Remaining useful life prediction method for rotating machinery based on long-term evolution learning

Abstract

The invention relates to the technical field of life prediction of rotating machinery, and specifically discloses a method for predicting the remaining service life of rotating machinery based on long-term evolutionary learning, comprising: entering the initial learning stage, constructing a prediction model based on a recurrent convolutional neural network using adaptive weight connections, obtaining a vibration signal data set of the current stage, training the prediction model, measuring the stability of the weight connections after the training is completed, and selecting representative samples of the current stage according to the informativeness and diversity of the samples; entering the evolutionary learning stage, obtaining the vibration signal data set of the current stage and combining it with the representative samples, and training the prediction model trained in the previous stage, in the current stage training, the update rate of the adaptive weight connection changes with the stability of the weight connection of the previous stage; after the current stage training is completed, measuring the stability of the weight connection, and selecting representative samples of the current stage from the vibration signal data set of the current stage according to the informativeness, diversity and novelty of the samples; performing multiple evolutionary learning on the prediction model until the training target is met. The prediction model of the invention can retain the acquired old knowledge for a long time and flexibly learn new knowledge, thereby realizing reliable and efficient remaining service life prediction for rotating machinery.

Description

Technical Field

[0001] The present invention relates to the technical field of rotating machinery life prediction, and more specifically to a method for predicting the remaining service life of rotating machinery based on long-term evolution learning. Background Art

[0002] As a key component of the mechanical system, the flexibility and efficiency of rotating parts not only enable the mechanical system to operate smoothly, but also provide a solid foundation for the realization of complex tasks. Predicting the remaining service life of mechanical rotating parts and implementing accurate preventive maintenance are of great significance to ensuring the efficiency and safety of equipment in long-term operation. In the era of "mechanical big data", data-driven intelligent life prediction methods are widely used due to their excellent prediction capabilities.

[0003] Since the data collection cycle from operation to failure of machinery is long and the failure modes and data distribution are diverse, traditional intelligent prediction methods are difficult to adapt to the dynamic changes of training data. Therefore, evolutionary learning technology came into being. The evolutionary learning method divides the learning of the prediction network into multiple stages and adopts special strategies to enable the network to learn new knowledge on the basis of retaining existing knowledge. This expands the upper limit of the learning ability of the prediction network and can better adapt to the changing environment and task requirements. However, the existing evolutionary learning technology still has certain limitations: 1) Evolutionary learning is a long and multi-stage process. Most studies only focus on the knowledge transfer between two adjacent learning stages, and have not established an effective memory retention mechanism to prevent the loss of key memory in the long-term evolution process. 2) When evolutionary learning enters a new stage, a small amount of old samples are needed to consolidate the old knowledge, but there are still irrelevant samples in the old samples. How to select representative samples from the old samples has not been fully considered.

[0004] To this end, the present invention proposes a method for predicting the remaining useful life of rotating machinery based on long-term evolution learning, and by introducing a long-term enhancement mechanism, the present invention provides a more reliable and efficient solution for life prediction of rotating machinery.

[0005] Therefore, how to retain the key memories in old samples and consolidate the old knowledge in a reasonable and efficient way to improve the reliability and efficiency of the life prediction of rotating machinery is an urgent problem that technicians in this field need to solve. Summary of the invention

[0006] In view of this, the present invention provides a method for predicting the remaining useful life of rotating machinery based on long-term evolutionary learning. The prediction model can retain the acquired old knowledge for a long time and flexibly learn new knowledge, thereby realizing reliable and efficient prediction of the remaining useful life of rotating machinery.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A method for predicting the remaining useful life of a rotating machinery based on long-term evolution learning comprises the following steps:

[0009] A prediction model based on a recurrent convolutional neural network is constructed using adaptive weight connections, where the weight connections represent trainable parameters in the prediction model;

[0010] Enter the initial learning stage, obtain the vibration signal data set of the current stage, and train the prediction model;

[0011] After the current phase of training is completed, its stability is measured by measuring the learning contribution of the weight connection;

[0012] Select representative samples of the current stage according to the informativeness and diversity of the samples and save them to the knowledge cache area;

[0013] Entering the evolutionary learning stage, after obtaining the current stage vibration signal data set and the representative samples in the knowledge cache area, the prediction model trained in the previous stage is trained. In the current stage training, the update rate of the adaptive weight connection changes with the stability of the weight connection in the previous stage;

[0014] After the current phase of training is completed, its stability is measured by measuring the learning contribution of the weight connection;

[0015] Select representative samples of the current stage from the vibration signal data set of the current stage according to the informativeness, diversity and novelty of the samples, and save them to the knowledge cache area;

[0016] The prediction model is subjected to multiple evolutionary learning until the training objectives are met.

