Method and device for predicting remaining life of thin film evaporator for Lyocell fiber production

Abstract

The present invention discloses a method and device for predicting the remaining life of a thin-film evaporator used in the production of Lyocell fibers, which relates to the technical field of mechanical fault diagnosis, and includes: collecting vibration data by vibration sensors in a preset area on the surface of the thin-film evaporator cylinder and performing preprocessing, and calculating the characteristic frequencies of key components; vectorially adding the vibration data of individual vibration sensors to construct a one-dimensional convolutional neural network model to obtain a characteristic curve; fitting with a plurality of high-order exponential functions to establish a remaining life characteristic curve; obtaining a data set to establish a one-dimensional convolutional neural network model for predicting the remaining life, inputting the data set into the one-dimensional convolutional neural network model to obtain a remaining life prediction model; and then using the vibration data collected in real time and the characteristic frequencies of key components after processing as the input of the prediction model. The present invention can reduce the complexity of the feature extraction network model, improve the real-time operation of the model, provide more accurate data labels, and effectively improve the accuracy of remaining life prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and particularly to a method and device for predicting the remaining life of a thin-film evaporator used in the production of Lyocell fibers. Background Art

[0002] As an efficient evaporation device, the thin-film evaporator is widely used in the spinning field, mainly for the concentration of spinning solutions, the recovery of solvents, and the treatment of heat-sensitive materials. Its working principle is to form a uniform liquid film of the feed liquid on the heating surface through a rotating scraper, and perform efficient evaporation under vacuum conditions.

[0003] However, during the long-term operation of the thin-film evaporator, various faults will inevitably occur. Among them, the main shaft of the large thin-film evaporator has a large axial dimension and is sensitive to vibration. Complex structures are connected to the main shaft, the system structure mode is complex, and the vibration of the main shaft is likely to excite the system structure mode, affecting the normal operation of the equipment and even damaging the equipment. At the same time, the locking / sticking of the main shaft is a common fault during the operation of the large thin-film evaporator, and the excessive vibration displacement of the main shaft is an important cause of this fault. The reasons for the abnormal vibration of the main shaft include imperfect equipment design, easy resonance during operation, as well as damage to the main shaft bearings, unbalance of the main shaft wear, and bending faults of the main shaft. These faults will lead to a decrease in the evaporation efficiency of the thin-film evaporator, unstable product quality, and even cause equipment shutdown, seriously affecting the continuity and economy of spinning production.

[0004] Traditional fault diagnosis methods mainly rely on manual inspection and experience judgment. This method has poor real-time performance and cannot detect potential fault hazards in a timely manner; and it has low accuracy, depends on personal experience, and is prone to misjudgment and missed judgment; it has insufficient predictability, cannot predict the development trend of faults, and is difficult to perform preventive maintenance.

[0005] A Chinese patent published on May 15, 2024, discloses a fault diagnosis method for mechanical equipment, with the publication number CN118194136B. It collects various types of information such as vibration time-domain information, acoustic time-domain information, and image information of the mechanical equipment to be diagnosed under the current operating state; and outputs vibration frequency-domain waveforms, acoustic frequency-domain waveforms, and image information; performs noise reduction processing; determines the network model and its transfer function; trains the network model based on historical data and completes the relearning of the network model; inputs the vibration frequency-domain waveforms, acoustic frequency-domain waveforms, and image information after noise reduction processing into the optimized network model; integrates the fault types of vibration time-domain information, acoustic time-domain information, and image information, calculates the correlation between each fault type and the fault classification results of historical data, and obtains the fault type of the mechanical equipment according to the correlation. It uses the measures of looking, listening, asking, and feeling the pulse to fuse and judge the fault type of the mechanical equipment, greatly improving the accuracy of fault diagnosis.

[0006] This method of diagnosing based on fault types cannot predict equipment faults in advance in real time. Moreover, this solution applies the interactive fusion of different data types and time data information to fault diagnosis. In the scenario of predicting the remaining life of equipment as a pre-judgment of equipment faults, especially in the scenario of predicting the remaining life with high real-time requirements, it is difficult to respond quickly. At the same time, when this solution extracts features, the connection between the features and the life labels is not strong, making it difficult to accurately construct the life curve. In addition, the key structural characteristics of the equipment are not fully extracted and integrated, which affects the prediction accuracy of important equipment.

[0007] Therefore, it is urgent to monitor the life cycle of the thin-film evaporator during use in real time. By collecting and analyzing the operation data of the thin-film evaporator in real time, a remaining life diagnosis model is established to achieve early warning of faults and predictive maintenance, ensure the effective operation of the thin-film evaporator, and improve the efficiency and product quality of spinning production. Summary of the Invention

[0008] Aiming at the deficiencies of the above-mentioned prior art, the present invention provides a method and device for predicting the remaining life of a thin-film evaporator for Lyocell fiber production, which can monitor the thin-film evaporator in real time and predict the remaining life, perform preventive maintenance accurately and in a timely manner, and ensure the effective operation of the thin-film evaporator.

[0009] To solve the above technical problems, the present invention adopts the following technical solutions:

[0010] The present invention provides a method for predicting the remaining life of a thin-film evaporator for Lyocell fiber production, and the prediction method includes the following steps:

[0011] S1. Obtain the vibration data and spindle speed data of the whole process from the normal operation of the thin-film evaporator to the failure of the main shaft and / or main shaft bearings, which causes the thin-film evaporator to be unable to operate normally;

[0012] The vibration data includes acceleration signals in three directions respectively collected by vibration sensors at the upper and lower ends and the middle region of the cylinder of the thin-film evaporator; among them, the vibration sensors at the upper and lower ends respectively monitor the fault conditions of the upper and lower bearings of the main shaft, and the vibration sensor in the middle region of the cylinder monitors the fault condition of the main shaft;

[0013] S2. Perform time-domain analysis and frequency-domain analysis on multiple data segments obtained by segmenting the acceleration signals at set time intervals; among them, time-domain analysis is used to extract the root mean square and kurtosis of each data segment, and frequency-domain analysis is used to perform the conversion from time domain to frequency domain on each data segment and extract the frequency centroid; obtain the fault characteristic frequency of the main shaft according to the spindle speed data, and obtain the fault characteristic frequency of the main shaft bearing according to the spindle speed data and the size parameters of the main shaft bearing;

[0014] S3. For the multiple vibration vector data obtained by vectorially adding the acceleration signals in three directions of each vibration sensor, build a one-dimensional convolutional neural network model that includes convolutional layers, pooling layers, activation functions, and fully connected layers; extract the response trend of the vibration response of each vibration vector data over time through the one-dimensional convolutional neural network model of each vibration vector data, and obtain the vibration vector data characteristic curves of the upper and lower bearings of the main shaft and the main shaft.

[0015] S4. Use multiple high-order exponential functions to fit the vibration vector data characteristic curves of the upper and lower bearings of the main shaft and the main shaft obtained in step S3 to form the remaining life characteristic curves of the upper bearing of the main shaft, the remaining life characteristic curve of the lower bearing of the main shaft, and the remaining life characteristic curve of the main shaft, and then express them in percentage form.

[0016] S5. Using the set time interval in step S2 as the time resolution, the root mean square, kurtosis, frequency centroid, fault characteristic frequencies of the main shaft bearings, and the fault characteristic frequencies of the main shaft obtained in step S2, as well as the data on the remaining life characteristic curves of the upper bearing of the main shaft, the remaining life characteristic curve of the lower bearing of the main shaft, and the remaining life characteristic curve of the main shaft obtained in step S4, jointly constitute a data set, where the data on the remaining life characteristic curves of the upper bearing of the main shaft, the remaining life characteristic curve of the lower bearing of the main shaft, and the remaining life characteristic curve of the main shaft are label data.

[0017] S6. Divide the data set obtained in step S5 into a training set, a validation set, and a test set; after training, validation, and testing, obtain a remaining life prediction model with an accuracy meeting the requirements.

[0018] S7. Obtain the real-time vibration data and the main shaft speed of the target thin-film evaporator; the root mean square, kurtosis, frequency centroid, fault characteristic frequencies of the main shaft, and the fault characteristic frequencies of the main shaft bearings obtained after processing are used as the inputs of the remaining life prediction model to predict the remaining life of the main shaft of the target thin-film evaporator and / or the remaining life of the main shaft bearings.

[0019] Further, in S2, the root mean square and kurtosis of each data segment are extracted by performing time-domain analysis on each data segment.

