Bearing remaining service life prediction method, device, electronic equipment and storage medium

Abstract

The embodiments of the present invention provide a method, device, electronic device and storage medium for predicting the remaining useful life of a bearing, and relate to the technical field of industrial equipment health management. By acquiring and processing the historical working condition signals of the first bearing and the second bearing under different working conditions, time signal matrices are obtained respectively, and multi-dimensional features are extracted and fused for each time signal matrix respectively, and the corresponding comprehensive feature distributions are obtained respectively, which are used as the source comprehensive feature distribution and the target comprehensive feature distribution respectively. The source comprehensive feature distribution and the target comprehensive feature distribution are feature aligned, and the distance loss is calculated to determine whether the distance loss meets the preset distance loss. If so, the prediction regressor is trained using the source comprehensive feature distribution, and the target comprehensive feature distribution is predicted to obtain the remaining useful life of the second bearing. Thus, the model is trained using the labeled first bearing data so that the model can accurately predict the remaining useful life of the second bearing.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment health management, and in particular to a method, device, electronic device and storage medium for predicting the remaining useful life of a bearing. Background Art

[0002] In the long run, industrial machinery is prone to failure, which may lead to downtime, economic losses, and even accidents. Bearings are often exposed to complex operating conditions, and more than 40% of mechanical failures are related to bearings. Therefore, it is necessary to predict the remaining useful life (RUL) of bearings and diagnose the current health status of bearings in advance, so as to reduce economic losses and major safety accidents.

[0003] In the prior art, a deep learning prediction method is used to predict the RUL of bearings. This method trains data under one working condition to obtain a trained model, and uses the model to predict the RUL of bearings under other working conditions. However, in actual engineering, bearings operate under different operating conditions, which causes changes in data distribution. For example, changes in factors such as load, speed, and temperature will affect the degradation pattern of the bearing, resulting in inconsistent distribution of source domain data and target domain data. Therefore, directly using the effect of model training in the source domain on target domain data will result in low accuracy in the model's prediction of the RUL of bearings under other working conditions. Summary of the invention

[0004] In view of this, an object of an embodiment of the present invention is to provide a method, device, electronic device and storage medium for predicting the remaining useful life of a bearing to at least partially improve the above-mentioned problem.

[0005] In order to achieve the above purpose, the technical solution adopted by the embodiment of the present invention is as follows:

[0006] In a first aspect, an embodiment of the present invention provides a method for predicting the remaining service life of a bearing, the method comprising:

[0007] Acquire a historical operating condition signal of the first bearing and a historical operating condition signal of the second bearing, and process each of the historical operating condition signals respectively to obtain a time signal matrix; wherein the first bearing and the second bearing are in different operating conditions, the historical operating condition signal of the first bearing includes a remaining service life label, and the time signal matrix includes a time dimension and a signal dimension;

[0008] Extract and fuse multi-dimensional features of each of the time signal matrices to obtain corresponding comprehensive feature distributions;

[0009] Taking the comprehensive feature distribution corresponding to the first bearing as the source comprehensive feature distribution, taking the comprehensive feature distribution corresponding to the second bearing as the target comprehensive feature distribution, aligning the source comprehensive feature distribution with the target comprehensive feature distribution, and calculating the distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution;

[0010] Determining whether the distance loss meets a preset distance loss;

[0011] If not, return to the step of performing multi-dimensional feature extraction on each of the time signal matrices and fusing them to calculate the distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution, until the latest distance loss meets the preset distance loss;

[0012] If so, the prediction regressor is trained using the source comprehensive feature distribution to obtain a mature prediction regressor, and the target comprehensive feature distribution is input into the mature prediction regressor to obtain the remaining service life of the second bearing.

[0013] Optionally, the historical operating condition signal includes signals at multiple time points, and the signal at each time point is a waveform. The historical operating condition signals are processed separately to obtain time signal matrices, including:

[0014] Performing index calculation on each of the waveforms of each of the historical operating condition signals respectively, and obtaining a plurality of index data corresponding to each of the waveforms respectively;

[0015] Perform time window processing on each of the historical operating condition signals according to a preset time length to obtain a plurality of time window data; the time window data includes a time step and indicator data;

[0016] The time window data of each of the historical operating condition signals are assembled in time sequence to obtain time signal matrices.

[0017] Optionally, performing multi-dimensional feature extraction and fusion on each of the time signal matrices to obtain a corresponding comprehensive feature distribution includes:

[0018] Adding position information to each of the time signal matrices;

[0019] Use multi-head attention to perform signal and time encoding on each of the time signal matrices after adding the position information to obtain signal features and time features;

[0020] The signal feature and the time feature are fused to obtain a comprehensive feature distribution; wherein the feature fusion formula is:

[0021]

[0022] in, is the signal characteristic, is the time feature, is the trainable parameter matrix.

