Intelligent prediction method for residual life of multi-stage degradation equipment

Abstract

The invention discloses an intelligent prediction method for the residual life of a multi-stage degradation device, and belongs to the technical field of reliability engineering, and the method comprises the steps: building and training a stage division model of the multi-stage degradation device based on a continuous deep belief network, and then generating a sample with a degradation label through a sliding time window; and a Bayesian bidirectional GRU is used to describe the performance evolution model of each stage. On the basis, the current multi-domain features of the in-service equipment are input into the stage division model to calculate the probability of each stage, and then the degradation features are extrapolated through the prediction model of each stage. Thirdly, a predicted degradation profile may be determined according to a total probability formula, similarly a degradation level at any moment is calculated iteratively, and a remaining life profile is determined in conjunction with a failure threshold. According to the method, a zero-life label is used as a typical feature, a degradation stage division and residual life prediction method is provided under the background of big data, and optimization of the number of degradation stages and improvement of residual life prediction precision are ensured.

Description

Technical Field

[0001] The invention belongs to the technical field of reliability engineering, and in particular relates to an intelligent prediction method for the remaining life of multi-stage degraded equipment. Background Art

[0002] Different from the failure mechanism method, the data-driven remaining life prediction method mainly uses mathematical tools to fit the monitoring data, which has been widely developed and generally favored in the field of remaining life prediction. At present, in the era of big data, the monitoring data collected by degraded equipment during operation is rapidly expanding and showing massive characteristics, which puts higher requirements on calculation. In this context, deep learning-based methods have received increasing attention in recent years. Representative models include autoencoders, deep belief networks, deep convolutional neural networks, deep recurrent neural networks, and their variants and combinations. Most studies focus on degraded equipment in a single stage. In other words, the same degradation mode should be met throughout the life cycle, and usually only a specific deep learning model is used for training.

[0003] Unfortunately, due to the inherent mechanism of component degradation, the external influence of the operating environment, and the coupling effect between components, the performance degradation process of some equipment is difficult to satisfy the degradation law of a single mode within the entire life profile. Therefore, multi-stage degradation characteristics may be often encountered in practice. Taking rolling bearings as an example, it can be inferred from engineering experience that the operation process is usually manifested in three typical stages, namely, the stable operation stage, the slow degradation stage, and the severe degradation stage. Obviously, if the existing method of a single model is used, the equipment cannot show the same degradation mode in different stages, so the prediction error of multi-stage degradation equipment is inevitable. The overall idea of ​​the remaining life prediction of multi-stage degradation equipment in the big data era is usually divided into the following two steps: degradation stage division and multi-stage prediction of remaining life. In particular, when the number of degradation stages is equal to 2, the degradation stage division can be converted into anomaly detection, and the degradation data after the anomaly point can be used to predict the remaining life, and many theoretical results have been achieved. However, by analyzing these studies, it is found that the field of remaining life prediction of multi-stage degradation equipment under big data will inevitably encounter three problems. The first problem is the unknown number of degradation stages, which often depends on empirical knowledge. Once the number of degradation stages is set inaccurately, the accuracy of subsequent remaining life prediction will be directly affected. The second problem is that these models rely heavily on life labels, which means that it is necessary to obtain life information in advance. Therefore, once truncated data are encountered in engineering practice, the existing prediction models with life as labels cannot be directly applied due to the lack of life labels. Although the prediction models with degradation labels can handle truncated data, there may be few related studies on the uncertainty quantification of remaining life under such models. The third problem is that the prediction results may be obtained based on a specific degradation stage, while ignoring the uncertainty of the results of the degradation stage. Taken together, these problems will inevitably lead to poor prediction results and limit their application in maintenance decisions. Therefore, the problem of uncertainty quantification in the remaining life prediction of multi-stage degraded equipment under zero life labels is still blank, and further research is urgently needed. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides an intelligent prediction method for the remaining service life of multi-stage degraded equipment.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0006] A method for intelligently predicting the remaining life of multi-stage degraded equipment comprises the following steps:

[0007] S1. Extract multi-domain features from the truncated monitoring data of historical multi-stage degradation equipment to characterize performance degradation, build a degradation stage division model and train it;

[0008] S2, based on the model trained in S1, build prediction models for different degradation stages and predict the remaining life of the equipment;

[0009] S3. The prediction results of the remaining life of the equipment at multiple stages are integrated to obtain the final predicted remaining life of the equipment.