[0020] Specifically, the root mean square of each data segment is extracted through time-domain analysis, and its calculation formula is:

[0021] ;

[0022] In the formula: is the root mean square of the data segment; is the number of samples of the data segment; is the 𝑖-th sampling value within the data segment;

[0023] Extract the kurtosis of each data segment through time-domain analysis. The calculation formula is as follows:

[0024] ;

[0025] In the formula: is the kurtosis of the data segment; is the average value of this data segment; is the 𝑖-th sampling value within the data segment.

[0026] Furthermore, in S2, perform the time-domain to frequency-domain conversion on each data segment using frequency-domain analysis, and then extract the frequency centroid;

[0027] Specifically, the time-domain to frequency-domain conversion of each data segment is calculated as follows:

[0028] ;

[0029] In the formula: is the transformed frequency-domain value of the data segment; is the number of discrete sampling points of the data segment; is the serial number of the time-domain discrete value of the data segment; is the n-th data sample of the data segment; is the imaginary unit;

[0030] The calculation formula for frequency centroid extraction is:

[0031] ;

[0032] In the formula: is the frequency centroid of the data segment; is the frequency value of the k-th spectral line; is the total number of spectral lines.

[0033] Furthermore, the set time interval is 0.5 s.

[0034] Furthermore, in S2, the fault characteristic frequencies of the bearing include:

[0035] Inner ring fault characteristic frequency:

[0036] ;

[0037] Outer ring fault characteristic frequency:

[0038] ;

[0039] Rolling element fault characteristic frequency:

[0040] ;

[0041] Cage fault characteristic frequency:

[0042] ;

[0043] Where: is the number of rolling elements; is the diameter of the rolling element; is the pitch diameter of the bearing; is the contact angle; is the rotational frequency of the main shaft, obtained based on the main shaft speed.

[0044] Furthermore, the fault characteristic frequency of the main shaft is:

[0045] ;

[0046] Where: is the multiple frequency number; is the rotational frequency of the main shaft, obtained based on the main shaft speed; is the rotational disturbance frequency of the main shaft.

[0047] Furthermore, in S2, the root mean square, kurtosis, and frequency centroid of each acceleration signal, as well as the fault characteristic frequency of the bearing and the fault characteristic frequency of the main shaft, are normalized and then output.

[0048] Further, in S3, the one-dimensional convolutional neural network model includes:

[0049] Build a one-dimensional convolutional neural network model containing convolutional, pooling, activation function, and fully connected layers for each vibration vector data:

[0050] Input sequence: ;

[0051] Where: is the input value at the t-th position in the sequence;

[0052] Convolution operation: ;

[0053] Where: is the output value at the t-th position after the convolution operation; is the size of the convolution kernel; is the weight parameter at the i-th position in the convolution kernel; is the element corresponding to the input sequence participating in the convolution calculation; is the bias term of the convolution operation;

[0054] Adopt the RELU activation function: ;

[0055] Where: is the output value after activation at the t-th position;

[0056] The max pooling operation is adopted: ;

[0057] In the formula: is the output value of the t-th pooling window; is the size of the pooling window; is the value at the i-th position in the output of the activation function;

[0058] Fully connected layer: ;

[0059] In the formula: is the value of the j-th output node; is the weight parameter connecting the i-th input node to the j-th output node; is the output value of the i-th node in the hidden layer; is the bias term of the j-th output node;

[0060] The cross-entropy loss function is adopted: ;

[0061] In the formula: is the cross-entropy loss value; is the total number of categories; the value of category in the true label; is the probability value of the -th category predicted by the model.

[0062] Furthermore, a remaining useful life prediction model is constructed based on the one-dimensional convolutional neural network model, and the root mean square, kurtosis, frequency centroid, fault characteristic frequency of the main shaft, and fault characteristic frequency of the main shaft bearing obtained by processing the real-time vibration data and the main shaft speed acquired in step S7 in the same way as in step S2 are input into the remaining useful life prediction model in real time.

[0063] Furthermore, in S3, the vibration vector data characteristic curve of the main shaft is obtained by taking the average of the vibration vector data characteristic curves corresponding to the vibration sensors in the middle region of the cylinder body.

[0064] Furthermore, in S4, the expressions of the remaining useful life characteristic curves of the upper bearing of the main shaft, the remaining useful life characteristic curve of the lower bearing of the main shaft, and the remaining useful life characteristic curve of the main shaft in the form of percentages are as follows:

[0065] ;

[0066] In the formula: , , , , is an adjustable parameter; is the system time of vibration data; is the defined time limit for the equipment to change from completely normal to abnormal; is the percentage of remaining life; is the natural base.

[0067] The present invention also provides a device for predicting the remaining life of a thin film evaporator for Lyocell fiber production, and the device includes:

[0068] A data acquisition module, configured to obtain real-time vibration data collected by vibration sensors at the upper and lower ends and the middle region of the cylinder body of the target thin film evaporator, and at the same time obtain the spindle speed from the central control platform of the target thin film evaporator;

[0069] A data preprocessing module, configured to segment the real-time vibration data obtained by the data acquisition module at a set time interval to obtain multiple vibration data segments at the set time interval; perform time-domain analysis and frequency-domain analysis on each data segment respectively to obtain the root mean square, kurtosis, and frequency centroid of each vibration data segment; obtain the fault characteristic frequency of the spindle according to the obtained spindle speed data, and then obtain the fault characteristic frequency of the spindle bearing according to the spindle speed data and the size parameters of the spindle bearing; finally, the data output by the data preprocessing module is the root mean square, kurtosis, frequency centroid of the acceleration signals in three directions of each vibration sensor, the fault characteristic frequency of the spindle bearing, and the fault characteristic frequency of the spindle; the root mean square, kurtosis, and frequency centroid of the acceleration signals in three directions of each vibration sensor, as well as the fault characteristic frequencies of the bearing and the spindle, are output after being normalized;

[0070] A remaining life prediction module, in which a remaining life prediction model in the above-mentioned method for predicting the remaining life of a thin film evaporator for Lyocell fiber production is encapsulated, and the data output by the data preprocessing module is used as the input data of the remaining life prediction model to predict the remaining life of the target thin film evaporator. The specifically predicted remaining life includes the remaining life of the spindle and / or the remaining life of the spindle bearing.

[0071] In summary, the present invention has the following beneficial effects:

[0072] (1) This solution forms a data matrix by using multi-component characteristic frequency data and big data feature extraction data, combines the vibration data collected from the bearings at the upper and lower ends and the spindle in the middle region of the thin film evaporator cylinder body with the characteristic frequencies of key components such as the spindle and the bearing. In practical applications, it can capture equipment degradation information more accurately, making the correlation between the characteristic data and the life label higher, and more effectively constructing a relationship model between the fault characteristics and the life label, and then forming a more accurate remaining life prediction model.

[0073] (2) This application collects vibration data during the single life cycle from normal to abnormal at the upper and lower ends and the middle region of the thin-film evaporator cylinder, and designs a remaining life characteristic curve by coupling a multi-term high-order exponential function empirical mathematical model with big data feature extraction, reducing the complexity of the feature extraction network model, improving the real-time operation of the model, meeting the real-time life prediction requirements, and providing more accurate data labels while effectively improving the accuracy of remaining life prediction.

[0074] (3) The average value of 4 characteristic curves of the middle region of the thin-film evaporator with respect to the main axis is taken to obtain the remaining life characteristic curve of the main axis, and characteristic curves reflecting the health status of the upper bearing, lower bearing, and main axis of the thin-film evaporator are constructed independently, simplifying the calculation process.

[0075] (4) Integrate the multi-component characteristic frequency data, big data feature extraction data, and remaining life characteristic curve data in the historical data to form a three-dimensional training dataset, train the remaining life prediction model designed based on the deep learning model, train the remaining life prediction model based on the dataset, establish the association mapping between the feature data and the life label, improve the prediction accuracy, and thus obtain a remaining life prediction model with the required accuracy.

[0076] (5) As the main equipment in Lyocell production, the large thin-film evaporator has a huge size and is sensitive to vibration during operation. If a failure occurs, it will have a huge impact on production, and it is very inconvenient to repair the large thin-film evaporator, especially for immediate repair after a failure. Therefore, it is very important to conduct real-time monitoring and remaining life prediction for the large thin-film evaporator. Therefore, a remaining life prediction model is established based on the thin-film evaporator, and the combination of multi-component characteristic frequency and big data feature extraction is used to monitor the vulnerable components of the thin-film evaporator in real time and output life prediction, which is beneficial to improving the fault tolerance rate and providing convenience for effective production management. Description of the Drawings

[0077] Figure 1 It is a layout diagram of vibration sensor measuring points on the thin-film evaporator cylinder.