[0017] Furthermore, the prediction model based on recurrent convolutional neural network is expressed as F 0 (x 0 |θ 0 ), where x 0 represents the input sample of the prediction model in the initial learning stage, θ 0 Predict the trainable parameters in the model for the initial learning phase, called weight connections;

[0018] The circular convolution prediction model includes a circular convolution layer, a pooling layer and a fully connected layer;

[0019] The recurrent convolution layer adds a recurrent connection between the output and the input to form an information transfer loop, so that the recurrent convolution layer feeds the output back to the input;

[0020] The pooling layer uses a pooling function to perform a downsampling operation to obtain a feature vector;

[0021] The fully connected layer outputs a predicted value of the remaining useful life of the rotating machinery according to the feature vector.

[0022] Furthermore, in the initial learning stage, the learning contribution calculation formula of the weight connection is as follows:

[0023]

[0024] in, represents the learning contribution of the weight connection in stage 0, N0 is the number of training samples in stage 0, and D 0 is the vibration signal dataset of stage 0, x 0 is the input sample of the prediction model at stage 0, θ 0* is the connection weight of the prediction model trained in the initial learning phase, F 0* (x 0 ,θ 0* ) is the prediction model trained in the initial learning phase, y 0 is the remaining service life label of the rotating machinery in the current state at stage 0; L(·) is the square error loss function, is the jth trainable weight connection in stage 0;

[0025] The stability calculation formula of weight connection is as follows:

[0026]

[0027] in, is the stability of the jth weight connection in stage 0, and Z is the number of weight connections.

[0028] Furthermore, in the initial learning stage, the informativeness of the sample is calculated as:

[0029]

[0030] in, is the informativeness of the sample, indicating the degree of degradation information contained in the sample. The network gradient obtained from a high-informative sample is closer to the average gradient generated by the entire data set; D 0 is the vibration signal dataset of stage 0, is the i-th sample in the vibration signal data set of stage 0; ||·|| is the Euclidean norm; and Represent the gradient vector generated by a single sample and the average gradient vector generated by the entire data set respectively;

[0031] The calculation formula for sample diversity is:

[0032]

[0033] in, The diversity of samples indicates the dispersion of samples in the entire life cycle of the machine. Samples at different moments in the life of the machine will cause different gradients in the optimized network. are samples of the machine at different life moments in stage 0, and N0 is the number of samples participating in training in stage 0.

[0034] Furthermore, in the initial learning stage, the representativeness of the samples is quantified based on the two indicators of sample informativeness and diversity, expressed as:

[0035]

[0036] in, To select representative samples according to the order of their representativeness scores, λ is the ratio of the number of representative samples to the size of the training data set;

[0037] The final selected representative sample is saved in the knowledge cache area, denoted as in, represents the i-th representative sample selected in stage 0, The label of the actual remaining service life of the machine corresponding to the i-th representative sample in stage 0, n' 0 Indicates the number of representative samples selected.

[0038] Furthermore, in the evolutionary learning stage, the update formula of the weight connection is:

[0039]

[0040] Among them, θ j is the jth trainable weight connection in the first evolutionary learning stage, D 1 is the vibration signal dataset of the first evolutionary learning stage, ε 0 is the representative sample selected in stage 0, N1 represents the number of samples participating in the training, that is, the sum of the number of vibration data set samples in the current stage and the representative samples in the previous stage; x 1 is the input sample of the prediction model in the first evolutionary learning stage, y 1 is the remaining service life label of the rotating machinery in the current state in the first evolutionary learning stage; F(x 1 ,θ 1 ) is the prediction model trained in the previous stage used before the first evolutionary learning stage training;

[0041] is the update rate of the jth weight connection in the first evolutionary learning stage, and its calculation formula is:

[0042]

[0043] Among them, α is the initial learning rate, ξ is a value greater than 1, is the stability of the weight connection in stage 0.