[0010] Furthermore, the S1 specifically includes the following steps:

[0011] S11. Use continuous deep belief network to build degradation stage division model and optimize the number of degradation stages;

[0012] S12, add Gaussian noise units to train the constructed continuous deep belief network layer by layer, update the weights and the parameters controlling the slope of the sigmoid function through pre-training, introduce the softmax regression layer for supervised training during the last layer training, and adjust the parameters through the BP loss function;

[0013] S13. Record the current loss function and determine whether the current degradation stage reaches a predefined maximum number. When the degradation stage reaches the predefined maximum number, the training loop will end. By comparing the loss functions of different stage numbers, the optimal stage number with the smallest loss function is determined.

[0014] Furthermore, the specific method of updating the weights and controlling the parameters of the slope of the sigmoid function by pre-training in S12 is:

[0015]

[0016] Where W ij is the weight, c j is the parameter that controls the slope of the sigmoid function, Δ represents the parameter change, <·> represents the mathematical expectation of the training data, and ε w and ε c W ij and c j The learning rate, and Once means s i and j One-step reconstruction state, s i s j is the product of the current states of neuron i and neuron j, It is the product of the one-step reconstruction states of neuron i and neuron j.

[0017] Furthermore, the BP loss function in S12 is expressed as:

[0018]

[0019] Where Θ is the set of all parameters in the continuous deep belief network, Q and n are the number of training samples and the number of degraded labels respectively; I ic is the indicator function. When the degradation stage result of the constructed stage division model is consistent with the actual degradation label, I ic is equal to 1, otherwise I ic =0; p ic is the probability of the input sample in the degradation stage c.

[0020] Furthermore, the S2 specifically includes the following steps:

[0021] S21, using Bayesian bidirectional GRU to build feature prediction models at different degradation stages;

[0022] S22, extract representative features from the monitoring time as input time series, input the obtained input time series into the feature prediction model obtained in S21 from both forward and reverse directions, and obtain the output states of the last forward layer and the last reverse layer respectively;

[0023] S23, connecting the output states of the last forward layer and the last reverse layer obtained in S22, and inputting the connected vector into a fully connected layer to obtain the prediction feature.

[0024] Furthermore, the specific method of extracting representative features from the monitoring time as the input time series in S22 and inputting the obtained input time series into the feature prediction model obtained in S21 from both positive and negative directions is as follows:

[0025]

[0026] In the formula, σ(·) represents the sigmoid function, → and ← represent the information transfer of the forward and backward processes respectively, L is the index of the forward layer and the backward layer, r t Represented by weight W rh , W rx and bias b r is the reset gate of the parameters, W rh With W rx Respectively represent the weights between the reset gate and the state, and between the reset gate and the input, Indicates time t k The new state of the storage unit is and bias and Respectively represent the weights between the new state and the input, the new state and the reset gate, z t Represented by weight W zh , W zx and bias b z is the parameter update gate, W zh With W zxRepresent the weights between the update gate and the state, and between the update gate and the input respectively.

[0027] Furthermore, the prediction feature in S23 is expressed as:

[0028]

[0029] In the formula, is the prediction feature, W hy is the weight of the output layer, b y is the bias of the output layer, and h is the vector of the concatenated output states of the last forward layer and the last reverse layer, expressed as:

[0030]

[0031] Where P0 is the time step of each training sample of the bidirectional GRU, is the output state of the last forward layer, The output state of the final reverse layer, t is the time.

[0032] Furthermore, the S3 specifically includes the following steps:

[0033] S31, inputting the multi-domain features of the stage-degraded equipment at the current time into the constructed prediction model of different degradation stages to obtain the probability of each degradation stage, and calculating the feature prediction distribution of multiple steps in the future;

[0034] S32, averaging the predicted distribution of the features for multiple steps in the future obtained in S31, and obtaining the predicted results of the future features of the in-service multi-stage degraded equipment;

[0035] S33. Integrate the multi-stage prediction results to obtain feature prediction distribution and point estimation, and use the performance feature distribution and fault threshold at the corresponding moment to calculate the failure probability of the equipment at the future moment.