[0078] Figure 2 It is a schematic flow diagram of the data preprocessing module.

[0079] Figure 3 It is a schematic flow diagram of the remaining life prediction module.

[0080] Figure 4 It is a schematic flow diagram of the design of the remaining life prediction and the remaining life label data.

[0081] Figure 5 It is a schematic flow diagram of the design of the remaining life prediction model.

[0082] Figure 6 It is a vibration vector data diagram.

[0083] Figure 7 It is a characteristic curve of vibration vector data. Specific implementation mode

[0084] The present invention will be further described in detail below in conjunction with embodiments.

[0085] Embodiment 1:

[0086] A method for predicting the remaining life of a thin-film evaporator used in the production of Lyocell fibers specifically includes the following steps:

[0087] S1. Obtain the vibration data of the whole process of the thin-film evaporator from normal operation to the failure of the main shaft and / or the main shaft bearing, resulting in the abnormal operation of the thin-film evaporator, and obtain the main shaft speed data from the central control platform;

[0088] The vibration data includes the acceleration signals in three directions respectively collected by the vibration sensors at the upper and lower ends and the middle region of the cylinder body of the thin-film evaporator; among them, the vibration sensors at the upper and lower ends respectively monitor the failure conditions of the upper and lower bearings of the main shaft, and the vibration sensor in the middle region of the cylinder body monitors the failure condition of the main shaft;

[0089] S2. Perform time-domain analysis and frequency-domain analysis on the multiple data segments obtained by segmenting the acceleration signals at a set time interval; among them, time-domain analysis is used to extract the root mean square and kurtosis of each data segment, and frequency-domain analysis is used to perform the conversion from time domain to frequency domain on each data segment and extract the frequency centroid; obtain the fault characteristic frequency of the main shaft according to the main shaft speed data, and obtain the fault characteristic frequency of the main shaft bearing according to the main shaft speed data and the size parameters of the main shaft bearing;

[0090] During implementation, set the time interval to 0.5 s.

[0091] In S2, the root mean square and kurtosis of each data segment are extracted through time-domain analysis of each data segment; specifically,

[0092] The root mean square of each data segment is extracted through time-domain analysis, and its calculation formula is:

[0093] ;

[0094] In the formula: is the root mean square of the data segment; is the number of samples of the data segment; is the 𝑖-th sampling value within the data segment.

[0095] The kurtosis of each data segment is extracted through time-domain analysis, and its calculation formula is:

[0096] ;

[0097] Wherein: is the kurtosis of the data segment; is the average value of this data segment; is the 𝑖-th sampling value within the data segment.

[0098] In S2, frequency domain analysis is used to perform the conversion from time domain to frequency domain for each data segment, and then the extraction of the frequency centroid is carried out;

[0099] Specifically, the conversion from time domain to frequency domain for each data segment is calculated by the formula:

[0100] ;

[0101] Wherein: is the transformed frequency domain value of the data segment; is the number of discrete sampling points of the data segment; is the serial number of the time domain discrete value of the data segment; is the n-th data sample of the data segment; is the imaginary unit.

[0102] The calculation formula for frequency centroid extraction is:

[0103] ;

[0104] Wherein: is the frequency centroid of the data segment; is the frequency value of the k-th spectral line; is the total number of spectral lines.

[0105] In S2, the fault characteristic frequencies of the bearing include:

[0106] Inner race fault characteristic frequency:

[0107] ;

[0108] Outer race fault characteristic frequency:

[0109] ;

[0110] Rolling element fault characteristic frequency:

[0111] ;

[0112] Cage fault characteristic frequency:

[0113] ;

[0114] Wherein: is the number of rolling elements; is the diameter of the rolling element; is the pitch diameter of the bearing; is the contact angle; is the rotational frequency of the main shaft, obtained according to the main shaft speed.

[0115] The fault characteristic frequency of the main shaft is:

[0116] ;

[0117] Wherein: is the multiple number; is the rotational frequency of the main shaft, obtained according to the main shaft speed; is the rotational disturbance frequency of the main shaft.

[0118] In S2, the root mean square, kurtosis, frequency centroid of each acceleration signal, as well as the fault characteristic frequency of the bearing and the fault characteristic frequency of the main shaft are normalized and then output.

[0119] The normalization process of the vibration data in this embodiment adopts the most conventional method, and the calculation method is: the difference between the magnitude of the current data and the minimum value of this type of data, divided by the difference between the maximum value and the minimum value of this type of data, to obtain a number between 0 and 1.

[0120] The obtained root mean square, kurtosis, frequency centroid, fault characteristic frequency of the main shaft, and fault characteristic frequency of the main shaft bearing are normalized and used as inputs, and input into the remaining life prediction model to predict the remaining life of the target thin film evaporator. The specifically predicted remaining life includes the remaining life of the main shaft and / or the remaining life of the main shaft bearing.

[0121] In S3, for the multiple vibration vector data obtained by vector addition of the acceleration signals in the three directions of each vibration sensor, a one-dimensional convolutional neural network model including convolution, pooling, activation function, and fully connected layers is built; the response trend of the vibration response of each vibration vector data changing with time is extracted through the one-dimensional convolutional neural network model of each vibration vector data to obtain the vibration vector data characteristic curves of the upper and lower bearings of the main shaft and the main shaft; the model is used to capture the changing trend presented by the vibration response of the vibration vector data over time, and this trend reflects the characteristic data of the equipment life degradation, presenting as a curve as shown in Figure 7 shown. Figure 7 In, the abscissa is time, and the data on the curve reflects the relationship between vibration data and life at the time point.

[0122] In S3, the one-dimensional convolutional neural network model includes:

[0123] Build a one-dimensional convolutional neural network model containing convolutional, pooling, activation function, and fully connected layers for each vibration vector data:

[0124] Input sequence: ;

[0125] The input layer receives vibration vector data, where: is the input value at the t-th position in the sequence;

[0126] Convolution operation: ;

[0127] where: is the output value at the t-th position after the convolution operation; is the size of the convolutional kernel; is the weight parameter at the i-th position in the convolutional kernel; is the corresponding element in the input sequence participating in the convolution calculation; is the bias term of the convolution operation.

[0128] Adopt the RELU activation function: ;

[0129] where: is the output value after activation at the t-th position.

[0130] Adopt the max pooling operation: ;

[0131] where: is the output value of the t-th pooling window; is the size of the pooling window; is the value at the i-th position in the output of the activation function.

[0132] Fully connected layer: ;

[0133] where: is the value of the j-th output node; is the weight parameter connecting the i-th input node to the j-th output node; is the output value of the i-th node in the hidden layer; is the bias term of the j-th output node.

[0134] Adopt the cross-entropy loss function: ;

[0135] where: is the cross-entropy loss value; is the total number of classes; The class in the true label value; is the Probability values of each category.

[0136] The output of the one-dimensional convolutional neural network model changes over time to obtain Figure 7 The changing trend of the vibration response of the vibration vector data shown over time.

[0137] S4. Use multiple high-order exponential functions to fit the vibration vector data characteristic curves of the upper and lower bearings of the main shaft and the main shaft obtained in step S3 respectively to form the remaining life characteristic curves of the upper bearing of the main shaft, the remaining life characteristic curves of the lower bearing of the main shaft, and the remaining life characteristic curves of the main shaft, and then express them in percentage form;

[0138] In S4, the vibration vector data characteristic curve of the main shaft is obtained by taking the average value of the vibration vector data characteristic curves corresponding to the vibration sensors in the middle area of the cylinder.

[0139] In S4, the formula for expressing the remaining life characteristic curves of the upper bearing of the main shaft, the remaining life characteristic curves of the lower bearing of the main shaft, and the remaining life characteristic curves of the main shaft in percentage form by combining their expressions is as follows:

[0140] ;

[0141] In the formula: , , , , are adjustable parameters. The specific debugging process is as follows: First, set the initial values to 1. Further observe the fitting situation between the vibration vector data characteristic curve and the remaining life characteristic curve obtained after fitting. Then adjust , , , , until the vibration vector data characteristic curve fits well with the remaining life characteristic curve obtained after fitting; where is the system time of the vibration data; is the time limit defined for the equipment to change from completely normal to abnormal; is the remaining life percentage; is the natural base.