[0023] Optionally, the using of multi-head attention to perform signal and time encoding on each of the time signal matrices after adding the position information to obtain signal features and time features includes:

[0024] The time signal matrix is ​​processed by the following formula to generate a query matrix, a key matrix and a value matrix:

[0025]

[0026]

[0027]

[0028] in, is the query matrix, is the bond matrix, is the value matrix, is a trainable parameter, ;

[0029] The attention weight matrix of the query matrix and the key matrix in the time signal matrix is ​​calculated by the attention weight matrix calculation formula; the attention weight matrix calculation formula is:

[0030]

[0031] in, , , represents the time dimension of the time signal matrix, Represents the time dimension time step, represents the signal dimension of the time signal matrix, is the scaling factor;

[0032] According to the attention weight matrix and the value matrix, the head output is calculated by the head output calculation formula; the head output calculation formula is:

[0033]

[0034] By using the attention weight matrix calculation formula and the head output calculation formula, multiple head outputs are calculated, and the multiple head outputs are spliced ​​together to obtain signal features and time features; wherein the splicing formula is:

[0035]

[0036] in, is the parameter matrix, is the number of header outputs, .

[0037] Optionally, the feature alignment of the source comprehensive feature distribution and the target comprehensive feature distribution, and the calculated distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution, includes:

[0038] Using an alignment formula to shorten the feature distance between the source comprehensive feature distribution and the target comprehensive feature distribution, and calculating the overall distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution;

[0039] Calculating the feature distance loss of each feature of the source comprehensive feature distribution and each feature of the target comprehensive feature distribution;

[0040] The distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution is calculated based on the overall distance loss and the feature distance loss.

[0041] Optionally, the alignment formula is as follows:

[0042]

[0043]

[0044] in, is the overall distance loss, are the source comprehensive feature distribution and the target comprehensive feature distribution respectively, and Respectively represent the sample numbers of the source comprehensive feature distribution and the target comprehensive feature distribution, is the feature kernel, characterizes the internal alignment of the source composite feature distribution, characterizing the alignment of the source comprehensive feature distribution with the target comprehensive feature distribution, Characterize the internal alignment of the target comprehensive feature distribution.

[0045] Optionally, calculating the feature distance loss of each feature of the source comprehensive feature distribution and each feature of the target comprehensive feature distribution includes:

[0046] For the source comprehensive feature distribution and the target comprehensive feature distribution, each feature is taken as the feature center, and the average feature of all features in the neighborhood of each feature center is calculated; the calculation formula is:

[0047]

[0048] in, is the average feature, Before and after the feature Neighborhood features, for neighbor characteristics;

[0049] According to each of the average features, the first feature distance loss of each of the average features is calculated by a first distance loss calculation formula; the first distance loss calculation formula is:

[0050]

[0051] in, is the first feature distance loss, and Respectively and The average characteristics, is the total number of the average features, ;

[0052] According to the second distance loss calculation formula, the second feature distance loss between each feature in the neighborhood of each feature center and the average feature is calculated; the second distance loss calculation formula is:

[0053]

[0054] in, For each feature in the neighborhood of the feature center, is the average feature in the neighborhood of the feature center;

[0055] The characteristic distance loss is calculated according to the first characteristic distance loss and the second characteristic distance loss corresponding to the source comprehensive characteristic distribution and the target comprehensive characteristic distribution; the calculation formula of the characteristic distance loss is:

[0056]

[0057] in, is a penalty parameter used to balance the contribution of source features and target features. and is the first characteristic distance loss and the second characteristic distance loss of the source comprehensive characteristic distribution, and A first feature distance loss and a second feature distance loss are obtained for the target comprehensive feature distribution.

[0058] In a second aspect, an embodiment of the present invention provides a device for predicting the remaining service life of a bearing, the device comprising:

[0059] A time signal matrix acquisition unit is used to acquire a historical operating condition signal of the first bearing and a historical operating condition signal of the second bearing, and process each of the historical operating condition signals respectively to obtain a time signal matrix respectively; wherein the first bearing and the second bearing are in different operating conditions, the historical operating condition signal of the first bearing includes a remaining service life label, and the time signal matrix includes a time dimension and a signal dimension;

[0060] A feature extraction and fusion unit is used to extract and fuse multi-dimensional features of each of the time signal matrices to obtain corresponding comprehensive feature distributions;

[0061] a distance loss calculation unit, configured to use the comprehensive feature distribution corresponding to the first bearing as a source comprehensive feature distribution, use the comprehensive feature distribution corresponding to the second bearing as a target comprehensive feature distribution, perform feature alignment on the source comprehensive feature distribution and the target comprehensive feature distribution, and calculate the distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution;

[0062] A distance loss judgment unit is used to judge whether the distance loss meets the preset distance loss; if not, return to the step of performing multi-dimensional feature extraction on each of the time signal matrices and fusing them to calculate the distance loss of the source comprehensive feature distribution and the target comprehensive feature distribution, until the latest distance loss meets the preset distance loss;

[0063] The remaining service life prediction unit is used to, if yes, use the source comprehensive feature distribution to train the prediction regressor to obtain a mature prediction regressor, input the target comprehensive feature distribution into the mature prediction regressor, and obtain the remaining service life of the second bearing.

[0064] In a third aspect, an embodiment of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the above-described methods when executing the program.

[0065] In a fourth aspect, an embodiment of the present invention provides a storage medium having a computer program stored thereon, wherein the computer program implements any of the above-described methods when executed by a processor.

[0066] The embodiments of the present invention provide a method, device, electronic device and storage medium for predicting the remaining useful life of a bearing. The method aligns the source comprehensive features corresponding to the first bearing with the target comprehensive feature distribution corresponding to the second bearing so that the feature distributions of the two are similar. The source comprehensive features are used to train the prediction regressors respectively, and the trained prediction regressors are used to predict the target comprehensive feature distribution to obtain the remaining useful life of the second bearing. In this way, the historical operating condition signals of the bearing under only one operating condition can be RUL marked to accurately predict the remaining useful life of the bearing under other operating conditions.