[0044] Furthermore, in the evolutionary learning stage, the learning contribution calculation formula of the weight connection is as follows:

[0045]

[0046] Among them, N1 is the number of samples participating in the first evolutionary learning stage, F(x 1 ,θ 1* ) is the prediction model trained in the first evolutionary learning stage, y 1 is the current state remaining useful life label of the training data in the first evolutionary learning phase, θ j is the jth trainable weight connection in the first evolutionary learning stage, θ 1* The connection weights of the prediction model trained in the first evolutionary learning stage;

[0047] The stability calculation formula of weight connection is as follows:

[0048]

[0049] in, is the stability of the weight connection in the initial learning stage, U() represents the function symbol; f is the inhibition rate and f∈[0,1] is used to inhibit the update of the weight connection; is the learning contribution of the i-th weight connection in the first evolutionary learning stage, and Z is the number of weight connections.

[0050] Furthermore, in the evolutionary learning stage, the calculation method of sample informativeness and diversity is the same as that in the initial learning stage, and the calculation formula of sample novelty is:

[0051]

[0052] in, and They represent the gradient vector generated by a single sample and the average gradient vector generated by representative samples in the first evolutionary learning stage respectively; θ 1* The connection weights of the prediction model trained in the first evolutionary learning stage.

[0053] Furthermore, in the evolutionary learning stage, the update formula for the representative samples in the knowledge buffer is:

[0054]

[0055] Among them, ε 0 is the knowledge buffer sample in the initial stage, τ is the weight factor, For informational purposes, For diversity, For novelty.

[0056] Furthermore, representative samples selected from all previous learning stages accumulated in the knowledge cache are used to participate in the training of the prediction model in the subsequent stage, so that the prediction model gradually accumulates the knowledge learned in the previous stages.

[0057] It can be seen from the above technical solutions that, compared with the prior art, the present invention has the following beneficial effects:

[0058] The present invention aims to solve the problem of forgetting key memories in the long-term learning process under the framework of evolutionary learning, and at the same time provides an effective way to consolidate old knowledge for the new stage of learning process. Through the long-term learning of the degradation performance of rotating machinery by the prediction model, the high-precision prediction of the remaining service life of rotating machinery is achieved. At the same time, the present invention plans the propagation of key memories between stages from a long-term perspective, so that the propagation of memory is no longer limited to two adjacent evolutionary stages. By introducing the long-term enhancement mechanism, in the process of evolutionary learning, the weight connection is given certain adaptive characteristics, and its importance is measured by measuring the learning contribution of the weight connection. The plasticity of the weight connection is continuously adjusted according to the importance of the memory. The important weight connection is more stable and has a lower update rate, thereby highlighting the important sample memory, so that the key memory remains stable for a long time, and the "forgetfulness" problem of the model is alleviated.

[0059] At the same time, when the model enters a new evolutionary stage, the present invention does not simply randomly select several old samples for subsequent training, but introduces an association replay technology to select representative samples in a reasonable and efficient manner, save the representative samples in the knowledge buffer, and associate the new samples with the representative samples in the new evolutionary learning stage, and use them together to train the prediction model, so that the model can "review the old to learn the new", consolidate typical degraded knowledge in each learning stage, and enable the network to capture the complete degraded knowledge embedded in the previous stage and the current learning stage, thereby improving the learning efficiency of new knowledge and obtaining better generalization.

[0060] The method of the present invention simulates the memory and learning mode of the human brain and can perform reliable and efficient remaining useful life prediction in the face of dynamic changes in data sets, which is of great significance to the development of evolutionary learning in practical engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0062] Figure 1 A flowchart of a method for predicting the remaining useful life of a rotating machinery based on long-term evolution learning provided by the present invention;

[0063] Figure 2 A network structure diagram of the long-term evolution learning provided by the present invention;

[0064] Figure 3 A schematic diagram of the association replay mechanism provided by the present invention;

[0065] Figure 4 It is a schematic diagram of the effectiveness verification result of the method of the present invention;

[0066] Figure 5 The figure is a schematic diagram for comparing the performance of the method of the present invention with the existing method. DETAILED DESCRIPTION

[0067] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. 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.

