and improve the accuracy of diagnosis and prediction.

[0062] Through the combination of the above basic technical solutions and improved technical solutions, the method proposed in this embodiment can achieve efficient and accurate equipment fault diagnosis and action prediction in practical applications, and significantly improve the reliability and production efficiency of the equipment.

[0063] The fault detection and action prediction models are constructed and trained as follows:

[0064] Step 1: collecting and preprocessing equipment torque signal data and vibration signal data to obtain preprocessed equipment torque signal data and preprocessed vibration signal data;

[0065] Data acquisition module: The servo motor of the device reads the torque signal data of the current device and the previous month , is the number of samples of torque signal data. The vibration signal data of the equipment at the current time and the previous month are collected through sensors installed on key components. n is the number of samples of vibration data.

[0066] Data cleaning module: performs data cleaning on the collected torque signal data Dataset_T0 and vibration signal data Dataset_V0, including missing value processing and noise processing:

[0067] (1) Processing of missing values: If the sample Y in Dataset_T0 and Dataset_V0 i and X i If the number of consecutive missing points is less than 5, the missing values ​​are filled with the average of the 5 data before and after the missing point. If the number of consecutive missing points is greater than or equal to 5, the sample data most similar to the sample to be processed is found based on the similarity measurement, and the data at the corresponding position is used to fill in the missing value to obtain the equipment torque signal data after the missing value processing. and vibration signal data

[0068] Noise processing: Torque signal data after missing value processing and vibration signal data There is a lot of noise, so the sliding average filter is used in turn to filter the samples Y in the data set i 1 and Continuous data points should be averaged using a sliding window to reduce the impact of random noise.

[0069] By processing the missing values ​​and noise in Dataset_V0 and Dataset_T0, the data is cleaned and the data quality is improved. The preprocessed torque signal data is obtained. And the preprocessed vibration data

[0070] Step 2: Extract features from the preprocessed equipment torque signal data and perform SOM clustering to build a fault diagnosis model to diagnose whether the equipment is faulty, output the fault diagnosis results of the faulty equipment, and provide a maintenance strategy based on the fault diagnosis results;

[0071] Feature extraction module: (1) Torque signal data Where l is the length of each torque sample. (2) Extract the time-frequency domain features of Dataset_T2: Extract the maximum value (f1), maximum absolute value (f2), minimum value (f3), mean (f4), peak-to-peak value (f5), absolute mean (f6), root mean square value (f7), standard deviation (f8), root mean square amplitude (f9), kurtosis (f10), and other parameters of each torque sample of the torque signal data Dataset_T2. 10 ), skewness(f 11 ), margin index (f 12 )、Waveform Index(f 13 )、Pulse index(f 14 ), peak index (f 15 ), and obtain the torque feature data set after feature extraction Y i 3 ={f i1 ,f i2 ,...,f i15}, where Y i is the i-th sample, f i is the feature of the i-th sample.

[0072] The feature extraction calculation process is as follows:

[0073] f1=max(y j ),j=1,2,...,l;

[0074]

[0075] f5 = f1 - f3;

[0076]

[0077] Fault diagnosis module: Torque feature dataset after feature extraction Y i 3 ={f i1 ,f i2 ,...,f i15} Perform SOM clustering to diagnose whether the torque sample is abnormal. The specific diagnosis process is as follows (a1) to (a6):

[0078] (a1) extracting the time-frequency domain features of the preprocessed equipment torque signal data to obtain the torque feature dataset Dataset_T3 after feature extraction;

[0079] (a2) Determine the size of the SOM: Where M is the number of neurons and N is the number of data instances;

[0080] Randomly initialize the SOM neurons: the weights (ω cv ) is initialized to a random number between 0 and 1, where c = 1, 2, ..., k(h) and v = 1, 2, ..., k(w), where k(h) and k(w) represent the number of rows and columns of the SOM graph, and the initial neuron field radius and learning rate are established;

[0081] (a3) Y i 3 ={f i1 ,f i2 ,...,f i15 After normalization, the Euclidean distance is used as the discriminant function to find the input torque sample Y in the SOM neuron i The best matching node is used as the output node, and the weight of the output node satisfies: data = argmin||Y d -w (d) ||, where data is the weight, d is the number of iterations, and Y d is the torque sample of the dth iteration, w (d) is the most matching node of the dth iteration, i∈[1,m], m is the number of samples of torque signal data;