[0036] Furthermore, the feature prediction distribution of the future multiple steps in S31 is expressed as:

[0037]

[0038] In the formula, z d+L represents the feature level of the next L steps, p(z d+L ) is the feature prediction distribution for the next L steps, p c is the probability of the cth degradation stage and c∈[1,n], n is the number of stages; p(z d+L |c) is the predicted distribution of the features in the future L steps under the cth degradation stage, n * The best stage number with the minimum loss function.

[0039] Furthermore, the specific method of calculating the failure probability of the device at a future time by using the performance characteristic distribution and the failure threshold at the corresponding time in S33 is:

[0040]

[0041] In the formula, P(t k ) for the future k The failure probability of the equipment at time z k Indicates t k The characteristic level of the equipment at the moment, w represents the failure threshold of the characteristic level.

[0042] The present invention has the following beneficial effects:

[0043] The model of the present invention not only optimizes the optimal number of degradation stages, but also quantifies multiple uncertainties including stage categories and performance degradation when facing a zero-life label, effectively ensuring prediction accuracy and providing accurate information support for subsequent maintenance arrangements and spare parts ordering. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic flow chart of the intelligent prediction method for the remaining life of multi-stage degraded equipment of the present invention.

[0045] Figure 2 It is the loss function of the continuous deep belief network under different degradation stage numbers in the embodiment of the present invention.

[0046] Figure 3a Schematic diagram of the degradation stage division result of the bearing 1_1 according to the embodiment of the present invention.

[0047] Figure 3b Schematic diagram of the degradation stage division results of the bearing 3_3 according to the embodiment of the present invention.

[0048] Figure 4a : is the probability of the bearing 1_1 in different stages at four prediction moments according to the embodiment of the present invention.

[0049] Figure 4b It is the probability of the bearing 3_3 in different stages at four prediction moments according to the embodiment of the present invention.

[0050] Figure 5 Schematic diagram of the CDF of the remaining life of the bearing 1_1 according to the embodiment of the present invention.

[0051] Figure 6 FIG. 1 is a PDF diagram of the remaining life of the bearing 1_1 according to an embodiment of the present invention.

[0052] Figure 7 Schematic diagram of the remaining life CDF of the bearing 3_3 according to the embodiment of the present invention.

[0053] Figure 8 Schematic diagram of the remaining life PDF of the bearing 3_3 according to the embodiment of the present invention. DETAILED DESCRIPTION

[0054] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0055] An intelligent prediction method for the remaining life of multi-stage degraded equipment, such as Figure 1 As shown, the following steps are included:

[0056] S1. Extract multi-domain features from the truncated monitoring data of historical multi-stage degradation equipment to characterize performance degradation, build a degradation stage division model and train it;

[0057] In this embodiment, the following steps are specifically included:

[0058] S11. Use continuous deep belief network to build degradation stage division model and optimize the number of degradation stages;

[0059] Multi-domain features are extracted from the truncated monitoring data of the historical multi-stage degradation equipment to characterize the performance degradation. Considering that the division stage is unknown, it can be initialized to 2, and the incomplete data of the historical multi-stage degradation equipment can be divided into two stages with the middle position as the division point. Therefore, the degradation stage label can be further obtained based on the original division. Subsequently, the multi-domain features including time domain, frequency domain and time-frequency domain are extracted from the monitoring data using the existing feature extraction theory. For ease of explanation, the multi-domain features and the associated degradation stage labels are represented as It can be regarded as the training basis of the degradation stage partition model.