[0067] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0069] Figure 1 A schematic structural block diagram of an electronic device provided by an embodiment of the present invention;

[0070] Figure 2 A schematic diagram of a flow chart of a method for predicting the remaining useful life of a bearing provided by an embodiment of the present invention;

[0071] Figure 3 A schematic diagram of feature extraction fusion provided by an embodiment of the present invention;

[0072] Figure 4 Another schematic diagram of a flow chart of a method for predicting the remaining useful life of a bearing provided by an embodiment of the present invention;

[0073] Figure 5 Another schematic diagram of a flow chart of a method for predicting the remaining useful life of a bearing provided by an embodiment of the present invention;

[0074] Figure 6 A schematic diagram of a technical solution of a method for predicting the remaining useful life of a bearing provided by an embodiment of the present invention;

[0075] Figure 7 A schematic diagram of feature extraction alignment provided by an embodiment of the present invention;

[0076] Figure 8 A schematic structural block diagram of a bearing remaining service life prediction device provided in an embodiment of the present invention.

[0077] Icon: 100-electronic device; 101-memory; 102-communication interface; 103-processor; 104-bus; 300-bearing remaining service life prediction device; 310-time signal matrix acquisition unit; 320-feature extraction and fusion unit; 330-distance loss calculation unit; 340-distance loss judgment unit; 350-remaining service life prediction unit. DETAILED DESCRIPTION

[0078] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0079] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0080] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0081] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0082] In the long run, industrial machinery is prone to failure, which may lead to downtime, economic losses, and even accidents. Bearings are often exposed to complex operating conditions, and more than 40% of mechanical failures are related to bearings. Therefore, it is necessary to predict the RUL of bearings and diagnose the current health status of bearings in advance to reduce economic losses and major safety accidents.

[0083] At present, the research on bearing RUL can be summarized into three categories: model-based methods, data-based methods, and hybrid methods.

[0084] (1) Model-based approach

[0085] Model-based approaches use statistical models to describe the degradation process of bearings and thus predict their RUL. These models usually rely on the degradation behavior of the equipment, physical laws, or expert knowledge. Statistical degradation models: These models build prediction models using historical degradation data of the equipment. Degradation models were built using random processes ranging from Brownian motion to fractional Brownian motion to predict the RUL of bearings. Long-range dependency degradation model: A new degradation model with long-range dependency based on the generalized Cauchy process was introduced, which achieved the maximum prediction range of the generalized Cauchy degradation model in the degradation sequence. A health index was created by combining a bidirectional gated recurrent unit with a Wiener process, which successfully predicted future degradation. Although this approach combines the advantages of deep learning and statistical degradation models, its accuracy relies on the prior knowledge of experts and the accuracy of the description of the bearing failure mechanism, which limits its application in dealing with complex systems.

[0086] (2) Data-based approach

[0087] Data-based methods predict the RUL of bearings by analyzing large amounts of historical data and advanced algorithms. These methods mainly rely on machine learning and deep learning techniques to automatically extract features and patterns from the data without the need for detailed physical models. Deep learning methods: In recent years, deep learning has become increasingly popular in processing time series data and has been successfully applied to RUL prediction. This method does not require a deep understanding of the physical properties of the system, can handle complex and high-dimensional data, and automatically learn models from the data. However, it requires a large amount of high-quality data, the model may lack interpretability, and it relies heavily on the quality and quantity of the data. Improper data processing may lead to inaccurate predictions. In addition, the construction of data-driven models relies only on historical data and does not have the ability to understand the principles of degradation.

[0088] (3) Hybrid approach

[0089] Hybrid methods combine model-based methods and data-based methods, using the advantages of both to build a more powerful prediction system. Model fusion: Improve the accuracy and robustness of predictions by combining physical models and data-driven models. Hybrid methods can make up for the shortcomings of a single method and improve the accuracy and interpretability of the system by combining physical models and data-driven models. However, their design and implementation are more complex and may require more computing resources and expertise.

[0090] Another method for predicting the RUL of bearings uses a deep learning prediction method. By training data under one working condition, a trained model is obtained, and the model is used to predict the RUL of bearings under other working conditions. However, in actual engineering, bearings operate under different operating conditions, which causes changes in data distribution. For example, changes in factors such as load, speed, and temperature will affect the degradation mode of the bearing, resulting in inconsistent distribution of source domain data and target domain data. Therefore, directly using the effect of model training in the source domain on target domain data will result in low accuracy in the model's prediction of the RUL of bearings under other working conditions.

[0091] Based on the above situation, an embodiment of the present invention provides a method, device, electronic device and storage medium for predicting the remaining useful life of a bearing. By acquiring and processing the historical operating condition signals of the first bearing and the second bearing under different operating conditions, corresponding comprehensive feature distributions are obtained respectively, which are used as the source comprehensive feature distribution and the target comprehensive feature distribution respectively. The source comprehensive features are respectively aligned with the target comprehensive feature distribution so that the feature distributions of the two tend to be similar. The prediction regressors are trained with the source comprehensive features respectively, and the trained prediction regressors are used to predict the target comprehensive feature distribution to obtain the remaining useful life of the second bearing. In this way, the historical operating condition signals of the bearing under only one operating condition can be RUL marked to accurately predict the remaining useful life of the bearing under other operating conditions.