[0082] (a4) Update weights: Use Gaussian function as neighborhood function, and update the weights of the winning neuron and its neighboring neurons according to the neighborhood function; the purpose of weight update is to make the weights of the winning neuron and its neighboring neurons closer to the input torque sample;

[0083] (a5) based on steps (a1) to (a4), iterating all input torque samples until a maximum number of iterations is reached, and then stopping the iteration to obtain a fault diagnosis model;

[0084] (a6) Determine the distance D = {d1, d2, ..., d m}, perform Box-Cox transformation on distance D to obtain distance D′, and take the 95% confidence level upper limit of distance D′ as threshold r. If the distance between the new torque signal data and the most matching node (BMU) exceeds threshold r, it is judged that a fault has occurred and an alarm is issued.

[0085] Step 3: Based on the preprocessed vibration signal data, a prediction model of a long-term and short-term neural network based on quantile regression is trained. When the equipment does not fail, the equipment action prediction result is output based on the trained prediction model, and a maintenance strategy is given based on the prediction result;

[0086] Model prediction module: collects vibration data after data preprocessing Including vibration data from the current moment to the previous month, using the prediction model of long-term and short-term neural network based on quantile regression Make a forecast, predict the data for the next week, and output the forecast results.

[0087] The training and validation process of the prediction model for the target device is as follows:

[0088] (b1) Divide the vibration signal data Dataset_V2 into a training set Train and a test set Test;

[0089] (b2) dividing the training set and the test set into two groups, one group as the response variable of the vibration sample data, and the other group as the explanatory variable of the vibration sample data, thereby constructing a training data set including the response variable and the explanatory variable, and a test data set including the response variable and the explanatory variable;

[0090] The sample data of the vibration sample data on the tth day in the training set Train is Dataset′={Q′(t)|t=1,2,...,30}, W′(t) represents the vibration data of the sample data on the P(t)th day in the training set Train; the vibration data from the 24th to the 30th day is Dataset″ i= {Q′ i (t)|t=24,...,30} as the response variable of the training set vibration sample data; the remaining 23 days of sample data in the training set Train Dataset″ i = {Q′ i (t)|t=1,2,...,23} are used as the explanatory variables of the i-th group of vibration sample data, thereby constructing a training data set containing 23 explanatory variables and 7 response variables, denoted as {X Train ,Y Train};X Train represents the explanatory variable of the vibration sample data in the training set Train, Y Train is the response variable of the vibration sample data in the training set Train; and X(i)={x β |β=1,2,...,23},x γ It is the βth explanatory variable of the vibration sample data in the training set Train.

[0091] According to the construction process of the training data set, the test set Test also constructs a data set including 23 explanatory variables and 7 response variables, denoted as {X test ,Y test}, where X test is the explanatory variable of the vibration sample data in the test set Test, Y test It is the response variable of the vibration sample data in the test set Test.

[0092] (b3) constructing a prediction model of a long-term and short-term neural network based on quantile regression, training the prediction model using a training data set, and continuously adjusting the prediction model weights and biases until a maximum number of iterations is reached, thereby obtaining a trained prediction model;

[0093] Use formula (1) to build a prediction model:

[0094]

[0095] Among them, Q is the prediction model, represents the zth quantile, and z=1,2,…,Z, where Z is the number of quantiles; N represents the number of groups of vibration sample data in the training set Train; Represents the weight parameter matrix set at the zth quantile; represents the set of connection bias vectors at the zth quantile, Y Train is the response variable of the vibration sample data constructed in the training set Train, X Train is the explanatory variable of the vibration sample data constructed in the training set Train, represents quantile, and f(.) represents LSTM function mapping.

[0096] (b4) The test data set {X test ,Y test The explanatory variable X of the vibration sample data in test Input it into the trained prediction model to obtain the conditional quantiles {G1, G2, ..., G z ,...,G Z}; Among them, G z Represents the conditional quantile of the test set at the zth quantile; the conditional quantile at the Zth quantile is used as the input of the Gaussian kernel function to calculate the probability density prediction result of the vibration data of any day e in the test data set

[0097]

[0098] Where h is the window width, K(·) is the Gaussian kernel function, G Z Represents the conditional quantile of the test set at the zth quantile, where Z is the number of quantiles.