[0060] S12, add Gaussian noise units to train the constructed continuous deep belief network layer by layer, update the weights and the parameters controlling the slope of the sigmoid function through pre-training, introduce the softmax regression layer for supervised training during the last layer training, and adjust the parameters through the BP loss function;

[0061] Since the basic element of the deep belief network is the binary restricted Boltzmann machine, there may be large reconstruction errors when modeling continuous data, which can be regarded as a defect of the deep belief network in representing the continuous monitoring data of degraded equipment. In order to solve this defect, a continuous deep belief network stacked by continuous restricted Boltzmann machines is proposed, which has more advantages than the deep belief network when processing continuous monitoring data by adding Gaussian noise units. In the training process of the continuous deep belief network, there are two steps: pre-training and back-adjustment. The first step is to calculate the initial values ​​of weights and biases based on the minimization of contrastive divergence algorithm. This pre-training is implemented layer by layer, indicating that the output of the continuous restricted Boltzmann machine of the previous layer is regarded as the input of the continuous restricted Boltzmann machine of the next layer. For a continuous restricted Boltzmann machine, the states of the visible layer and the hidden layer can be expressed as and where i and j represent the neuron indices of the visible and hidden layers, respectively. The key unknown parameters include the weights W ij and the parameter c that controls the slope of the sigmoid function j , updated as follows:

[0062]

[0063]

[0064] Where Δ represents the parameter change and <·> represents the mathematical expectation of the training data. w and ε c Respectively represent W ij and c j The learning rate is and Respectively represent s i and j One-step reconstruction state. i s j is the product of the current states of neuron i and neuron j, is the product of the one-step reconstruction states of neuron i and neuron j. Subsequently, the model parameters can be updated according to equations (1) and (2).

[0065] Assume that the number of continuous restricted Boltzmann machines in the continuous deep belief network is R, and each continuous restricted Boltzmann machine is trained through the above mechanism. After the last continuous restricted Boltzmann machine is trained, it is necessary to introduce a softmax regression layer for supervised training. In other words, the second step of the continuous deep belief network is to use multiple domain features and associated degradation stage labels to adjust parameters through the BP algorithm. The loss function can be written as:

[0066]

[0067] Where Θ represents all parameter sets in the continuous deep belief network, Q and n represent the number of training samples and the number of degraded labels. ic is the indicator function. When the degradation stage result of the constructed stage division model is consistent with the actual degradation label, I ic is equal to 1. On the contrary, once the result is inconsistent with the actual degradation label, I ic is equal to 0. ic is the probability of the input sample in the degradation stage c, which can be calculated as:

[0068]

[0069] Among them, r ic Represents the output of neuron c in the last continuous restricted Boltzmann machine for input sample i. The present invention adopts the Adam optimization method to minimize the above loss function to adjust the parameters

[0070] S13. Record the current loss function and determine whether the current degradation stage reaches a predefined maximum number. When the degradation stage reaches the predefined maximum number, the training loop will end. By comparing the loss functions of different stage numbers, the optimal stage number with the smallest loss function is determined.

[0071] The degradation stage partitioning model based on the continuous deep belief network can be trained by implementing the above process, recording the loss function and associating the stage number. Based on the stage number, the middle position of each stage is selected as a candidate partitioning point. In each stage, the trend factors of the data before and after the middle position can be monitored separately, and then the difference between the two trend factors can be obtained. The stage with the largest difference is selected as the further partitioning stage, so the degradation stage number can become n+1. In addition, the above process can be re-executed to generate multiple domain features and associate degradation stage labels, and train the degradation stage partitioning model. Next, record the current loss function and determine whether the current degradation stage reaches the predefined maximum number. When the degradation stage reaches the predefined maximum number, the training loop will end. By comparing the loss functions of different stage numbers, the optimal stage number with the smallest loss function is determined: n * =argminl(n).

[0072] S2, based on the model trained in S1, build prediction models for different degradation stages and predict the remaining life of the equipment;

[0073] In this embodiment, the following steps are specifically included:

[0074] S21, using Bayesian bidirectional GRU to build feature prediction models at different degradation stages;

[0075] Since the degradation mechanisms and evolution laws at different stages are different, it is necessary to construct prediction models for different degradation stages. To this end, representative features should be extracted from the monitoring time in advance, which can be expressed as: Z = [Z1, Z2, …, Z M ], where Z i is the representative feature sequence of the i-th multi-stage degraded device.