[0142] S5. Taking the time interval set in step S2 as the time resolution, the root mean square, kurtosis, frequency centroid, fault characteristic frequencies of the main shaft bearings, and fault characteristic frequencies of the main shaft obtained in step S2, as well as the data on the remaining life characteristic curves of the upper main shaft bearing, lower main shaft bearing, and main shaft obtained in step S4, these data together constitute a data set, where the data on the remaining life characteristic curves of the upper main shaft bearing, lower main shaft bearing, and main shaft are label data;

[0143] In practical applications, for the sake of easy understanding, the root mean square, kurtosis, frequency centroid, fault characteristic frequencies of the main shaft bearings, and fault characteristic frequencies of the main shaft obtained in step S2, as well as the data on the remaining life characteristic curves of the upper main shaft bearing, lower main shaft bearing, and main shaft obtained in step S4, are plotted on the same two-dimensional coordinate. Each moment on the two-dimensional coordinate corresponds to a series of data, and these data together constitute a data set.

[0144] S6. Establish a remaining life prediction model, and divide the data set obtained in step S5 into a training set, a validation set, and a test set; according to the training process of the deep learning model, use the divided data set to train, validate, and test the remaining life prediction model. Taking the ratio of 8:1:1 as an example for dividing the training set, validation set, and test set, finally obtain a remaining life prediction model with the accuracy meeting the requirements.

[0145] In S6, based on the one-dimensional convolutional neural network model, a remaining life prediction model is constructed, and the root mean square, kurtosis, frequency centroid, fault characteristic frequencies of the main shaft, and fault characteristic frequencies of the main shaft bearings obtained by processing the real-time vibration data and main shaft speed acquired in step S7 in the same way as in step S2 are input into the remaining life prediction model in real time.

[0146] In the one-dimensional convolutional neural network model, the input sequence: ;

[0147] ;

[0148] Among them, represents the remaining life corresponding to the t-th data segment calculated according to the remaining life characteristic curve; , , , , , , , They respectively correspond to the normalized root mean square, kurtosis, frequency centroid, inner race fault characteristic frequency, outer race fault characteristic frequency, rolling element fault characteristic frequency, cage fault characteristic frequency, and spindle fault characteristic frequency corresponding to the t-th data segment.

[0149] S7. Obtain the real-time vibration data and spindle speed of the target thin film evaporator; the root mean square, kurtosis, frequency centroid, spindle fault characteristic frequency, and spindle bearing fault characteristic frequency obtained after processing are used as the inputs of the remaining life prediction model to predict the remaining life of the target thin film evaporator spindle and / or the remaining life of the spindle bearing.

[0150] Since the mechanical system vibrates during operation, the vibration will change when components fail, and different degrees of failure result in different vibration responses. Therefore, it is possible to predict the time of component failure or even failure by analyzing the characteristic changes of the vibration response, that is, to achieve life prediction. The root mean square, kurtosis, frequency centroid, spindle fault characteristic frequency, and spindle bearing fault characteristic frequency reflect the characteristics of the vibration response, and these characteristics can be used to describe the vibration situation. The remaining life prediction model is established through a deep learning model, and this establishment process is achieved by training the deep learning model with the characteristic data of the vibration response in the whole life cycle of the equipment components. The deep learning model learns the characteristics of the component degradation process through a large amount of data training, and then realizes the prediction of the component life.

[0151] Embodiment 2:

[0152] This embodiment provides a method for establishing a remaining life prediction model for a large thin film evaporator. During implementation, the real-time vibration data includes 1 vibration sensor respectively arranged at the upper and lower ends of the thin film evaporator cylinder body and 4 vibration sensors in four areas of the cylinder body, a total of 18 acceleration signals in three directions collected by 6 vibration sensors. Here, the three directions refer to the x-axis, y-axis, and z-axis directions; among them, the vibration sensor at the upper end of the cylinder body is used to monitor the fault condition of the upper bearing of the spindle, the vibration sensor at the lower end of the cylinder body is used to monitor the fault condition of the lower bearing of the spindle, and the 4 vibration sensors in the four areas of the cylinder body are used to monitor the fault condition of the spindle. The obtained 18 acceleration signals are respectively segmented according to a set time interval to obtain multiple data segments with the set time interval; time domain analysis and frequency domain analysis are respectively performed on each data segment to obtain the root mean square, kurtosis, and frequency centroid of each vibration data segment; the fault characteristic frequency of the spindle is obtained according to the obtained spindle speed data, and then the fault characteristic frequency of the spindle bearing is obtained according to the spindle speed data and the size parameters of the spindle bearing.

[0153] The method for establishing the remaining life prediction model includes:

[0154] S1. Obtain the vibration data and spindle speed data of the thin-film evaporator during the whole process from normal operation to the failure of the spindle and spindle bearings, which leads to the abnormal operation of the thin-film evaporator.

[0155] The vibration data includes 18 acceleration signals in three directions collected by 6 vibration sensors, namely 1 vibration sensor respectively arranged at the upper and lower ends of the cylinder body of the thin-film evaporator and 4 vibration sensors in four areas of the cylinder body. Among them, the vibration sensor at the upper end of the cylinder body is used to monitor the failure condition of the upper-end bearing of the spindle, the vibration sensor at the lower end of the cylinder body is used to monitor the failure condition of the lower-end bearing of the spindle, and the 4 vibration sensors in four areas of the cylinder body are used to monitor the failure condition of the spindle.

[0156] S2. Segment the 18 acceleration signals obtained in step S1 at set time intervals to obtain multiple data segments at the set time intervals; perform time-domain analysis and frequency-domain analysis on each data segment respectively to obtain the root mean square, kurtosis and frequency centroid of each vibration data segment; obtain the fault characteristic frequency of the spindle according to the spindle speed data obtained in step S1, and then obtain the fault characteristic frequency of the spindle bearing according to the spindle speed data and the size parameters of the spindle bearing.

[0157] S3. Perform vector addition on the acceleration signals in three directions of each vibration sensor of the vibration data obtained in step S1 to obtain 6 vibration vector data, and build a one-dimensional convolutional neural network model containing convolution, pooling, activation function and fully connected layer for each vibration vector data; extract the response trend of the vibration response of each vibration vector data changing with time through the one-dimensional convolutional neural network model of each vibration vector data.

[0158] S4. Use multiple high-order exponential functions to fit the vibration vector data characteristic curves of the upper and lower bearings of the spindle and the spindle obtained in step S3 to form the remaining life characteristic curves of the upper-end bearing of the spindle, the remaining life characteristic curves of the lower-end bearing of the spindle and the remaining life characteristic curves of the spindle, and express the remaining life characteristic curves of the upper-end bearing of the spindle, the remaining life characteristic curves of the lower-end bearing of the spindle and the remaining life characteristic curves of the spindle in percentage form after fitting.

[0159] Taking the set time interval in step S2 as the time resolution, using the response time of the vibration data as the abscissa, the root mean square, kurtosis, frequency centroid, the fault characteristic frequency of the spindle bearing and the fault characteristic frequency of the spindle obtained in step S2, and the data on the remaining life characteristic curves of the upper-end bearing of the spindle, the remaining life characteristic curves of the lower-end bearing of the spindle and the remaining life characteristic curves of the spindle obtained in step S4, these data together constitute a data set, and the data on the remaining life characteristic curves of the upper-end bearing of the spindle, the remaining life characteristic curves of the lower-end bearing of the spindle and the remaining life characteristic curves of the spindle are label data.

[0160] For the sake of easy understanding, in step S5, taking the time interval set in step S2 as the time resolution, using the response time of the vibration data as the abscissa, plotting the root mean square, kurtosis, frequency centroid, fault characteristic frequencies of the main shaft bearings, and fault characteristic frequencies of the main shaft obtained in step S2, as well as the remaining life characteristic curves of the upper bearing of the main shaft, the remaining life characteristic curve of the lower bearing of the main shaft, and the remaining life characteristic curve of the main shaft obtained in step S4 onto the same two-dimensional coordinate. Each moment on the two-dimensional coordinate corresponds to a series of data, and these data together constitute a data set, where the data on the remaining characteristic curves are labeled data.

[0161] S6. Establish a remaining life prediction deep learning model, divide the data set obtained in step S5 into a training set, a validation set, and a test set according to the ratio of 8:1:1; according to the training process of the deep learning model, use the divided data set to train, validate, and test the remaining life prediction deep learning model, and finally obtain a remaining life prediction model with the required accuracy.