[0068] like Figure 1 As shown, an embodiment of the present invention discloses a method for predicting the remaining service life of a rotating machinery based on long-term evolution learning, comprising the following steps:

[0069] A prediction model based on a recurrent convolutional neural network is constructed using adaptive weight connections, where the weight connections represent trainable parameters in the prediction model;

[0070] Enter the initial learning stage, obtain the vibration signal data set of the current stage, and train the prediction model;

[0071] After the current phase of training is completed, its stability is measured by measuring the learning contribution of the weight connection;

[0072] Select representative samples of the current stage according to the informativeness and diversity of the samples and save them to the knowledge cache area;

[0073] Entering the evolutionary learning stage, after obtaining the current stage vibration signal data set and the representative samples in the knowledge cache area, the prediction model trained in the previous stage is trained. In the current stage training, the update rate of the adaptive weight connection changes with the stability of the weight connection in the previous stage;

[0074] After the current phase of training is completed, its stability is measured by measuring the learning contribution of the weight connection;

[0075] Select representative samples of the current stage from the vibration signal data set of the current stage according to the informativeness, diversity and novelty of the samples, and save them to the knowledge cache area;

[0076] The prediction model is subjected to multiple evolutionary learning until the training objectives are met.

[0077] Below, each step of the present invention is further described.

[0078] 1. The current stage is the initial learning stage of the prediction model, denoted as s=0.

[0079] First, obtain the vibration signal dataset of the rotating machinery equipment from initial operation to failure in the current stage in, represents the samples obtained by the vibration sensor, A label indicating the actual remaining service life of the corresponding device in its current state. Y 0 The set of labels representing the remaining service life of the equipment in the current state at stage 0, n 0 Indicates D 0 The number of samples contained in .

[0080] Secondly, the adaptive weight connection is used to construct the recurrent convolutional neural network prediction model F 0 (x 0 |θ 0 ), where x 0 represents the input sample of the prediction model in the initial learning stage, θ 0 Predict the trainable parameters in the model for the initial learning phase, called weight connections;

[0081] The recurrent convolutional neural network is mainly composed of a stack of recurrent convolutional layers, pooling layers and fully connected layers;

[0082] The recurrent convolution layer adds a recurrent connection between the output and the input to form an information transmission loop, so that the recurrent convolution layer feeds the output back to the input and has the ability to remember time information. For the qth recurrent convolution layer, the state of the convolution layer at time t is It can be expressed as:

[0083]

[0084] In the formula, represents the input of the recurrent convolutional layer, represents the state of the convolutional layer at time t-1, and σ(·) represents the sigmoid activation function. For the pooling layer, the pooling function is used to perform the downsampling operation to obtain the feature vector.

[0085] For the pooling layer, a downsampling operation is performed using a pooling function to obtain a feature vector;

[0086] Afterwards, the feature vector is fed forward to the subsequent three fully connected layers. For the first two fully connected layers, the number of neurons in each layer is F, and the ReLU function is used as the activation function, where the ReLU function expression is: f(x) = max(0, x). The third fully connected layer is the output layer, which contains only one neuron to output the remaining service life prediction value of the rotating machinery.

[0087] At this stage, set the loss function L 0 , the loss function is the mean square error loss, and the specific formula is: Among them, n 0 is the number of samples involved in training, y k is the true value of the remaining useful life of the sample, is the predicted value of the remaining useful life of the sample. 0 (x 0 |θ 0 ) for training, the process is as follows:

[0088] i. Set the number of training iterations T, training batch bs, and initial learning rate α;

[0089] ii. Use the dataset D in this stage 0 Use the training batch as input for the prediction model;

[0090] iii. Automatically learn multi-level data representation from input data using multiple layers of recurrent convolutional layers and pooling layers;

[0091] iv. Feedforward calculates the output value of each layer until the last layer, and calculates the loss function L based on the output value of the last layer 0 ;

[0092] v. Back propagation calculates the error of each layer and uses the gradient descent algorithm to update the weight connection θ 0 , the update formula is:

[0093] vi. Repeat steps iii to v until the number of training iterations reaches T.

[0094] vii. Get the trained model F0* (x 0 |θ 0* ), where the updated weight connection is θ 0* .

[0095] In the initial learning stage, after the prediction model is trained, the weight connection stability measurement and representative sample selection are performed as follows:

[0096] 1) Learning contributions by measuring weight connections Get the stability of weight connection Z represents the number of weight connections. Weight connections with strong learning contributions have greater stability and smaller learning rates (update rates) in subsequent stages, thereby preventing the degradation of key memories during long-term evolutionary learning.

[0097] 1.1) Measure the learning contribution of the weight connection, the calculation formula is as follows:

[0098]

[0099] in, represents the learning contribution of the weight connection in stage 0, N0 is the number of training samples in stage 0, and D 0 is the vibration signal dataset of stage 0, x 0 is the input sample of the prediction model at stage 0, θ 0* is the weight connection of the prediction model after the initial learning phase, F 0* (x 0 ,θ 0* )Same as F 0* (x 0 |θ 0* ), represents the prediction model trained in the initial stage, y 0 is the remaining service life label of the rotating machinery in the current state at stage 0; L(·) is the square error loss function, is the jth trainable weight connection in stage 0.