[0076] S22, extract representative features from the monitoring time as input time series, input the obtained input time series into the feature prediction model obtained in S21 from both forward and reverse directions, and obtain the output states of the last forward layer and the last reverse layer respectively;

[0077] Taking bearings as an example, the root mean square should be selected as a representative feature. For specific degradation stages, the sliding time window technique is introduced to generate samples with feature labels, indicating that the goal of this training process is to describe the relationship between time series. As a typical variant of recurrent neural networks, GRU has been widely used in time series prediction in various fields. In general, this variant can be mathematically expressed as follows:

[0078]

[0079] Among them, r t Represented by weight W rh , W rx and bias b r is the reset gate for the parameters. Indicates time t k The new state of the storage unit is and bias z t Represented by weight W zh , W zx and bias b z is the update gate of the parameters. The traditional structure of GRU is difficult to characterize the relationship between the current representative features and the future representative features. Therefore, the bidirectional GRU is introduced to facilitate the processing of input time series in two directions, and to mine the forward and backward time dependencies. Its mathematical formula is expressed as:

[0080]

[0081] Where σ(·) represents the sigmoid function, → and ← represent the information transfer of the forward and backward processes respectively, and L is the index of the forward layer and the backward layer. The output states of the last forward layer and the last backward layer are defined as and

[0082] S23, connecting the output states of the last forward layer and the last reverse layer obtained in S22, and inputting the connected vector into a fully connected layer to obtain the prediction feature.

[0083] These two output states can then be connected as Here, P0 represents the time step of each training sample of the bidirectional GRU. The concatenated vector is further input into a fully connected layer to obtain the prediction feature, which can be calculated as:

[0084]

[0085] Among them, W hy and b y denote the weight and bias of the output layer respectively. Assume that the training data set of the cth degradation stage can be represented as input and the corresponding output here, and Represent the input and output of the mth training sample, Q c is the number of training samples in the cth degradation stage. d and L represent the time window length and prediction step length, respectively. In order to fully consider the uncertainty quantification of the prediction results, the idea of ​​Bayesian neural network is adopted, and the parameters of the bidirectional GRU are regarded as random variables, namely Bayesian bidirectional GRU. In addition, the variational inference technique is applied to approximately solve the distribution of the prognosis results. Therefore, the performance evolution model of each degradation stage can be constructed and the uncertainty of the representative features can be characterized.

[0086] S3. The prediction results of the remaining life of the equipment at multiple stages are integrated to obtain the final predicted remaining life of the equipment.

[0087] This embodiment specifically includes the following steps:

[0088] S31, inputting the multi-domain features of the stage-degraded equipment at the current time into the constructed prediction model of different degradation stages to obtain the probability of each degradation stage, and calculating the feature prediction distribution of multiple steps in the future;

[0089] Once the degradation stage division model and the performance prediction model of different degradation stages are trained, the remaining life prediction of multi-stage degradation equipment can be performed on the field. Specifically, the multi-domain features of the multi-stage degradation equipment at the current time are input into the constructed stage division model to obtain the probability of each degradation stage, that is, {p1,…,p n Subsequently, the samples obtained by the sliding time window technology at the current time of the multi-stage degradation equipment on site are input into the prediction models of different degradation stages respectively. The prediction distribution of the future L-step features under the corresponding degradation stage can be expressed as p(z d+L|c). According to the total probability formula, the feature prediction distribution of the next L steps can be further expressed as:

[0090]

[0091] S32, averaging the predicted distribution of the features for multiple steps in the future obtained in S31, and obtaining the predicted results of the future features of the in-service multi-stage degraded equipment;

[0092] After obtaining the prediction distribution in the above formula, the average value can be regarded as the prediction result of the future characteristics of the in-service multi-stage degradation equipment. Therefore, a new test sample with a specific time window can be generated, which can be input into the prediction model of different degradation stages respectively for the subsequent L-step feature prediction results. Similarly, the feature prediction distribution and point estimate can be obtained by fusing the multi-stage prediction results. This also shows that the above process can be implemented alternately by prediction supplementation. Therefore, this method can determine the distribution of performance characteristics at any time in the future.

[0093] S33. Integrate the multi-stage prediction results to obtain feature prediction distribution and point estimation, and use the performance feature distribution and fault threshold at the corresponding moment to calculate the failure probability of the equipment at the future moment.