[0162] After establishing the remaining life prediction model, obtain the vibration data collected by a total of 6 vibration sensors at the upper and lower ends of the cylinder body of the target thin film evaporator and in four areas of the cylinder body, and at the same time obtain the main shaft speed from the central control platform of the target thin film evaporator. The target thin film evaporator is the thin film evaporator for which the remaining life needs to be predicted.

[0163] Segment the obtained vibration data at a set time interval to obtain multiple vibration data segments of the set time interval; perform time-domain analysis and frequency-domain analysis on each data segment respectively to obtain the root mean square, kurtosis, and frequency centroid of each vibration data segment; obtain the fault characteristic frequency of the main shaft according to the main shaft speed data, and obtain the fault characteristic frequency of the main shaft bearing according to the main shaft speed data and the dimensional parameters of the main shaft bearing; normalize and output the root mean square, kurtosis, and frequency centroid of the 18 acceleration signals in three directions of the 6 vibration sensors, as well as the fault characteristic frequencies of the bearing and the fault characteristic frequencies of the shaft.

[0164] Take the root mean square, kurtosis, and frequency centroid of the 18 acceleration signals in three directions of the 6 vibration sensors of the obtained target thin film evaporator, as well as the fault characteristic frequencies of the main shaft of the target thin film evaporator and the fault characteristic frequencies of the main shaft bearing as inputs, and input them into the remaining life prediction model established by using the method for establishing the remaining life prediction model of the thin film evaporator in steps S1 - S6 of this embodiment to predict the remaining life of the target thin film evaporator. The specifically predicted remaining life includes the remaining life of the main shaft and / or the remaining life of the main shaft bearing.

[0165] As Figure 1 shown, a plurality of vibration sensors distributed in the preset areas on the surface of the cylinder body of the target thin film evaporator collect the vibration data of the upper and lower bearings of the main shaft and the vibration data of the main shaft in real time.

[0166] Figure 1 Among the six sensors, vibration sensor 1 is used to collect the vibration data of the upper bearing in the upper region of the thin-film evaporator, vibration sensor 6 is used to collect the vibration data of the lower bearing in the lower region, and vibration sensors 2, 3, 4, and 5 are used to collect the vibration data of the main shaft in the middle region. During implementation, preferably, vibration sensor 1 is installed on the cylinder body in the area corresponding to the upper bearing and the upper mounting plate, and vibration sensor 2 is installed on the cylinder body in the area corresponding to the lower bearing and the lower mounting plate. The vibration data includes 18 acceleration signals in three directions collected by a total of six vibration sensors at the upper and lower ends of the thin-film evaporator cylinder body and in four regions of the cylinder body; among them, the vibration sensors at the upper end of the cylinder body are used to monitor the fault conditions of the upper bearing of the main shaft, the vibration sensors at the lower end of the cylinder body are used to monitor the fault conditions of the lower bearing of the main shaft, and the four vibration sensors in the four regions of the cylinder body are used to monitor the fault conditions of the main shaft.

[0167] After preprocessing the vibration data, the root mean square, kurtosis, and frequency centroid are obtained, and the fault characteristic frequencies of the bearing and the main shaft are calculated. After normalizing these data, they are input into the remaining life prediction model established by the method for establishing the remaining life prediction model in real time. The remaining life prediction model predicts the main shaft and the main shaft bearing and outputs the remaining life of the main shaft and the main shaft bearing.

[0168] The normalization process of the vibration data in this embodiment adopts the most conventional method, and the calculation method is: the difference between the current data size and the minimum value of this type of data is divided by the difference between the maximum value and the minimum value of this type of data to obtain a number between 0 and 1.

[0169] Schematic diagram of the process of the remaining life prediction module Figure 3 As shown, this module takes the output of data preprocessing as input and then predicts the remaining life of the target thin-film evaporator. The specific remaining life predicted includes the remaining life of the main shaft and the remaining life of the main shaft bearing.

[0170] The remaining life prediction module adopts a remaining life prediction method driven by both knowledge and data and is designed based on a deep learning network. Schematic diagram of the design process of the remaining life label data in the remaining life prediction Figure 4 As shown, the design method is as follows: First, design the label data set. First, in multiple tests, 18 acceleration signals in three directions of six vibration sensors are obtained through the data acquisition module during the whole process from the equipment running in good condition to the equipment being unable to operate normally due to the failure of each of the main shaft and the main shaft bearing. Then, the vibration accelerations in three directions of each sensor are vectorially added to obtain 6 vibration vector data. A schematic diagram of one vibration vector data is as Figure 6As shown, the amplitude represents the magnitude of the vibration acceleration collected by the vibration sensor, and the sample points represent the number of sampling points of the vibration acceleration collected by the vibration sensor. The number of sampling points of the collected vibration acceleration.

[0171] Then, for each vibration vector data, build a one-dimensional convolutional neural network model that includes convolution, pooling, activation function, and fully connected layers:

[0172] Build a one-dimensional convolutional neural network model that includes convolution, pooling, activation function, and fully connected layers for each vibration vector data:

[0173] Input sequence: ;

[0174] Where: is the input value at the t-th position in the sequence;

[0175] Convolution operation: ;

[0176] Where: is the output value at the t-th position after the convolution operation; is the size of the convolution kernel; is the weight parameter at the i-th position in the convolution kernel; is the corresponding element in the input sequence participating in the convolution calculation; is the bias term of the convolution operation;

[0177] Adopt the RELU activation function: ;

[0178] Where: is the output value after activation at the t-th position;

[0179] Adopt the max pooling operation: ;

[0180] Where: is the output value of the t-th pooling window; is the size of the pooling window; is the value at the i-th position in the output of the activation function;

[0181] Fully connected layer: ;

[0182] Where: is the value of the j-th output node; is the weight parameter connecting the i-th input node to the j-th output node; is the output value of the i-th node in the hidden layer; is the bias term of the j-th output node;

[0183] Adopt the cross-entropy loss function: ;

[0184] In the formula: is the cross-entropy loss value; is the total number of categories; the value of category in the true label; is the probability value of the th category predicted by the model.

[0185] The response trend of the vibration response over time is extracted by this one-dimensional convolutional neural network model. Finally, six vibration vector data characteristic curves formed by the vibration response trends of the upper and lower bearings of the main shaft and the main shaft will be obtained. Figure 7 Figure 1 is a schematic diagram of one of the vibration vector data characteristic curves. Among them, the amplitude represents the magnitude of the vibration acceleration collected by the vibration sensor, and the sample points represent the sampling points of the vibration acceleration collected by the vibration sensor. The obtained vibration vector data characteristic curve is significantly characteristic. As time goes by, the damage of the main shaft and the main shaft bearings gradually accumulates and quickly collapses when it accumulates to a certain level. This phenomenon is the same as the damage phenomenon of mechanical parts in reality. Therefore, this curve can be used as a reference for estimating the remaining life of the main shaft and the main shaft bearings.

[0186] Among these six vibration vector data characteristic curves, two vibration vector data characteristic curves are derived from the data of the upper and lower sensors of the thin-film evaporator, respectively reflecting the health status of the upper and lower main shaft bearings. There are also vibration vector data characteristic curves derived from the data of the four vibration sensors on the cylinder of the thin-film evaporator, reflecting the health status of the main shaft. For the convenience of calculation, the data of these four vibration vector data characteristic curves are averaged to obtain a vibration vector data characteristic curve of the main shaft. Finally, a total of 3 vibration vector data characteristic curves reflecting the health status of the main shaft and the main shaft bearings of the thin-film evaporator will be obtained, namely 1 vibration vector data characteristic curve of the upper bearing of the main shaft, 1 vibration vector data characteristic curve of the lower bearing of the main shaft, and 1 vibration vector data characteristic curve of the main shaft.

[0187] However, the fluctuations of the obtained vibration vector data characteristic curve are relatively obvious, which does not conform to the damage accumulation of components leading to failure. In reality, the damage of components gradually accumulates over time and shows a changing trend, while the current curve has obvious fluctuations and does not match the progressive process of physical laws. This is because there are many interferences in the measured vibration signals. In view of this, a method combining an empirical mathematical model and big data feature extraction is used to design the remaining life characteristic curve. A multi-term high-order exponential function is used to fit the curve. At the same time, the mathematical models of the remaining life characteristic curves of the upper bearing of the main shaft, the lower bearing of the main shaft, and the remaining life characteristic curve of the main shaft expressed in percentage form are as follows:

[0188] ;

[0189] In the formula: 、 、 、 、 are adjustable parameters; is the abscissa of the vibration data, that is, the system time; is the defined time limit for the equipment to change from completely normal to abnormal; is the remaining life percentage; is the natural logarithm base. So far, with a time resolution of 0.5 s and the response time of the vibration data as the abscissa, the root mean square, kurtosis, frequency centroid, fault characteristic frequencies of the bearings, fault characteristic frequencies of the main shaft, and the data of the remaining life characteristic curve of the vibration data are plotted on the same two-dimensional coordinate. Each moment in this two-dimensional coordinate corresponds to a series of data, and these data together constitute a data set. Among them, the data on the remaining life characteristic curve are labeled data.