[0092] To implement the process steps and functions of each example of the present invention, please refer to Figure 1 , Figure 1 A schematic structural block diagram of an electronic device provided in an embodiment of the present invention. The electronic device 100 includes a memory 101 and a processor 103, and the memory 101 and the processor 103 are electrically connected to each other directly or indirectly to realize data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses 104 or signal lines. The memory 101 can be used to store software programs and modules, and the processor 103 executes various functional applications and data processing by executing the software programs and modules stored in the memory 101.

[0093] The electronic device 100 may be, but is not limited to, a personal computer (PC), a server, a distributed computer, etc. It is understandable that the electronic device 100 is not limited to a physical server, but may also be a virtual machine on a physical server, a virtual machine built on a cloud platform, etc., which can provide the same functions as the server or virtual machine. The operating system of the electronic device 100 may be, but is not limited to, a Windows system, a Linux system, etc.

[0094] Among them, the memory 101 can be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), etc.

[0095] The communication connection between the electronic device 100 and an external device is achieved through at least one communication interface 102 (which can be wired or wireless).

[0096] The processor 103 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the embodiment of the present invention can be completed by the hardware integrated logic circuit in the processor 103 or the instructions in the form of software. The processor 103 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.

[0097] Understandably, Figure 1 The structure shown is for illustration only. The electronic device 100 may also include Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown. Figure 1Each component shown in the figure can be implemented by hardware, software or a combination thereof.

[0098] The following is an exemplary description of the method for predicting the remaining service life of a bearing provided by the present invention. Specifically, Figure 2 A flow chart of a method for predicting the remaining useful life of a bearing provided by an embodiment of the present invention is shown in FIG. Figure 2 , the execution subject of this method can be the above Figure 1 The electronic device 100 shown in FIG. 1 includes: Figure 2 The following steps are described:

[0099] S210: Acquire a historical operating condition signal of the first bearing and a historical operating condition signal of the second bearing, and process each historical operating condition signal respectively to obtain a time signal matrix.

[0100] The first bearing and the second bearing are in different working conditions, the historical working condition signal of the first bearing includes a remaining service life label, and the time signal matrix includes a time dimension and a signal dimension.

[0101] The first bearing and the second bearing are in different working conditions. It can be understood that the first bearing and the second bearing can be bearings of the same model or bearings with certain differences. They run on different machines or in different environments, resulting in large differences in the distribution of the working condition signals generated. Among them, the working condition signal can be a vibration signal, a speed signal, a temperature signal, etc., and each signal data is a waveform with a time series. Preferably, the vibration signal is obtained to predict the RUL of the bearing.

[0102] It can be understood that the first bearing and the second bearing referred to above do not represent only two bearings, and the operating condition signals of more than two bearings can also be obtained. As long as the operating condition signal of one of the bearings is marked with RUL, the method can be used to predict the RUL of the other bearings and any other bearing.

[0103] The acquired working condition signal is processed into a time signal matrix, including the time dimension and the signal dimension, so that the weighted features of the two dimensions of signal and time step can be taken into account, ensuring that the extracted features cover different aspects of the signal and increase comprehensiveness and accuracy.

[0104] S220: Perform multi-dimensional feature extraction and fusion on each time signal matrix to obtain corresponding comprehensive feature distributions.

[0105] The time signal matrix includes the time dimension and the signal dimension. By extracting features from the time dimension and the signal dimension respectively and fusing the extracted features to obtain a comprehensive feature distribution, more comprehensive signal information can be obtained.

[0106] S230: Taking the comprehensive feature distribution corresponding to the first bearing as the source comprehensive feature distribution, taking the comprehensive feature distribution corresponding to the second bearing as the target comprehensive feature distribution, aligning the source comprehensive feature distribution with the target comprehensive feature distribution, and calculating the distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution.

[0107] Since the historical operating condition signals of the first bearing and the second bearing are quite different, the comprehensive feature distribution after feature extraction and fusion is also quite different, so the two need to be aligned. The distance loss of the two images is calculated to verify whether the two comprehensive feature distributions are aligned.

[0108] S240: Determine whether the distance loss meets the preset distance loss; if not, execute step S250; if so, execute step S260.

[0109] S250: Return to the step of extracting multi-dimensional features from each time signal matrix and fusing them to calculate the distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution, until the latest distance loss meets the preset distance loss.

[0110] That is, when the distance loss does not meet the preset distance loss, the process returns to executing steps S220 to S240.

[0111] S260: Using the source comprehensive feature distribution to train the prediction regressor to obtain a mature prediction regressor, inputting the target comprehensive feature distribution into the mature prediction regressor to obtain the remaining service life of the second bearing.

[0112] That is, when the distance loss meets the preset distance loss, it is considered that the source comprehensive feature distribution and the target comprehensive feature distribution are already similar, and the source comprehensive feature distribution can be used to train the prediction regressor, and then the trained prediction regressor can be used to predict the RUL of the second bearing.

[0113] This method extracts features from the operating signals of the first bearing and the second bearing under two different operating conditions, fuses them and aligns them, uses the data with the RUL label of the first bearing to train the prediction regressor, and then uses the trained prediction regressor to predict the RUL of the second bearing, so that the RUL of the second bearing can be accurately predicted.