[0100] 1.2) The stability of the weight connection is measured by the learning contribution of the weight connection. The calculation formula is as follows:

[0101]

[0102] in, is the stability of the jth weight connection in stage 0, and Z is the number of weight connections.

[0103] 2) Construct a knowledge buffer, which is used to store representative samples and propagate them in the future evolutionary learning process. In the process of selecting representative samples, the association replay technology is introduced to select representative samples in a reasonable and efficient way, and the representativeness of the samples is evaluated by the two indicators of informativeness and diversity.

[0104] 2.1) The calculation formula of the sample information is:

[0105]

[0106] in, is the informativeness of the sample, indicating the degree of degradation information contained in the sample. The network gradient obtained from a high-informative sample is closer to the average gradient generated by the entire data set; D 0 is the vibration signal dataset of stage 0, is the i-th sample in the vibration signal data set of stage 0; ||·|| is the Euclidean norm;

[0107] and They represent the gradient vector generated by a single sample and the average gradient vector generated by the entire data set, respectively. The calculation formula is as follows:

[0108]

[0109]

[0110] In the formula, Represents the first weight connection in stage 0, Represents the Zth weight connection in stage 0.

[0111] 2.2) The calculation formula of sample diversity is:

[0112]

[0113] in, The diversity of samples indicates the dispersion of samples in the entire life cycle of the machine. Samples at different moments in the life of the machine will cause different gradients in the optimized network. are samples of the machine at different life moments in stage 0, and N0 is the number of samples participating in training in stage 0.

[0114] Afterwards, the representativeness of the sample is quantified based on the two indicators of sample informativeness and diversity, expressed as:

[0115]

[0116] in, Representative samples are selected according to the order of their representativeness scores; λ is the ratio of the number of representative samples to the size of the training data set, which is usually set to 0.1. The number of representative samples is selected according to the size of λ.

[0117] Afterwards, the trained prediction model is saved and recorded as At the same time, a knowledge buffer is constructed to store the final selected representative samples in the knowledge buffer, denoted as in, represents the i-th representative sample selected in stage 0, The label of the actual remaining service life of the machine corresponding to the i-th representative sample in stage 0, n' 0 Indicates the number of representative samples selected.

[0118] 2. Enter the first evolutionary learning stage, denoted as s=1.

[0119] like Figure 2 As shown, first, obtain the vibration signal dataset of the rotating machinery equipment from initial operation to failure in the current stage in, represents the samples obtained by the vibration sensor, Indicates the current status of the corresponding device is really a label of the remaining service life. Y 1 represents the set of labels of the remaining useful life of the equipment in the current state in the first stage of evolutionary learning, and n 1 Indicates D 1 Then the newly acquired equipment life cycle data set With representative samples Combined and used together for prediction model training. The training in the evolutionary learning stage uses the model trained in the previous stage. A long-term enhancement mechanism is introduced to update the weight connection in combination with the stability of the weight connection calculated in the previous stage.

[0120] At this stage, the specific training process of the prediction model is as follows:

[0121] i. Set the number of training iterations T, training batch bs, and initial learning rate α;

[0122] ii. Use the dataset D in this stage 1 Use the training batch as input for the prediction model;

[0123] iii. Automatically learn multi-level data representation from input data using multiple layers of recurrent convolutional layers and pooling layers;

[0124] iv. Feedforward calculates the output value of each layer until the last layer, and calculates the loss function L based on the output value of the last layer 1 ;

[0125] v. Back propagation calculates the error of each layer and uses the gradient descent algorithm to update the weight connection θ 1 .

[0126] It should be noted that the update rate of the adaptive weight connection in the evolutionary learning stage varies with the stability of the weight connection. The update formula of the weight connection is:

[0127] In the evolutionary learning stage, the update formula of the weight connection is:

[0128]

[0129] Among them, θ j is the jth weight connection, D 1 is the vibration signal dataset of the first evolutionary learning stage, ε 0 is the representative sample selected in stage 0, N1 represents the number of samples participating in the training, that is, the sum of the number of vibration data set samples in the current stage and the representative samples in the previous stage; x 1 is the input sample of the prediction model in the first evolutionary learning stage, y 1 is the remaining service life label of the rotating machinery in the current state in the first evolutionary learning stage; F(x 1 ,θ 1 ) is the prediction model trained in the previous stage used before the first evolutionary learning stage training;

[0130] is the update rate of the jth weight connection in the first evolutionary learning stage, and its calculation formula is:

[0131]

[0132] Among them, α is the initial learning rate, ξ is a value greater than 1, is the stability of the weight connection in stage 0.