[0099] (b5) Output the value below the 99th percentile as the predicted value for the next week;

[0100] Result output module: The result output module is the terminal link of the entire prediction and fault diagnosis system. Its main function is to present the fault diagnosis results and action prediction results generated by the system to the user in an intuitive and easy-to-understand way. Specifically, the result output module will organize and summarize the results obtained by the fault detection and action prediction models, and based on the system operation and maintenance experience and the equipment manufacturer's maintenance strategy, it will then be presented to the user in the form of charts, reports or dashboards. At the same time, based on the fault diagnosis results and action prediction results, the result output module will also provide corresponding maintenance suggestions or early warning information to help users better understand the operating status of the equipment, take necessary maintenance measures in a timely manner, and avoid potential failure risks. In addition, in order to improve the user experience and ease of use of the system, the result output module may also support customized settings, allowing users to choose different output content and formats according to their needs.

[0101] This embodiment has the following advantages (c1) to (c6):

[0102] (c1) Application of unsupervised learning in fault diagnosis in this embodiment: The self-organizing map (SOM) clustering algorithm is used to perform unsupervised learning on the vibration signal of the equipment and the torque signal of the servo motor to determine whether the equipment has a fault. This method does not require the accumulation of a large amount of fault data, thereby reducing the difficulty and cost of data collection, making engineering applications more convenient and practical.

[0103] (c2) Multi-dimensional feature extraction and analysis: This embodiment combines the vibration signal data of the equipment and the torque signal data of the servo motor, and extracts features that can fully reflect the equipment status through time-frequency domain analysis. This multi-dimensional feature extraction method reduces the deviation of a single data source and provides more comprehensive equipment health status information, thereby improving the accuracy and reliability of fault diagnosis.

[0104] (c3) Application of Quantile Regression Long Short-Term Neural Network (QRLSTM) in Motion Prediction: For equipment diagnosed as fault-free, this embodiment further uses the QRLSTM network (i.e., prediction model) to perform time series analysis on vibration data to predict the future motion trend of the equipment. This method not only optimizes resource allocation and reduces false positives and false negatives, but also ensures that production equipment can continue to operate in the best condition and prolongs the service life of the equipment.

[0105] (c4) Real-time monitoring and active maintenance: This embodiment can accurately determine the current health status of the equipment and predict its future movement trend by real-time monitoring and analysis of the equipment's vibration signals and servo motor's torque signals. This real-time monitoring and active maintenance method helps to prevent potential failures in advance, reduce downtime caused by sudden failures, significantly reduce maintenance costs, and improve the stability and reliability of the production line.

[0106] (c5) Phased processing: This embodiment adopts a phased processing method, firstly diagnosing the equipment fault, and only when the equipment is diagnosed as fault-free, will the equipment action prediction be further performed. This processing method avoids unnecessary duplication of work, improves the operating efficiency of the system and the pertinence of maintenance work.

[0107] (c6) Optimizing resource allocation and reducing false alarms and missed alarms: The method of this embodiment not only improves the accuracy and reliability of fault diagnosis, but also optimizes resource allocation and reduces the possibility of false alarms and missed alarms. This helps ensure that production equipment can continue to operate in the best condition, prolong the service life of the equipment, and also saves maintenance costs for the enterprise.

[0108] Therefore, compared with CN201911077243.4, from the data point of view, CN201911077243.4 needs to collect the health and fault data of the data. The equipment fault diagnosis of this embodiment adopts an unsupervised clustering algorithm, which does not need to accumulate fault data, and is more conducive to engineering applications. The single feature data adopted by CN201911077243.4 and the multi-dimensional data in the time-frequency domain of this embodiment can reduce the deviation of the single data source and provide more comprehensive equipment health status information. From the method point of view, this embodiment involves a fault diagnosis and predictive maintenance method and system, which first diagnoses whether the equipment is faulty, and predicts the action of the equipment diagnosed as having no fault, which is conducive to preventing potential faults in advance and optimizing the operating efficiency of the equipment. By real-time monitoring and analysis of the vibration signal of the equipment and the torque signal of the servo motor, the system can accurately judge the current health status of the equipment. For those equipment diagnosed as fault-free, the system further uses historical data and machine learning models to predict their future action trends, thereby making maintenance work more proactive and efficient. This approach can not only reduce downtime caused by sudden failures, but also significantly reduce maintenance costs and improve the stability and reliability of the entire production line.