[0094] In order to achieve the remaining life prediction, it is necessary to first define the concept of failure. Generally, when the representative characteristics of the multi-stage degradation equipment exceed the failure threshold in engineering, it can be considered that the multi-stage degradation equipment has failed. According to reliability theory, the failure probability at a future moment can be expressed by an expression associated with the performance characteristic distribution and the failure threshold at the corresponding moment, that is,

[0095]

[0096] Then, the characteristic uncertainty can be converted into the remaining life uncertainty with the help of the failure rate. That is, the probability in the remaining life CDF will be regarded as the failure probability at this time. Finally, the PDF of the remaining life can be further calculated based on the remaining life CDF, thereby quantifying the remaining life prediction uncertainty of multi-stage degradation equipment and providing important information support for the formulation of scientific health management strategies.

[0097] Verification

[0098] The present invention uses the experimental data of bearings from the 2012 PHM Challenge to verify the effectiveness of the proposed method. The experimental platform has three key parts: the rotating part, the loading part and the measuring part. The rotating part includes an asynchronous motor with a gearbox and two shafts, which is specially used to generate and transmit rotational motion. The loading part is designed to generate radial forces applied to the test rolling bearing, while the measuring part consisting of two accelerometers and a temperature sensor is designed to measure the operating conditions. The present invention uses the vibration signal of the accelerometer on the horizontal axis to verify the proposed method.

[0099] Degradation experiments were conducted under three working conditions, in which the sampling frequency of the accelerometer was 25.6kHz and the sampling duration of each sampling was 0.1s. This shows that there are 2560 data points for a specific sampling. In addition, each sampling interval is 10s. In order to implement the proposed method, the training set and the test set should be classified in advance. Two datasets involving working condition 1 and working condition 3 are selected to form the test set. The remaining datasets under these two working conditions are regarded as training datasets for training the degradation stage division model and the prognosis model. Table 1 lists the specific bearing numbers of the training set and the test set. Although several datasets are run-to-failure, the life information is not utilized in the entire process of the proposed method to meet the actual situation of zero life labels.

[0100] Table 1 Bearing numbers of training set and test set

[0101] Training set Bearing 1_2 Bearing 1_3 Bearing 1_4 Bearing 1_5 Bearing 1_6 Bearing 1_7 Bearing 3_1 Bearing 3_2 Test Set Bearing 1_1 Bearing 3_3

[0102] Then, multi-domain features are extracted from the bearing vibration data to train the degradation stage division model based on the continuous deep belief network. Here, 16 time domain features and 13 frequency domain features are selected as the extracted features, such as mean, root mean square, center frequency, root mean square frequency, etc. Since the extracted multi-domain features are input into the continuous deep belief network, the dimension of the input layer is 29. The size of the output layer is equal to the number of degradation stages. In addition, through cross-validation, the size of the hidden layer is selected as 20, 14, and 10. The specific structure and initial parameters of the continuous deep belief network are shown in Table 2.

[0103] Table 2 Specific structure and initial parameters of the continuous deep belief network

[0104]

[0105] The initial number of degradation stages is set to 2, and the extracted features of the training set can be divided into two stages at the middle position. Therefore, the continuous deep belief network model can be trained by pre-training and reverse adjustment. Correspondingly, the loss function can be recorded together with the corresponding number of stages and division status. Depending on the way the number of degradation stages is increased, different division situations can be obtained from 2 to a predefined maximum value. In the experiment, the predefined maximum value is set to 8. For the division features of the training set, the degradation stage division model based on the continuous deep belief network should be retrained in a similar way. Therefore, the continuous deep belief network loss function under different numbers of degradation stages can be determined, such as Figure 2 shown.

[0106] It can be observed that when the degradation stage is equal to 4, the loss function is the smallest, which means the classification accuracy is the highest. Therefore, it is more scientific and reasonable to consider four degradation stages in the stage division.