[0114] Furthermore, in order to make the information displayed by the working condition signal more comprehensive, statistical indicators can be extracted from the working condition signal. The historical working condition signal is specifically a waveform with a time sequence. The historical working condition signal includes signals at multiple time points, and the signal at each time point is a small waveform. Processing each historical working condition signal separately to obtain a time signal matrix can include the following steps:

[0115] S211: Calculate the index for each waveform of each historical operating condition signal to obtain a plurality of index data corresponding to each waveform.

[0116] Among them, the indicators may include time domain indicators, frequency domain indicators, and time-frequency domain indicators.

[0117] Time domain indicators may include: maximum value, minimum value, mean value, peak-to-peak value, absolute mean value, root mean square value, root square amplitude, standard deviation, kurtosis, skewness, margin indicator, waveform indicator, pulse indicator, and peak indicator.

[0118] Frequency domain indicators may include: center of gravity frequency, frequency root mean square, average frequency, and frequency variance.

[0119] Time-frequency domain indicators: We perform wavelet packet signal analysis on the operating condition signal, and then use the energy percentage of the last layer of sub-bands as the extracted time-frequency domain indicators.

[0120] The specific calculation methods of each indicator are shown in Table 1.

[0121]

[0122] S212: Perform time window processing on each historical operating condition signal according to a preset time length to obtain a plurality of time window data; wherein the time window data includes a time step and indicator data.

[0123] As calculated by the above indicators, after the processing of the historical operating condition signal is completed, each time point corresponds to 26 indicators. Now the historical operating condition signal is processed by time window, and the signal is divided into multiple time window data of preset duration. The signal of each time window data is taken as a sample, that is, each time window data contains indicators of multiple time points.

[0124] S213: Assemble the time window data of each historical operating condition signal in time sequence to obtain time signal matrices.

[0125] After processing the time window data, it is processed into a time signal matrix, whose horizontal axis can be the time dimension and the vertical axis can be the indicator dimension.

[0126] After the historical working condition signal is processed into a matrix of two dimensions, time and index, in order to extract more comprehensive signal information, a model can be used to extract and fuse features in the two dimensions. The model can be an ordered dual-attention transfer network (ODASTN) model, which is a neural network architecture that combines dual attention mechanism and knowledge transfer, and is used to process ordered data or sequences. Specifically, the above step S220 may include the following steps:

[0127] S221: Add position information to each time signal matrix.

[0128] Transformer can be used to process the time signal matrix and add position information, so that the model can capture the temporal relationship in the sequence data through position encoding.

[0129] S222: Use multi-head attention to perform signal and time encoding on each time signal matrix after adding position information to obtain signal features and time features.

[0130] S223: Fusing the signal features and the time features to obtain a comprehensive feature distribution.

[0131] For details, see Figure 3 , Figure 3 A schematic diagram of feature extraction fusion provided by an embodiment of the present invention, encoding the horizontal and vertical dimensions of the time signal matrix respectively is equivalent to encoding the signal and time respectively, and the structure of the time coding layer is the same as the signal coding layer. The only difference between them is that the time coding layer extracts features along the time step (horizontal) dimension, and the signal coding layer extracts features along the signal (vertical) dimension. After extracting features from the signal and time dimensions, the ODASTN model performs feature fusion on the two, and the features extracted from the signal encoder and the time encoder are recorded as and The formula for feature fusion is:

[0132]

[0133] in, is the trainable parameter matrix.

[0134] See again Figure 3 , we set the signal encoder and the time encoder in parallel. In this setting, the features of the signal dimension and the time step dimension can be extracted simultaneously. This design effectively avoids the mutual influence of the information of the two dimensions and helps to improve the performance of RUL prediction.

[0135] In a possible implementation, step S222 may include the following steps:

[0136] S2221: Process the time signal matrix to generate a query matrix, a key matrix, and a value matrix using the following formula:

[0137]

[0138]

[0139]

[0140] in, is the query matrix, is the key matrix, is the value matrix, is a trainable parameter, .

[0141] This step is the self-attention mechanism in the Transformer model, which enables the model to capture long-distance dependencies in sequence data.

[0142] S2222: Calculate the attention weight matrix of the query matrix and the key matrix in the time signal matrix using the attention weight matrix calculation formula; the attention weight matrix calculation formula is:

[0143]

[0144] in, , , represents the time dimension of the time signal matrix, Represents the time dimension time step, represents the signal dimension of the time signal matrix, is the scaling factor.

[0145] In this way, the correlation between different time and different signal dimensions in the sequence can be captured, thereby achieving a deep understanding of the sequence data. The computation of is the core of the self-attention mechanism, which enables the model to dynamically focus on the most important parts of the sequence.

[0146] S2223: Calculate the head output according to the attention weight matrix and the value matrix through the head output calculation formula; the head output calculation formula is:

[0147]

[0148] This formula can be used to calculate the head output of the self-attention mechanism.

[0149] S2224: Calculate multiple head outputs using the attention weight matrix calculation formula and the head output calculation formula, and splice the multiple head outputs together to obtain signal features and time features; wherein the splicing formula is:

[0150]

[0151] in, is the parameter matrix, is the number of header outputs, .

[0152] For the extracted source comprehensive feature distribution and target comprehensive feature distribution, since the two bearings are in different working conditions, it is necessary to align the features of the two. For details, see Figure 4 , Figure 4 Another schematic flow chart of a method for predicting the remaining service life of a bearing provided by an embodiment of the present invention, step S230 may include the following steps:

[0153] S231: Use the alignment formula to shorten the feature distance between the source comprehensive feature distribution and the target comprehensive feature distribution, and calculate the overall distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution.