[0109] Compared with CN202010784496.1, on the one hand, compared with CN202010784496.1, this embodiment extracts multi-dimensional features from the data, which is not limited to the analysis of a single signal, but combines the vibration signal data of the equipment and the torque signal data of the servo motor, and extracts features that can fully reflect the state of the equipment through advanced time-frequency domain analysis methods. This method makes the data characterize the state of the equipment more detailed and accurate, which helps to improve the accuracy and reliability of fault diagnosis. On the other hand, after cleaning the data, this embodiment first performs equipment fault diagnosis, and only when the equipment is diagnosed as fault-free will the equipment action prediction be further performed. This staged processing method avoids unnecessary duplication of work, and there is no need to redefine the maintenance points before each prediction, thereby greatly improving the operating efficiency of the system and the pertinence of maintenance work. This is conducive to optimizing resource allocation, reducing false alarms and missed alarms, ensuring that production equipment can continue to operate in the best state, and also extending the service life of the equipment.

[0110] As an embodiment

[0111] (d1) Data acquisition module: The servo motor of the equipment reads the torque signal data. The vibration signal data is collected through sensors installed on key components.

[0112] (d2) Data cleaning module: performs data cleaning on the collected torque signal data and vibration signal data, including missing value processing and noise processing. Figure 3This is the original data graph of the torque signal data. It can be seen that there are obvious missing values ​​(discontinuous data) and noisy data (random and abnormally fluctuating data).

[0113] (d3) Feature extraction module: extract features from the equipment torque signal data after data cleaning.

[0114] (d4) Fault diagnosis module: Use the feature-extracted data set to perform SOM clustering to diagnose whether the sample is abnormal. Figure 4 This is a mapping diagram of the clustered torque data on f1 and f10, where the horizontal axis is f10 and the vertical axis is f1.

[0115] (d5) Model prediction module: predicts the machine's movement based on vibration signal data. Figure 5 and Figure 6 This is a graph showing motion prediction based on vibration data.

[0116] (d6) Result output module: The result output module is the terminal link of the entire prediction and fault diagnosis system. Its main function is to present the prediction results and fault diagnosis results generated by the system to the user in an intuitive and easy-to-understand manner. Figure 7 It is the interface diagram of the entire system.

[0117] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for equipment fault diagnosis and predictive maintenance, characterized in that: Inputting equipment torque signal data and vibration signal data into a fault detection and action prediction model to output a fault diagnosis result and an action prediction result; Get equipment maintenance recommendations based on system operation and maintenance experience, equipment manufacturer maintenance strategies, fault diagnosis results, and action prediction results; The fault detection and action prediction model is constructed and trained as follows: Step 1: collecting and preprocessing the torque signal data and the vibration signal data of the equipment to obtain the preprocessed torque signal data and the preprocessed vibration signal data of the equipment; Step 2: Extract features from the preprocessed equipment torque signal data and perform SOM clustering to build a fault diagnosis model to diagnose whether the equipment is faulty, output the fault diagnosis results of the faulty equipment, and provide a maintenance strategy based on the fault diagnosis results; Step 3: Based on the preprocessed vibration signal data, a prediction model of the long-term and short-term neural network based on quantile regression is trained. When the equipment does not fail, the equipment action prediction result is output based on the trained prediction model, and a maintenance strategy is given based on the prediction result.

2. The equipment fault diagnosis and predictive maintenance method according to claim 1, characterized in that: In step 1, the preprocessing process of the equipment torque signal data Dataset_T0 and the vibration signal data Dataset_V0 is as follows: Missing value processing: If the number of consecutive missing points in the samples in Dataset_T0 and Dataset_V0 is less than 5, the missing values ​​are filled with the average of the 5 data before and after the missing points; if the number of consecutive missing points in the samples is greater than or equal to 5, the sample data most similar to the sample to be processed is found according to the similarity measurement, and the data at the corresponding position is used to fill in the missing value processing equipment torque signal data Dataset_T1 and vibration signal data Dataset_V1; Noise processing: Using a sliding average filter, apply a sliding window to the continuous data points of the samples in Dataset_T1 and Dataset_V1 to calculate the average value in turn to reduce the impact of random noise, and obtain the preprocessed equipment torque signal data Dataset_T2 and the preprocessed vibration signal data Dataset_V2.