[0107] In addition, the multi-domain features of the test set are input into the trained continuous deep belief network to determine the stage division results at different times with the minimum loss function. In order to facilitate stage display and subsequent prediction, the present invention provides degradation stage division for the test samples in the form of root mean square. Figure 3a and Figure 3b As shown, it can be seen that the constructed stage division model can maintain the correct stage results for most test samples. However, for some test samples, the obtained stage results may be inconsistent with the stage labels. By statistics, it can be found that the number of inconsistent degradation stage division results for bearing 1_1 and bearing 3_3 is 42 and 15 respectively. Therefore, the accuracy of the stage division model of these two bearings is 98.57% and 97.7%, which verifies the correctness of the proposed degradation stage division model. In addition, considering internal and external factors, there may be a certain degree of uncertainty in the stage attribution of each time period. According to the constructed softmax regression layer, the corresponding probabilities of the four degradation stages can be calculated. Four prediction moments are selected as representatives for each bearing to demonstrate this uncertainty, namely the 1700th cycle, 2000th cycle, 2300th cycle and 2600th cycle of bearing 1_1, and the 250th, 300th, 350th and 400th cycles of bearing 3_3. Figure 4a and Figure 4b The probabilities of the two bearings at different stages at the four prediction moments are described. Therefore, the stage categories and the corresponding uncertainties can be identified through the constructed partitioning model.

[0108] After obtaining the degradation stage division results of the training set, prediction models based on different stages of the Bayesian bidirectional GRU will be constructed and trained to describe different evolution laws. To achieve this goal, the root mean square of the bearing is extracted as a representative feature and the remaining life prediction is performed in advance. Based on the root mean square value of a specific degradation stage, the training samples of the Bayesian bidirectional GRU can be obtained by the sliding time window technology. The corresponding parameters, including the step size and prediction step size, can be set to 10 and 1. The specific structure is 50-50, and the learning rate is 0.005. Once the prognostic model of the Bayesian bidirectional GRU is well trained, the uncertainty of the root mean square of bearing 1_1 and bearing 3_3 at any time in the future can be characterized, and the remaining life distribution including CDF and PDF can be obtained. Figure 5-Figure 8 The CDF and PDF of the remaining life at four prediction moments are shown respectively.

[0109] It is obvious that the CDF of the remaining life of bearing 1_1 and bearing 3_3 at the four prediction moments increases monotonically from 0 to 1. When the CDF of the remaining life is equal to 0.5, the median of the remaining life can be obtained. Based on this idea, the median of the remaining life of bearing 1_1 at these four prediction moments can be approximately obtained as 1104 cycles, 801 cycles, 504 cycles and 203 cycles, while the median of the remaining life of bearing 3_3 at these four prediction moments can be approximately obtained as 183 cycles, 135 cycles, 84 cycles and 35 cycles. In addition, it can be observed that the range of the CDF curve of the remaining life at these four prediction moments can cover the true remaining life, and all prediction results are close to the true remaining service life. Therefore, the method proposed in the present invention can ensure the correctness of the remaining life prediction, which is mainly because the proposed method can consider the different evolution processes of different degradation stages in the future.

[0110] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0111] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0112] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0113] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0114] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for intelligently predicting the remaining life of multi-stage degraded equipment, characterized in that: The steps include: S1. Extract multi-domain features from the truncated monitoring data of historical multi-stage degradation equipment to characterize performance degradation, build a degradation stage division model and train it; S2, based on the model trained in S1, build prediction models for different degradation stages and predict the remaining life of the equipment; S3. The prediction results of the remaining life of the equipment at multiple stages are integrated to obtain the final predicted remaining life of the equipment.

2. According to claim 1, a method for intelligently predicting the remaining life of multi-stage degraded equipment is characterized in that: The S1 specifically includes the following steps: S11. Use continuous deep belief network to build degradation stage division model and optimize the number of degradation stages; S12, add Gaussian noise units to train the constructed continuous deep belief network layer by layer, update the weights and the parameters controlling the slope of the sigmoid function through pre-training, introduce the softmax regression layer for supervised training during the last layer training, and adjust the parameters through the BP loss function; S13. Record the current loss function and determine whether the current degradation stage reaches a predefined maximum number. When the degradation stage reaches the predefined maximum number, the training loop will end. By comparing the loss functions of different stage numbers, the optimal stage number with the smallest loss function is determined.

3. The method for intelligently predicting the remaining life of multi-stage degraded equipment according to claim 2 is characterized in that: The specific method of updating the weights and the parameters controlling the slope of the sigmoid function by pre-training in S12 is: Where W ij is the weight, c j is the parameter that controls the slope of the sigmoid function, Δ represents the parameter change, <·> represents the mathematical expectation of the training data, and ε w and ε c W ij and c j The learning rate, and Once means s i and j One-step reconstruction state, s i s j is the product of the current states of neuron i and neuron j, It is the product of the one-step reconstruction states of neuron i and neuron j.