[0154] The alignment formula may include multiple distance alignment formulas, such as Euclidean distance or Manhattan distance. Preferably, the Multi-Kernel Maximum Mean Discrepancy (MK-MMD) formula is used to shorten the feature distances between different domains. The MK-MMD formula is as follows:

[0155]

[0156]

[0157] in, is the overall distance loss, are the source comprehensive feature distribution and the target comprehensive feature distribution respectively, and Respectively represent the sample numbers of the source comprehensive feature distribution and the target comprehensive feature distribution, is the feature kernel, characterizes the internal alignment of the source composite feature distribution, characterizing the alignment of the source comprehensive feature distribution with the target comprehensive feature distribution, Characterize the internal alignment of the target comprehensive feature distribution.

[0158] Using this MK-MMD metric can reduce the cross-domain distance by learning invariant features in the high-level embedding space, which helps improve the model's generalization ability to different working conditions.

[0159] S232: Calculate the feature distance loss of each feature of the source comprehensive feature distribution and each feature of the target comprehensive feature distribution.

[0160] S233: Calculate the distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution according to the overall distance loss and the feature distance loss.

[0161] First, the alignment formula is used to calculate the overall distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution, then the feature distance loss between each feature is calculated, and finally the overall distance loss and the feature distance loss are combined to obtain the distance loss.

[0162] In one possible implementation, in order to improve the ability of the ODASTN model to learn high entropy feature representations, an ordered entropy regularizer can be introduced to calculate the feature distance loss. For details, see Figure 5 , Figure 5 Another schematic flow chart of a method for predicting the remaining useful life of a bearing provided by an embodiment of the present invention, step S232 may include the following steps:

[0163] S2321: For the source comprehensive feature distribution and the target comprehensive feature distribution, take each feature as the feature center and calculate the average feature of all features in the neighborhood of each feature center; the calculation formula is:

[0164]

[0165] in, is the average feature, Before and after the feature Neighborhood features, Neighborhood characteristics.

[0166] It is understandable that any feature is taken as the feature center, and its front and back The neighbor characteristics are , we can calculate the average features of all neighbors centered on any feature.

[0167] S2322: Calculate the first feature distance loss of each average feature according to the first distance loss calculation formula.

[0168] Since the feature space is unbounded, the features must be normalized. This can be done using Normalize all features:

[0169]

[0170] Although regularization It can indeed disperse the features to a greater extent, but it also breaks the regularity of the feature space. Therefore, it is preferred to use right The feature norms in are weighted. That is, the first distance loss calculation formula is:

[0171]

[0172] in, is the first feature distance loss, and Respectively and The average characteristics, is the total number of average features, .

[0173] By first distance loss calculation, It can also expand features while maintaining averageness.

[0174] S2323: Calculate the second feature distance loss between each feature and the average feature in the neighborhood of each feature center according to the second distance loss calculation formula; the second distance loss calculation formula is:

[0175]

[0176] in, For each feature in the neighborhood of the feature center, is the average feature in the neighborhood of the feature center.

[0177] By calculating the second feature distance loss, an additional compactness term is introduced to further minimize the conditional entropy and directly consider each feature in the feature space. With its center The distance between.

[0178] S2324: Calculate the feature distance loss according to the first feature distance loss and the second feature distance loss corresponding to the source comprehensive feature distribution and the target comprehensive feature distribution; the calculation formula of the feature distance loss is:

[0179]

[0180] in, is a penalty parameter used to balance the contribution of source features and target features. and is the first feature distance loss and the second feature distance loss of the source comprehensive feature distribution, and It is the first feature distance loss and the second feature distance loss of the target comprehensive feature distribution.

[0181] Final Combination and , and obtain the feature distance loss.

[0182] In step S233, and , and the total distance loss is obtained. The distance loss can be expressed as:

[0183]

[0184] in, and It is used for balance and The penalty parameter of the action, is the overall distance loss, is the feature distance loss.

[0185] In order to facilitate the understanding of the remaining service life prediction method of the bearing, the method is introduced again from the overall perspective. Figure 6 , Figure 6 A schematic diagram of the technical solution of a method for predicting the remaining useful life of a bearing provided in an embodiment of the present invention is provided. The historical operating condition signals of the first bearing and the second bearing are respectively subjected to index calculation to obtain time domain features, frequency domain features and time-frequency domain features. The ODASTN model is used to extract features of each indicator, and the source domain features and the target domain features are aligned using ordered entropy and multi-core maximum mean difference. The source domain features are used to train the prediction regressor, and finally the regressor is used to predict the target domain features to obtain the RUL of the second bearing.

[0186] Among them, feature extraction and alignment refer to Figure 7 , Figure 7 A schematic diagram of feature extraction alignment provided for an embodiment of the present invention, starting from source domain data and target domain data, the raw data is converted into an embedded representation suitable for model processing by input embedding, and then position information and position coding are added. Signal encoding and time encoding are performed using a signal coding layer and a time coding layer to obtain signal features and time features. The signal coding layer and the time coding layer have the same structure, both including multi-head attention, residual connection and layer normalization, feedforward layer and residual connection and layer normalization. After the signal features and time features are fused through the feature fusion layer, the source domain features and the target domain features are obtained, and feature alignment is performed using MK-MMD and ordered entropy.