[0133] vi. Repeat steps iii to v until the number of training iterations reaches T.

[0134] vii. Get the trained model F 1* (x 1 |θ 1* ), where the updated weight connection is θ 1* .

[0135] After the model training is completed in this stage, weight connection stability measurement and representative sample selection are also required, as follows:

[0136] 1) By measuring the learning contribution of weight connections Get the stability of weight connection Z represents the number of weighted connections. The stability calculation in the evolutionary learning phase is different from that in the initial learning phase. Specifically, the weighted connection stability is calculated as follows:

[0137]

[0138] in, is the stability of the weight connection in the initial learning stage, U() represents the function symbol; f is the inhibition rate and f∈[0,1] is used to inhibit the update of the weight connection; Z is the number of weight connections; I j 1 is the learning contribution of the jth weight connection in the first learning stage, and the calculation formula is:

[0139]

[0140] Among them, N1 is the number of samples participating in the first evolutionary learning stage, F(x 1 ,θ 1* )Same as F 1* (x 1 |θ 1* ), is the prediction model trained in the first evolutionary learning stage, y 1 is the current state remaining useful life label of the training data in the first evolutionary learning phase, θ j is the jth trainable weight connection in the first evolutionary learning stage, θ 1* The weight connections of the prediction model after the first evolutionary learning phase is completed.

[0141] 2) Update the knowledge buffer. More new samples are obtained in the evolutionary learning phase, and the buffer needs to be updated to improve the representativeness of the buffer samples. Specifically, Figure 3 As shown, in the process of selecting representative samples, the association replay technology is introduced to select representative samples in a reasonable and efficient way. On the basis of the two evaluation indicators of informativeness and diversity, another indicator is added, which is novelty, so that the information contained in the selected samples should be as different as possible from the information contained in the knowledge buffer.

[0142] The calculation formula of the novelty of the sample is:

[0143]

[0144] in, and They represent the gradient vector generated by a single sample and the average gradient vector generated by representative samples in the first evolutionary learning stage respectively; The dataset D is the first evolutionary learning stage1 The i-th sample in .

[0145] At this stage, the update formula for the representative samples in the knowledge buffer is:

[0146]

[0147] Among them, ε 0 is the knowledge buffer sample in the initial stage, τ is the weight factor, For informational purposes, For diversity, For novelty.

[0148] 3. After that, the prediction model has been in the evolutionary learning stage. As time goes by, the data of rotating machinery is constantly updated. The representative samples selected from all previous learning stages accumulated in the knowledge cache are used to participate in the training of the prediction model in the subsequent stages, so that the prediction model gradually accumulates the knowledge learned in the previous stages. When the prediction model completes a stage of learning, it is necessary to calculate the stability of the weight connection and use the association replay technology to select representative samples to update the knowledge buffer. When the model enters a new stage, it uses the long-term memory enhancement strategy and combines the weight connection stability to update the model to achieve long-term evolutionary learning. The detailed method is the same as 2.

[0149] The effectiveness of the method of the present invention is further illustrated by an experiment below.

[0150] Taking the accelerated degradation experiment of rolling bearings as a case, the run-to-failure data collected in the experiment are used to evaluate the effectiveness of the method of the present invention. The rolling bearing tested in this experiment is LDKUER204 type. The bearing is driven by an AC motor and the speed is set by the motor speed controller. The radial load is provided by a hydraulic loading device and applied to the housing of the tested bearing. The vibration signal of the tested bearing is collected by vertical and horizontal accelerometers with a sampling frequency of 25.6kHz, that is, 2560 data points are recorded every 0.1 second and collected once every 12 seconds. Three different working conditions were simulated by setting different speeds and different radial forces. The speeds were set to 2100rpm, 2400rpm, and 2700rpm, respectively, and the radial forces were set to 12KN, 11KN, and 10KN, respectively, and 30 bearings were tested.