[0190] Furthermore, a remaining life prediction model is designed. This model is designed based on a deep learning model, and the design process is as shown in Figure 5 . 80% of the data set is used as the training set, 10% as the validation set, and the remaining 10% as the test set. Then, a convolutional deep learning network model for remaining life prediction is established. According to the conventional deep learning model development process, the divided data set is used to train, validate, and test the remaining life prediction deep learning model, and finally a remaining life prediction model with the required accuracy is obtained.

[0191] According to the above method, the design of the remaining life prediction module proposed in this application can be realized. In practical applications, vibration data in a preset area on the target thin evaporator cylinder is collected in real time through vibration sensors, the vibration data collected by the vibration sensors is preprocessed, and the characteristic frequencies of the key components of the thin film evaporator are calculated.

[0192] The preprocessing includes segmenting the vibration data. The data segmentation follows the principle of segmenting by time, dividing the continuously collected vibration data into multiple data segments at time intervals. The data segmentation divides the continuously collected vibration data into multiple data segments at a time interval of 0.5 seconds.

[0193] The preprocessing also includes time-domain analysis;

[0194] The time-domain analysis completes the extraction of the root mean square of the vibration data for each data segment. The calculation formula is:

[0195] ;

[0196] In the formula: is the root mean square of the data segment; is the number of samples in the data segment; is the 𝑖-th sampling value within the data segment;

[0197] The time-domain analysis completes the extraction of the kurtosis of the vibration data for each data segment. The calculation formula is:

[0198] ;

[0199] In the formula: is the kurtosis of the data segment; is the average value of this data segment; is the 𝑖-th sampling value within the data segment.

[0200] The preprocessing also includes frequency-domain analysis;

[0201] The frequency-domain analysis completes the conversion of the vibration data from the time domain to the frequency domain for each data segment. The calculation formula is:

[0202] ;

[0203] In the formula: is the transformed frequency-domain value of the data segment; is the number of discrete sampling points in the data segment; is the sequence number of the time-domain discrete value of the data segment; is the 𝑛-th data sample of the data segment; is the imaginary unit;

[0204] The frequency-domain analysis also includes the extraction of the frequency centroid. The calculation formula is:

[0205] ;

[0206] In the formula: is the frequency centroid of the data segment; is the frequency value of the 𝑘-th spectral line; is the total number of spectral lines.

[0207] Finally, the root mean square, kurtosis, and frequency centroid obtained from each vibration data are arranged in order and output.

[0208] The calculation of key characteristic frequencies includes the fault characteristic frequencies of the bearing, and the fault characteristic frequencies of the bearing include:

[0209] Inner ring fault characteristic frequency:

[0210] ;

[0211] Outer ring fault characteristic frequency:

[0212] ;

[0213] Rolling element fault characteristic frequency:

[0214] ;

[0215] Cage fault characteristic frequency:

[0216] ;

[0217] In the formula: is the number of rolling elements; is the diameter of the rolling element; is the pitch diameter of the bearing; is the contact angle; is the rotational frequency of the shaft, obtained from the spindle speed.

[0218] The calculation of the key component characteristic frequency also includes the fault characteristic frequency of the spindle, and the fault characteristic frequency model of the spindle is:

[0219] ;

[0220] In the formula: is the multiple frequency number; is the rotational frequency of the spindle, and the rotational frequency per minute is obtained from the spindle speed; is the rotational disturbance frequency of the spindle.

[0221] The remaining life prediction module inputs the root mean square, kurtosis, frequency centroid, bearing fault characteristic frequency, and spindle fault characteristic frequency of the normalized vibration response data output from the data preprocessing module in real time, and outputs the remaining life of the key vulnerable parts of the target thin film evaporator, thereby evaluating the remaining life of the entire device.

[0222] Example 3:

[0223] This embodiment provides a remaining life prediction device for a thin film evaporator used in the production of Lyocell fiber. The device includes:

[0224] A data acquisition module, which is used to obtain real-time vibration data collected by 6 vibration sensors, including 1 measuring point at the upper end of the cylinder body of the target thin-film evaporator, 1 measuring point at the lower end, and 4 measuring points in the middle of the cylinder body. At the same time, the spindle speed is obtained from the central control platform of the target thin-film evaporator.

[0225] A data preprocessing module, which is used to segment the vibration data obtained by the data acquisition module at set time intervals to obtain multiple vibration data segments of set time intervals; perform time-domain analysis and frequency-domain analysis on each data segment respectively to obtain the root mean square, kurtosis and frequency centroid of each vibration data segment; obtain the fault characteristic frequency of the spindle according to the obtained spindle speed data, and obtain the fault characteristic frequency of the spindle bearing according to the spindle speed data and the size parameters of the spindle bearing; finally, the data output by the data preprocessing module is the root mean square, kurtosis, frequency centroid, fault characteristic frequency of the spindle bearing and fault characteristic frequency of the spindle of 18 acceleration signals in three directions of 6 vibration sensors; the root mean square, kurtosis and frequency centroid of 18 acceleration signals in three directions of 6 vibration sensors, as well as the fault characteristic frequencies of the bearing and the spindle are normalized and then output.

[0226] A remaining life prediction module, in which the remaining life prediction model in Embodiment 1 or Embodiment 2 is encapsulated, and the data output by the data preprocessing module is used as the input data of the remaining life prediction model to predict the remaining life of the target thin-film evaporator. The specifically predicted remaining life includes the remaining life of the spindle and / or the remaining life of the spindle bearing.

[0227] This embodiment is also provided with a status display and alarm module, configured as:

[0228] Visualize the real-time remaining life prediction value;

[0229] If the remaining life of the spindle or the bearing is lower than the preset threshold, trigger an audible and visual alarm and push an alarm message. For example, during implementation, an alarm is issued when the remaining life of the target thin-film evaporator and its vulnerable parts is lower than 5% of the set remaining life threshold.

[0230] The vibration data is 18 acceleration signals in three directions of 6 vibration sensors. The normalization process of the vibration data uses the most conventional method. The calculation method is: the difference between the current data size and the minimum value of this type of data, divided by the difference between the maximum value and the minimum value of this type of data, to obtain a number between 0 and 1.

[0231] Schematic diagram of the process of the remaining life prediction module Figure 3As shown, this module takes the output of data preprocessing as input and then predicts the remaining life of the target thin-film evaporator. The specifically predicted remaining life includes the remaining life of the main shaft and the remaining life of the main shaft bearing.

[0232] The remaining life prediction module adopts a remaining life prediction method driven by both knowledge and data, and is designed relying on a deep learning network. Schematic diagram of the design process of the remaining life label data in the remaining life prediction Figure 4 As shown, the design method is as follows: Design the label data set. First, obtain 18 acceleration signals in three directions of six vibration sensors during the whole process of the equipment running from intact to the failure of each of the main shaft and the main shaft bearing, which causes the equipment to malfunction, in multiple tests. Then, vectorially add the vibration accelerations in the three directions of each sensor to obtain 6 vibration vector data. Then, for each vibration vector data, build a one-dimensional convolutional neural network model containing convolutional, pooling, activation function, and fully connected layers:

[0233] Input sequence: ;

[0234] In the formula: is the input value at the t-th position in the sequence;

[0235] Convolution operation: ;

[0236] In the formula: is the output value at the t-th position after the convolution operation; is the size of the convolution kernel; is the weight parameter at the i-th position in the convolution kernel; is the element corresponding to the convolution calculation in the input sequence; is the bias term of the convolution operation;

[0237] Adopt the RELU activation function: ;

[0238] In the formula: is the output value after activation at the t-th position;

[0239] Adopt the max pooling operation: ;

[0240] In the formula: is the output value of the t-th pooling window; is the size of the pooling window; is the value at the i-th position in the output of the activation function;

[0241] Fully connected layer: ;

[0242] In the formula: is the value of the j-th output node; is the weight parameter connecting the i-th input node to the j-th output node; is the output value of the i-th node in the hidden layer; is the bias term of the j-th output node;

[0243] The cross-entropy loss function is adopted: ;

[0244] In the formula: is the cross-entropy loss value; is the total number of categories; the value of category in the true label; is the probability value of the -th category predicted by the model.