[0187] Furthermore, the present invention also provides a bearing remaining service life prediction device 300, see Figure 8 , Figure 8 A schematic structural block diagram of a bearing remaining service life prediction device provided in an embodiment of the present invention, the bearing remaining service life prediction device 300 includes:

[0188] The time signal matrix acquisition unit 310 is used to acquire the historical operating condition signal of the first bearing and the historical operating condition signal of the second bearing, and process each historical operating condition signal separately to obtain a time signal matrix respectively; wherein the first bearing and the second bearing are in different operating conditions, the historical operating condition signal of the first bearing includes a remaining service life label, and the time signal matrix includes a time dimension and a signal dimension.

[0189] The feature extraction and fusion unit 320 is used to extract and fuse multi-dimensional features of each time signal matrix to obtain corresponding comprehensive feature distributions.

[0190] The distance loss calculation unit 330 is used to use the comprehensive feature distribution corresponding to the first bearing as the source comprehensive feature distribution, the comprehensive feature distribution corresponding to the second bearing as the target comprehensive feature distribution, align the source comprehensive feature distribution with the target comprehensive feature distribution, and calculate the distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution.

[0191] The distance loss judgment unit 340 is used to judge whether the distance loss meets the preset distance loss; if not, return to execute the step of performing multi-dimensional feature extraction on each time signal matrix and fusing them to calculate the distance loss of the source comprehensive feature distribution and the target comprehensive feature distribution, until the latest distance loss meets the preset distance loss.

[0192] The remaining service life prediction unit 350 is used to train the prediction regressor using the source comprehensive feature distribution to obtain a mature prediction regressor, input the target comprehensive feature distribution into the mature prediction regressor, and obtain the remaining service life of the second bearing.

[0193] In summary, the embodiment of the present invention provides a method, device, electronic device and storage medium for predicting the remaining useful life of a bearing. The historical operating condition signals of the first bearing and the second bearing under different operating conditions are obtained and processed to obtain the corresponding comprehensive feature distributions, which are used as the source comprehensive feature distribution and the target comprehensive feature distribution respectively. Through parallel self-attention encoders, the signal features from the time domain, frequency domain and time-frequency domain are simultaneously integrated in the time dimension and signal dimension to extract more comprehensive signal information. The source comprehensive features are respectively aligned with the target comprehensive feature distribution. The feature alignment method combining MK-MMD and ordered entropy (OE) is used. In combination with MK-MMD and OE regularizers, it is possible to reduce the cross-domain feature differences while retaining the integrity and structure of the features within the domain, thereby improving the cross-domain adaptability of the model. The OE regularizer can expand the distance between features within the domain and weight their distances to maintain the ordinal relationship, while the MK-MMD metric can reduce the differences in cross-domain feature distributions. In view of these two advantages, MK-MMD and OE are combined. When the MK-MMD indicator reduces the difference between cross-domain features, the OE regularizer will maintain a certain distance between features in the domain like classification, and maintain the order relationship between them like regression. While reducing the impact of inter-domain differences on model performance, the life prediction ability of the model is guaranteed. Therefore, the historical operating condition signal of the bearing under one working condition can be marked with RUL only, and the remaining service life of the bearing under other working conditions can be accurately predicted.

[0194] In the embodiments provided by the present invention, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of the code, and a module, a program segment or a part of the code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0195] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0196] If the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a computer-readable storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0197] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0198] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A method for predicting the remaining service life of a bearing, characterized in that: The method comprises: Acquire a historical operating condition signal of the first bearing and a historical operating condition signal of the second bearing, and process each of the historical operating condition signals respectively to obtain a time signal matrix; wherein the first bearing and the second bearing are in different operating conditions, the historical operating condition signal of the first bearing includes a remaining service life label, and the time signal matrix includes a time dimension and a signal dimension; The multi-dimensional features of each time signal matrix are extracted and fused respectively to obtain corresponding comprehensive feature distributions, including: adding position information to each time signal matrix; using multi-head attention to perform signal and time encoding on each time signal matrix after adding the position information to obtain signal features and time features; performing feature fusion on the signal features and the time features to obtain a comprehensive feature distribution; wherein the formula for feature fusion is: in, is the signal characteristic, is the time feature, is the trainable parameter matrix; Taking the comprehensive feature distribution corresponding to the first bearing as the source comprehensive feature distribution, taking the comprehensive feature distribution corresponding to the second bearing as the target comprehensive feature distribution, aligning the source comprehensive feature distribution with the target comprehensive feature distribution, and calculating the distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution; Determining whether the distance loss meets a preset distance loss; If not, return to the step of performing multi-dimensional feature extraction on each of the time signal matrices and fusing them to calculate the distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution, until the latest distance loss meets the preset distance loss; If so, the prediction regressor is trained using the source comprehensive feature distribution to obtain a mature prediction regressor, and the target comprehensive feature distribution is input into the mature prediction regressor to obtain the remaining service life of the second bearing.

2. The method according to claim 1, characterized in that The historical operating condition signal includes signals at multiple time points, and the signal at each time point is a waveform. The historical operating condition signals are processed separately to obtain time signal matrices, including: Performing index calculation on each of the waveforms of each of the historical operating condition signals respectively, and obtaining a plurality of index data corresponding to each of the waveforms respectively; Perform time window processing on each of the historical operating condition signals according to a preset time length to obtain a plurality of time window data; the time window data includes a time step and indicator data; The time window data of each of the historical operating condition signals are assembled in time sequence to obtain time signal matrices.