[0151] In order to construct an evolutionary learning scenario, the tested bearings were arranged in 8 stages to be learned by the prediction model, in which the test data set remained unchanged, and the remaining marked run-to-failure data were assigned to different stages in turn. The experiment compared the method of the present invention with the existing evolutionary learning method. Method A adopted the same model structure as the method of the present invention, but did not adopt the long-term enhancement mechanism and the associated replay technology. Method B constructed the same model structure as the method of the present invention, used the associated replay technology, but did not adopt the long-term enhancement mechanism in the evolutionary learning process. Method C constructed the same model structure as the method of the present invention, used the long-term enhancement mechanism to retain the memory of important samples, but did not adopt the associated replay technology. In the experimental results, the present invention uses the root mean square error (RMSE) and the symmetric mean absolute percentage error (SMAPE) to evaluate the performance of the proposed continuous learning framework. In addition, the average forgetting value (AF) is designed to measure the forgetting of the prediction model, and its calculation formula is Where f s k =max t∈{k,...s-1} (SMAPE s,k -SMAPE t,k ), represents the amount of forgetting after the prediction model completes the sth stage of learning, SMAPE s,k is the cumulative mean absolute percentage error on the kth training set after the prediction model completes the sth stage.

[0152] The experimental condition configuration is shown in Table 1, and the constructed evolutionary learning scenario is shown in Table 2.

[0153] Table 1. Bearing working condition configuration in the embodiment

[0154] Simulation conditions Speed Radial force Bearing No. Condition 1 2100rpm 12KN Bearing 1-1 to Bearing 1-10 Condition 2 2400rpm 11KN Bearing 2-1 to Bearing 2-10 Condition 3 2700rpm 10KN Bearing 3-1 to Bearing 3-10

[0155] Table 2 Evolutionary learning scenarios constructed in the embodiment

[0156]

[0157]

[0158] The experimental results are summarized in Table 3. Figure 4 and Figure 5 .

[0159] Table 3. Evaluation results of the performance of the method of the present invention

[0160] Evaluation Metrics Phase 1 Phase 3 Phase 5 Phase 7 RMSE 184.9008 133.8080 132.5821 79.3123 SMAPE 1.2123 0.8986 0.7156 0.5161

[0161] It can be seen from the experimental results that the present invention has a lower forgetting rate and better remaining useful life prediction performance in the evolutionary learning process, and with the changes in each stage, the method of the present invention has obvious advantages over other methods.

[0162] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0163] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting the remaining useful life of rotating machinery based on long-term evolution learning, characterized in that: The following steps are involved: A prediction model based on a recurrent convolutional neural network is constructed using adaptive weight connections, where the weight connections represent trainable parameters in the prediction model; Enter the initial learning stage, obtain the vibration signal data set of the current stage, and train the prediction model; After the current phase of training is completed, its stability is measured by measuring the learning contribution of the weight connection; Select representative samples of the current stage according to the informativeness and diversity of the samples and save them to the knowledge cache area; Entering the evolutionary learning stage, after obtaining the current stage vibration signal data set and the representative samples in the knowledge cache area, the prediction model trained in the previous stage is trained. In the current stage training, the update rate of the adaptive weight connection changes with the stability of the weight connection in the previous stage; After the current phase of training is completed, its stability is measured by measuring the learning contribution of the weight connection; Select representative samples of the current stage from the vibration signal data set of the current stage according to the informativeness, diversity and novelty of the samples, and save them to the knowledge cache area; Perform multiple evolutionary learning on the prediction model until the training objectives are met; In the initial learning stage, the learning contribution of the weight connection is calculated as follows: in, represents the learning contribution of the weight connection in stage 0, N0 is the number of training samples in stage 0, and D 0 is the vibration signal dataset of stage 0, x 0 is the input sample of the prediction model at stage 0, θ 0* is the connection weight of the prediction model trained in the initial learning phase, F 0* (x 0 ,θ 0* ) is the prediction model trained in the initial learning phase, y 0 is the remaining service life label of the rotating machinery in the current state at stage 0; L(·) is the square error loss function, is the jth trainable weight connection in stage 0; The stability calculation formula of weight connection is as follows: in, } is the stability of the jth weight connection at stage 0, and Z is the number of weight connections; In the initial learning stage, the informativeness of the sample is calculated as: in, is the informativeness of the sample, indicating the degree of degradation information contained in the sample. The network gradient obtained from a high-informative sample is closer to the average gradient generated by the entire data set; D 0 is the vibration signal dataset of stage 0, is the i-th sample in the vibration signal data set of stage 0; ||·|| is the Euclidean norm; and Represent the gradient vector generated by a single sample and the average gradient vector generated by the entire data set respectively; The calculation formula for sample diversity is: in, The diversity of samples indicates the dispersion of samples in the entire life cycle of the machine. Samples at different moments in the life of the machine will cause different gradients in the optimized network. are samples of the machine at different life moments in stage 0, and N0 is the number of samples participating in training in stage 0.