[0245] The response trend of the vibration response over time is extracted through this model. Finally, six vibration vector data feature curves will be obtained. Figure 7 is a schematic diagram of a vibration vector data feature curve. The obtained vibration vector data feature curves are distinct. As time goes by, the damage to the main shaft and main shaft bearings gradually accumulates and rapidly collapses when it accumulates to a certain level. This phenomenon is the same as the damage phenomenon of mechanical parts in actual situations. Therefore, the vibration vector data feature curves can be used as a reference for estimating the remaining life of the main shaft and main shaft bearings.

[0246] Among these six vibration vector data feature curves, two vibration vector data feature curves are derived from the data of the upper and lower sensors of the thin-film evaporator, respectively reflecting the health conditions of the upper and lower main shaft bearings. The remaining four vibration vector data feature curves are derived from the data of the four vibration sensors in the middle area of the thin-film evaporator cylinder, reflecting the health condition of the main shaft. For the convenience of calculation, the data of these four vibration vector data feature curves are averaged to obtain a vibration vector data feature curve of the main shaft. Finally, a total of 3 vibration vector data feature curves reflecting the health conditions of the main shaft and main shaft bearings of the thin-film evaporator will be obtained, namely 1 vibration vector data feature curve of the upper main shaft bearing, 1 vibration vector data feature curve of the lower main shaft bearing, and 1 vibration vector data feature curve of the main shaft. A vibration vector data feature curve of the main shaft.

[0247] However, the obtained vibration vector data characteristic curve still has defects. For example, the fluctuations are obvious, which does not conform to the damage accumulation of components leading to failure. This is because there are many interferences in the measured vibration signals. In view of this, a method combining an empirical mathematical model and big data feature extraction is used to design the remaining life characteristic curve. A multi-term high-order exponential function is used to fit this curve. At the same time, combined with the expression of this vibration vector data characteristic curve, the mathematical model expressing the remaining life characteristic curve in percentage form is as follows:

[0248] ;

[0249] In the formula: 、 、 、 、 are adjustable parameters; is the system time of the vibration data; is the defined time limit for the equipment to change from completely normal to abnormal; is the remaining life percentage; is the natural base. So far, with a time resolution of 0.5 s, taking the response time of the vibration data as the abscissa, the root mean square, kurtosis, frequency centroid of the vibration data, the fault characteristic frequency of the bearing, the fault characteristic frequency of the main shaft, and the remaining life characteristic curve data are plotted on the same two-dimensional coordinate. Each moment in this two-dimensional coordinate corresponds to a series of data, and these data together constitute a data set. Among them, the data on the remaining life characteristic curve are label data.

[0250] Further design a remaining life prediction model. This model is designed based on a deep learning model, and the design process is as Figure 5 . Take 80% of the data set as the training set, 10% as the validation set, and the remaining 10% as the test set. Then, establish a one-dimensional convolutional deep learning network model for remaining life prediction. According to the conventional deep learning model development process, use the divided data set to train, validate, and test the remaining life prediction deep learning model, and finally obtain a remaining life prediction model with the required accuracy.

[0251] According to the above method, the design of the remaining life prediction module proposed in this application can be realized. In practical applications, vibration data in a preset area on the target thin evaporator cylinder is collected in real time through a vibration sensor, the vibration data collected by the vibration sensor is preprocessed, and the characteristic frequencies of the key components of the thin film evaporator are calculated.

[0252] The preprocessing includes segmenting the vibration data. The data segmentation adopts the principle of segmenting by time, and the continuously collected vibration data is divided into multiple data segments according to the time interval. The data segmentation divides the continuously collected vibration data into multiple data segments according to a time interval of 0.5 seconds.

[0253] The preprocessing also includes time-domain analysis;

[0254] The time-domain analysis completes the extraction of the root mean square of the vibration data for each data segment, and its calculation formula is:

[0255] ;

[0256] In the formula: is the root mean square of the data segment; is the number of samples of the data segment; is the 𝑖-th sampling value within the data segment;

[0257] The time-domain analysis completes the extraction of the kurtosis of the vibration data for each data segment, and its calculation formula is:

[0258] ;

[0259] In the formula: is the kurtosis of the data segment; is the average value of this data segment; is the 𝑖-th sampling value within the data segment.

[0260] The preprocessing also includes frequency-domain analysis;

[0261] The frequency-domain analysis completes the conversion of the vibration data of each data segment from the time domain to the frequency domain, and its calculation formula is:

[0262] ;

[0263] In the formula: is the transformed frequency-domain value of the data segment; is the number of discrete sampling points of the data segment; is the serial number of the time-domain discrete value of the data segment; is the 𝑛-th data sample of the data segment; is the imaginary unit;

[0264] The frequency-domain analysis also includes the extraction of the frequency centroid, and its calculation formula is:

[0265] ;

[0266] In the formula: is the frequency centroid of the data segment; is the frequency value of the 𝑘-th spectral line; is the total number of spectral lines.

[0267] Finally, the root mean square, kurtosis, and frequency centroid obtained from each vibration data are arranged in order and output.

[0268] The calculation of key feature frequencies includes the fault feature frequencies of the bearing, and the fault feature frequencies of the bearing include:

[0269] Inner race fault feature frequency:

[0270] ;

[0271] Outer race fault feature frequency:

[0272] ;

[0273] Rolling element fault feature frequency:

[0274] ;

[0275] Cage fault feature frequency:

[0276] ;

[0277] Where: is the number of rolling elements; is the diameter of the rolling element; is the pitch diameter of the bearing; is the contact angle; is the rotational frequency of the shaft, obtained from the spindle speed.

[0278] The calculation of key component feature frequencies also includes the fault feature frequency of the spindle, and the fault feature frequency model of the spindle is:

[0279] ;

[0280] Where: is the multiple frequency number; is the rotational frequency of the spindle, obtained from the spindle speed; is the disturbance frequency.

[0281] Integrate the root mean square, kurtosis, and frequency centroid of each acceleration signal, as well as the fault feature frequencies of the bearing and the fault feature frequencies of the spindle into a data matrix, and output after normalizing the root mean square, kurtosis, frequency centroid, bearing fault feature frequencies, and spindle fault feature frequencies of each acceleration signal in the data matrix.

[0282] The remaining life prediction module will input in real time the root mean square, kurtosis, frequency centroid of each acceleration signal output from the data preprocessing module and normalized, the fault feature frequencies of the bearing, and the fault feature frequencies of the spindle, and output the remaining life of the key vulnerable parts of the target thin film evaporator, and then evaluate the remaining life of the entire equipment.