3. The method according to claim 1, characterized in that The multi-head attention is used to perform signal and time encoding on each of the time signal matrices after adding the position information to obtain signal features and time features, including: The time signal matrix is ​​processed by the following formula to generate a query matrix, a key matrix and a value matrix: in, is the query matrix, is the bond matrix, is the value matrix, is a trainable parameter, ; The attention weight matrix of the query matrix and the key matrix in the time signal matrix is ​​calculated by the attention weight matrix calculation formula; the attention weight matrix calculation formula is: in, , , represents the time dimension of the time signal matrix, Represents the time dimension time step, represents the signal dimension of the time signal matrix, is the scaling factor; According to the attention weight matrix and the value matrix, the head output is calculated by the head output calculation formula; the head output calculation formula is: By using the attention weight matrix calculation formula and the head output calculation formula, multiple head outputs are calculated, and the multiple head outputs are spliced ​​together to obtain signal features and time features; wherein the splicing formula is: in, is the parameter matrix, is the number of header outputs, .

4. The method according to claim 1, characterized in that: The feature alignment of the source comprehensive feature distribution and the target comprehensive feature distribution, and the calculated distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution, includes: Using an alignment formula to shorten the feature distance between the source comprehensive feature distribution and the target comprehensive feature distribution, and calculating the overall distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution; Calculating the feature distance loss of each feature of the source comprehensive feature distribution and each feature of the target comprehensive feature distribution; The distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution is calculated based on the overall distance loss and the feature distance loss.

5. The method according to claim 4, characterized in that The alignment formula is as follows: in, is the overall distance loss, are the source comprehensive feature distribution and the target comprehensive feature distribution respectively, and Respectively represent the sample numbers of the source comprehensive feature distribution and the target comprehensive feature distribution, is the feature kernel, characterizes the internal alignment of the source composite feature distribution, characterizing the alignment of the source comprehensive feature distribution with the target comprehensive feature distribution, Characterize the internal alignment of the target comprehensive feature distribution.

6. The method according to claim 4, characterized in that The calculating the feature distance loss of each feature of the source comprehensive feature distribution and each feature of the target comprehensive feature distribution includes: For the source comprehensive feature distribution and the target comprehensive feature distribution, each feature is taken as the feature center, and the average feature of all features in the neighborhood of each feature center is calculated; the calculation formula is: in, is the average feature, Before and after the feature Neighborhood features, for neighbor characteristics; According to each of the average features, the first feature distance loss of each of the average features is calculated by a first distance loss calculation formula; the first distance loss calculation formula is: in, is the first feature distance loss, and Respectively and The average characteristics, is the total number of the average features, ; According to the second distance loss calculation formula, the second feature distance loss between each feature in the neighborhood of each feature center and the average feature is calculated; the second distance loss calculation formula is: in, For each feature in the neighborhood of the feature center, is the average feature in the neighborhood of the feature center; The characteristic distance loss is calculated according to the first characteristic distance loss and the second characteristic distance loss corresponding to the source comprehensive characteristic distribution and the target comprehensive characteristic distribution; the calculation formula of the characteristic distance loss is: in, is a penalty parameter used to balance the contribution of source features and target features. and is the first characteristic distance loss and the second characteristic distance loss of the source comprehensive characteristic distribution, and A first feature distance loss and a second feature distance loss are obtained for the target comprehensive feature distribution.

7. A bearing remaining service life prediction device, characterized in that: The device comprises: A time signal matrix acquisition unit is used to acquire a historical operating condition signal of the first bearing and a historical operating condition signal of the second bearing, and process each of the historical operating condition signals respectively to obtain a time signal matrix respectively; wherein the first bearing and the second bearing are in different operating conditions, the historical operating condition signal of the first bearing includes a remaining service life label, and the time signal matrix includes a time dimension and a signal dimension; The feature extraction and fusion unit is used to extract and fuse multi-dimensional features of each of the time signal matrices to obtain corresponding comprehensive feature distributions, including: adding position information to each of the time signal matrices; using multi-head attention to perform signal and time encoding on each of the time signal matrices after adding the position information to obtain signal features and time features; performing feature fusion on the signal features and the time features to obtain a comprehensive feature distribution; wherein the formula for feature fusion is: in, is the signal characteristic, is the time feature, is the trainable parameter matrix; a distance loss calculation unit, configured to use the comprehensive feature distribution corresponding to the first bearing as a source comprehensive feature distribution, use the comprehensive feature distribution corresponding to the second bearing as a target comprehensive feature distribution, perform feature alignment on the source comprehensive feature distribution and the target comprehensive feature distribution, and calculate the distance loss between the source comprehensive feature distribution and the target comprehensive feature distribution; A distance loss judgment unit is used to judge whether the distance loss meets the preset distance loss; if not, return to the step of performing multi-dimensional feature extraction on each of the time signal matrices and fusing them to calculate the distance loss of the source comprehensive feature distribution and the target comprehensive feature distribution, until the latest distance loss meets the preset distance loss; The remaining service life prediction unit is used to, if yes, use the source comprehensive feature distribution to train the prediction regressor to obtain a mature prediction regressor, input the target comprehensive feature distribution into the mature prediction regressor, and obtain the remaining service life of the second bearing.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images