3. The equipment fault diagnosis and predictive maintenance method according to claim 1, characterized in that: In step 2, specifically: (a1) extracting the time-frequency domain features of the preprocessed equipment torque signal data to obtain the torque feature dataset Dataset_T3 after feature extraction; (a2) Determine the size of the SOM, initialize the weights between SOM neurons to random numbers between 0 and 1, and establish the initial neuron field radius and learning rate; (a3) After normalizing the feature dataset Dataset_T3, the Euclidean distance is used as the discriminant function to find the input torque sample Y in the SOM neuron i The best matching node is used as the output node, and the weight of the output node satisfies: data = argmin||Y d -w (d) ||, where data is the weight, d is the number of iterations, and Y d is the torque sample of the dth iteration, w (d) is the most matching node of the dth iteration, i∈[1,m], m is the number of samples of torque signal data; (a4) Update weights: Use the Gaussian function as the neighborhood function and update the weights of the winning neuron and its neighboring neurons according to the neighborhood function; (a5) based on steps (a1) to (a4), iterating all input torque samples until a maximum number of iterations is reached, and then stopping the iteration to obtain a fault diagnosis model; (a6) Determine the distance D from all input torque data to the most matching node in the fault diagnosis model, and perform Box-Cox transformation on the distance D to obtain the distance D ′ , with a distance D ′ The 95% confidence level upper limit is taken as the threshold r. If the distance between the new torque sample and the best matching node exceeds the threshold r, it is judged that a fault has occurred and an alarm is issued.

4. The equipment fault diagnosis and predictive maintenance method according to claim 2, characterized in that: In step 3, the training and validation process of the prediction model for the target device is as follows: Divide the preprocessed vibration signal data Dataset_V2 into a training set and a test set; The training set and the test set are divided into two groups, one group is used as the response variable of the vibration sample data, and the other group is used as the explanatory variable of the vibration sample data, so as to construct a training data set including the response variable and the explanatory variable, and a test data set including the response variable and the explanatory variable; Construct a prediction model based on long-term and short-term neural network of quantile regression, train the prediction model with training data set, and continuously adjust the prediction model weight and bias until the maximum number of iterations is reached, so as to obtain a trained prediction model; The explanatory variables of the vibration sample data in the test data set are input into the trained prediction model to obtain the conditional quantiles of the test data set at the Z quantiles. The conditional quantiles at the Z quantiles are used as the input of the Gaussian kernel function to calculate the probability density prediction results of the vibration data of any day e in the test data set. Output the value below the 99th percentile as the predicted value for the next week.

5. The equipment fault diagnosis and predictive maintenance method according to claim 4, characterized in that: Construct a prediction model of long-term and short-term neural network based on quantile regression, specifically: Among them, Q is the prediction model, represents the zth quantile, and z=1,2,…,Z, where Z is the number of quantiles; N represents the number of groups of vibration sample data in the training set Train; Represents the weight parameter matrix set at the zth quantile; represents the set of connection bias vectors at the zth quantile, Y train is the response variable of the vibration sample data constructed in the training set Train, X Train is the explanatory variable of the vibration sample data constructed in the training set Train, represents quantile, and f(.) represents LSTM function mapping.

6. The equipment fault diagnosis and predictive maintenance method according to claim 4, characterized in that: Probability density prediction results The calculation formula is as follows: Where h is the window width, K(·) is the Gaussian kernel function, G z Represents the conditional quantile of the test set at the zth quantile, where Z is the number of quantiles.

7. An equipment fault diagnosis and predictive maintenance system, characterized in that: Inputting equipment torque signal data and vibration signal data into a fault detection and action prediction model to output a fault diagnosis result and an action prediction result; Get equipment maintenance recommendations based on system operation and maintenance experience, equipment manufacturer maintenance strategies, fault diagnosis results, and action prediction results; The construction and training of the fault detection and action prediction model includes an acquisition module, a fault diagnosis module and a prediction module; The acquisition module is used to acquire and preprocess the equipment torque signal data and vibration signal data to obtain the preprocessed equipment torque signal data and preprocessed vibration signal data; The fault diagnosis module is used to extract features and perform SOM clustering on the preprocessed equipment torque signal data, build a fault diagnosis model to diagnose whether the equipment has a fault, output the fault diagnosis results of the faulty equipment, and provide a maintenance strategy based on the fault diagnosis results; The prediction module trains a prediction model of a long-term and short-term neural network based on quantile regression based on the preprocessed vibration signal data, outputs the equipment action prediction results based on the trained prediction model, and provides a maintenance strategy based on the prediction results.

8. 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 equipment fault diagnosis and predictive maintenance method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of classification programs, and the plurality of classification programs are used to be called by a processor and execute the equipment fault diagnosis and predictive maintenance method as described in any one of claims 1 to 6.

Images