4. The method for intelligently predicting the remaining life of multi-stage degraded equipment according to claim 2 is characterized in that: The BP loss function in S12 is expressed as: Where Θ is the set of all parameters in the continuous deep belief network, Q and n are the number of training samples and the number of degraded labels respectively; I ic is the indicator function. When the degradation stage result of the constructed stage division model is consistent with the actual degradation label, I ic is equal to 1, otherwise I ic =0; pi c is the probability of the input sample in the degradation stage c.

5. The method for intelligently predicting the remaining life of multi-stage degraded equipment according to claim 1 is characterized in that: The S2 specifically includes the following steps: S21, using Bayesian bidirectional GRU to build feature prediction models at different degradation stages; S22, extract representative features from the monitoring time as input time series, input the obtained input time series into the feature prediction model obtained in S21 from both forward and reverse directions, and obtain the output states of the last forward layer and the last reverse layer respectively; S23, connecting the output states of the last forward layer and the last reverse layer obtained in S22, and inputting the connected vector into a fully connected layer to obtain the prediction feature.

6. The method for intelligently predicting the remaining life of multi-stage degraded equipment according to claim 5 is characterized in that: In S22, representative features are extracted from the monitoring time as input time series, and the input time series is input into the feature prediction model obtained in S21 from both positive and negative directions as follows: In the formula, σ(·) represents the sigmoid function, → and ← represent the information transfer of the forward and backward processes respectively, L is the index of the forward layer and the backward layer, r t Represented by weight W rh , W rx and bias b r is the reset gate of the parameters, W rh With W rx Respectively represent the weights between the reset gate and the state, and between the reset gate and the input, Indicates time t k The new state of the storage unit is and bias and Respectively represent the weights between the new state and the input, the new state and the reset gate, z t Represented by weight W zh , W zx and bias b z is the parameter update gate, W zh With W zx Represent the weights between the update gate and the state, and between the update gate and the input respectively.

7. The method for intelligently predicting the remaining life of multi-stage degraded equipment according to claim 5, characterized in that: The prediction feature in S23 is expressed as: In the formula, is the prediction feature, W hy is the weight of the output layer, b y is the bias of the output layer, and h is the vector of the concatenated output states of the last forward layer and the last reverse layer, expressed as: Where PP0 is the time step of each training sample of bidirectional GRU, is the output state of the last forward layer, The output state of the final reverse layer, t is the time.

8. The method for intelligently predicting the remaining life of multi-stage degraded equipment according to claim 1 is characterized in that: The S3 specifically includes the following steps: S31, inputting the multi-domain features of the stage-degraded equipment at the current time into the constructed prediction model of different degradation stages to obtain the probability of each degradation stage, and calculating the feature prediction distribution of multiple steps in the future; S32, averaging the predicted distribution of the features for multiple steps in the future obtained in S31, and obtaining the predicted results of the future features of the in-service multi-stage degraded equipment; S33. Integrate the multi-stage prediction results to obtain feature prediction distribution and point estimation, and use the performance feature distribution and fault threshold at the corresponding moment to calculate the failure probability of the equipment at the future moment.

9. The method for intelligently predicting the remaining life of multi-stage degraded equipment according to claim 8, characterized in that: The feature prediction distribution of the future multiple steps in S31 is expressed as: In the formula, z d+L represents the feature level of the next L steps, p(z d+L ) is the feature prediction distribution for the next L steps, p c is the probability of the cth degradation stage and c∈[1,n], n is the number of stages; p(z d+L |c) is the predicted distribution of features for the next L steps under the cth degradation stage condition, and the best stage number with the minimum n*loss function.

10. The intelligent prediction method for the remaining life of multi-stage degraded equipment according to claim 8, characterized in that: The specific method of calculating the failure probability of the equipment at the future time by using the performance characteristic distribution and the failure threshold at the corresponding time in S33 is: In the formula, P(t k ) for the future k The failure probability of the equipment at time z k Indicates t k The characteristic level of the equipment at the moment, w represents the failure threshold of the characteristic level.

Images