[0283] The above are only the preferred embodiments of the present invention, and the present invention is not limited thereto. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining life of a thin-film evaporator used in the production of Lyocell fibers, characterized in that, It includes the following steps: S1. Obtain the vibration data and spindle speed data of the thin-film evaporator during the whole process from normal operation to the failure of the main shaft and / or main shaft bearings, which leads to the abnormal operation of the thin-film evaporator; The vibration data includes the acceleration signals in three directions respectively collected by the vibration sensors at the upper and lower ends and the middle area of the cylinder body of the thin-film evaporator; among them, the vibration sensors at the upper and lower ends respectively monitor the failure conditions of the upper and lower bearings of the main shaft, and the vibration sensors in the middle area of the cylinder body monitor the failure conditions of the main shaft; S2. Perform time-domain analysis and frequency-domain analysis on multiple data segments obtained by segmenting the acceleration signals at set time intervals; among them, time-domain analysis is used to extract the root mean square and kurtosis of each data segment, and frequency-domain analysis is used to perform the conversion from time domain to frequency domain on each data segment and extract the frequency centroid; obtain the fault characteristic frequency of the main shaft according to the spindle speed data, and obtain the fault characteristic frequency of the main shaft bearings according to the spindle speed data and the size parameters of the main shaft bearings; S3. Build a one-dimensional convolutional neural network model including convolutional layer, pooling layer, activation function and fully connected layer for each of the multiple vibration vector data obtained by vector addition of the acceleration signals in three directions of each vibration sensor; extract the response trend of the vibration response of each vibration vector data changing with time through the one-dimensional convolutional neural network model of each vibration vector data, and obtain the vibration vector data characteristic curves of the upper and lower bearings of the main shaft and the main shaft. The vibration vector data characteristic curve of the main shaft is obtained by taking the average value of the vibration vector data characteristic curves corresponding to the vibration sensors in the middle area of the cylinder body; S4. Use multiple high-order exponential functions to fit the vibration vector data characteristic curves of the upper and lower bearings of the main shaft and the main shaft obtained in step S3 respectively to form the remaining life characteristic curves of the upper bearing of the main shaft, the remaining life characteristic curve of the lower bearing of the main shaft and the remaining life characteristic curve of the main shaft, and express them in percentage form; S5. With the set time interval in step S2 as the time resolution, the root mean square, kurtosis, frequency centroid, the fault characteristic frequency of the main shaft bearings and the fault characteristic frequency of the main shaft obtained in step S2, and the data on the remaining life characteristic curves of the upper bearing of the main shaft, the remaining life characteristic curve of the lower bearing of the main shaft and the remaining life characteristic curve of the main shaft obtained in step S4 together constitute a data set, where the data on the remaining life characteristic curves of the upper bearing of the main shaft, the remaining life characteristic curve of the lower bearing of the main shaft and the remaining life characteristic curve of the main shaft are label data; S6. Divide the data set obtained in step S5 into a training set, a validation set and a test set; after training, validation and testing, obtain a remaining life prediction model with the required accuracy; S7. Obtain the real-time vibration data and spindle speed of the target thin-film evaporator; the root mean square, kurtosis, frequency centroid, the fault characteristic frequency of the main shaft and the fault characteristic frequency of the main shaft bearings obtained after processing are used as the input of the remaining life prediction model to predict the remaining life of the main shaft of the thin-film evaporator and / or the remaining life of the main shaft bearings; 2. The method for predicting the remaining life of the thin film evaporator for Lyocell fiber production according to claim 1, characterized in that, In S2, the root mean square and kurtosis of each data segment are extracted by performing time-domain analysis on each data segment; Specifically, the root mean square of each data segment is extracted through time-domain analysis, and its calculation formula is: Where: Y tRMS is the root mean square of the data segment; n n is the number of samples of the data segment; x i is the i-th sampling value within the data segment; The kurtosis of each data segment is extracted through time-domain analysis, and its calculation formula is: Where: Y tK is the kurtosis of the data segment; u is the average value of this data segment; x i is the i-th sampling value within the data segment.

3. The method for predicting the remaining life of the thin-film evaporator for Lyocell fiber production according to claim 2, wherein In S2, frequency-domain analysis is used to perform time-domain to frequency-domain conversion on each data segment, and then the frequency centroid is extracted; Specifically, the time-domain to frequency-domain conversion of each data segment is performed, and its calculation formula is: Where: S (k) is the transformed frequency-domain value of the data segment; N is the number of discrete sampling points of the data segment; k k is the sequence number of the discrete time-domain value of the data segment; x[n] is the nth data sample of the data segment; j is the imaginary unit; The calculation formula for frequency centroid extraction is: Where: F zx is the frequency centroid of the data segment; f k is the frequency value of the k-th spectral line; K is the total number of spectral lines.

4. The method for predicting the remaining life of the thin film evaporator for the production of Lyocell fibers according to any one of claims 1-3, characterized in that, The set time interval is 0.5 s.

5. The method for predicting the remaining life of the thin-film evaporator used in the production of Lyocell fibers according to claim 3, wherein In S2, the fault characteristic frequencies of the bearing include: Inner race fault characteristic frequency: Outer race fault characteristic frequency: Rolling element fault characteristic frequency: Cage fault characteristic frequency: Where: Q is the number of rolling elements; d is the diameter of the rolling elements; D is the pitch circle diameter of the bearing; α is the contact angle; f r is the rotational frequency of the main shaft, obtained based on the main shaft speed.

6. The method for predicting the remaining life of the thin-film evaporator for Lyocell fiber production according to claim 5, characterized in that, The fault characteristic frequency of the main shaft is: f PA = n1f r + Δf; Where: n1 is the multiple frequency; f r is the rotation frequency of the main shaft, obtained according to the main shaft speed; Δf is the disturbance frequency of the main shaft rotation.

7. The method for predicting the remaining life of the thin-film evaporator for the production of Lyocell fibers according to claim 6, characterized in that, In S2, the root mean square, kurtosis, and frequency centroid of each acceleration signal, as well as the fault characteristic frequencies of the bearing and the main shaft, are normalized and then output.

8. The method for predicting the remaining life of the thin-film evaporator for producing lyocell fibers according to claim 1 or 7, characterized in that, In S3, the one-dimensional convolutional neural network model includes: A one-dimensional convolutional neural network model including convolutional, pooling, activation function, and fully connected layers is built for each vibration vector data: Input sequence: x = [x1, x2, …, x t ; where: x t is the input value at the t-th position in the sequence; Convolution operation: where: y t is the output value at the t-th position after the convolution operation; R is the size of the convolution kernel; w i is the weight parameter at the i-th position in the convolution kernel; x t+i-1 is the element corresponding to the input sequence that participates in the convolution calculation; b is the bias term of the convolutional operation; Use the ReLU activation function: a t = max(0, y t ); where: a t is the output value after activation at the t-th position; Adopt the maximum pooling operation: where: z t is the output value of the t-th pooling window; P is the size of the pooling window; a i is the value at the i-th position in the output of the activation function; Fully connected layer: O j = ∑ i W ji · h i + b j ; where: o j is the value of the j-th output node; W ji is the weight parameter connecting the i-th input node to the j-th output node; h i is the output value of the i-th node in the hidden layer; b j is the bias term of the j-th output node; The cross-entropy loss function is adopted: Where: L is the cross-entropy loss value; C is the total number of categories; y c is the value of category C in the true label; is the probability value of the C-th category predicted by the model.

9. The method for predicting the remaining life of the thin-film evaporator for the production of Lyocell fibers according to claim 8, characterized in that, Based on the one-dimensional convolutional neural network model, a remaining life prediction model is constructed. The real-time vibration data and the main shaft speed obtained in step S7 are input into the remaining life prediction model in real time after being processed in the same way as in step S2 to obtain the root mean square, kurtosis, frequency centroid, the fault characteristic frequency of the main shaft, and the fault characteristic frequency of the main shaft bearing.

10. The method for predicting the remaining life of the thin-film evaporator for the production of lyocell fibers according to claim 9, wherein In S4, the expressions of the remaining life characteristic curves of the upper bearing of the main shaft, the remaining life characteristic curve of the lower bearing of the main shaft, and the remaining life characteristic curve of the main shaft in percentage form are as follows: Where: a, b, c, d, and e1 are adjustable parameters; x is the system time of the vibration data; x T is the defined time limit for the device to change from completely normal to abnormal; y y is the percentage of remaining life; e is the natural base.

11. A device for predicting the remaining life of a thin-film evaporator used in the production of lyocell fibers, characterized in that, The device includes: A data acquisition module for acquiring real-time vibration data collected by vibration sensors at the upper and lower ends and the middle region of the cylinder of the target thin-film evaporator, and simultaneously obtaining the main shaft speed from the central control platform of the target thin-film evaporator; A data preprocessing module for segmenting the real-time vibration data acquired by the data acquisition module at a set time interval to obtain multiple vibration data segments of the set time interval; performing time-domain analysis and frequency-domain analysis on each data segment respectively to obtain the root mean square, kurtosis, and frequency centroid of each vibration data segment; obtaining the fault characteristic frequency of the main shaft according to the obtained main shaft speed data, and then obtaining the fault characteristic frequency of the main shaft bearing according to the main shaft speed data and the size parameters of the main shaft bearing; finally, the data output by the data preprocessing module is the root mean square, kurtosis, frequency centroid, the fault characteristic frequency of the main shaft bearing, and the fault characteristic frequency of the main shaft of the acceleration signals in three directions of each vibration sensor; the root mean square, kurtosis, and frequency centroid of the acceleration signals in three directions of each vibration sensor, as well as the fault characteristic frequencies of the bearing and the main shaft, are normalized and then output; Remaining life prediction module, in which a remaining life prediction model in the remaining life prediction method for the thin film evaporator used in the production of Lyocell fibers described in any one of the above claims 1-10 is encapsulated. Using the data output by the data preprocessing module as the input data of the remaining life prediction model, the remaining life of the target thin film evaporator is predicted. The specifically predicted remaining life includes the remaining life of the main shaft and / or the remaining life of the main shaft bearing.

Images