2. The method for predicting the remaining useful life of rotating machinery based on long-term evolution learning according to claim 1 is characterized in that: The prediction model based on recurrent convolutional neural network is represented as F 0 (x 0 |θ 0 ), where x 0 represents the input sample of the prediction model in the initial learning stage, θ 0 Predict the trainable parameters in the model for the initial learning phase, called weight connections; The recurrent convolutional neural network includes a recurrent convolutional layer, a pooling layer and a fully connected layer; The recurrent convolution layer adds a recurrent connection between the output and the input to form an information transfer loop, so that the recurrent convolution layer feeds the output back to the input; The pooling layer uses a pooling function to perform a downsampling operation to obtain a feature vector; The fully connected layer outputs a predicted value of the remaining useful life of the rotating machinery according to the feature vector.

3. The method for predicting the remaining useful life of rotating machinery based on long-term evolution learning according to claim 1, characterized in that: In the initial learning stage, the representativeness of the sample is quantified based on the two indicators of sample informativeness and diversity, expressed as: in, To select representative samples according to the order of their representativeness scores, λ is the ratio of the number of representative samples to the size of the training data set; The final selected representative sample is saved in the knowledge cache area, denoted as Among them, x′ i 0 represents the i-th representative sample selected in stage 0, y′ i 0 The label of the actual remaining service life of the machine corresponding to the i-th representative sample in stage 0, n' 0 Indicates the number of representative samples selected.

4. The method for predicting the remaining useful life of rotating machinery based on long-term evolution learning according to claim 1, characterized in that: In the evolutionary learning stage, the update formula of the weight connection is: Among them, θ j is the jth trainable weight connection in the first evolutionary learning stage, D 1 is the vibration signal dataset of the first evolutionary learning stage, ε 0 is the representative sample selected in stage 0, N1 represents the number of samples participating in the training, that is, the sum of the number of vibration data set samples in the current stage and the representative samples in the previous stage; x 1 is the input sample of the prediction model in the first evolutionary learning stage, y 1 is the remaining service life label of the rotating machinery in the current state in the first evolutionary learning stage; F(x 1 ,θ 1 ) is the prediction model trained in the previous stage used before the first evolutionary learning stage training; is the update rate of the jth weight connection in the first evolutionary learning stage, and its calculation formula is: Among them, α is the initial learning rate, ξ is a value greater than 1, is the stability of the weight connection in stage 0.

5. The method for predicting the remaining useful life of rotating machinery based on long-term evolution learning according to claim 4 is characterized in that: In the evolutionary learning stage, the learning contribution calculation formula of the weight connection is as follows: Among them, N1 is the number of samples participating in the first evolutionary learning stage, F(x 1 ,θ 1* ) is the prediction model trained in the first evolutionary learning stage, θ 1* The connection weights of the prediction model trained in the first evolutionary learning stage; The stability calculation formula of weight connection is as follows: in, is the stability of the weight connection in the initial learning stage, U() represents the function symbol; f is the inhibition rate and f∈[0,1] is used to inhibit the update of the weight connection; is the learning contribution of the i-th weight connection in the first evolutionary learning stage, and Z is the number of weight connections.

6. The method for predicting the remaining useful life of rotating machinery based on long-term evolution learning according to claim 1, characterized in that: In the evolutionary learning stage, the calculation method of sample informativeness and diversity is the same as that in the initial learning stage, and the calculation formula of sample novelty is: in, and They represent the gradient vector generated by a single sample and the average gradient vector generated by representative samples in the first evolutionary learning stage respectively; is the i-th sample in the dataset in the first evolutionary learning stage, θ 1* The connection weights of the prediction model trained in the first evolutionary learning stage.

7. The method for predicting the remaining useful life of rotating machinery based on long-term evolution learning according to claim 1, characterized in that: In the evolutionary learning stage, the update formula of the representative samples in the knowledge buffer is: Among them, ε 0 is the knowledge buffer sample in the initial stage, τ is the weight factor, For informational purposes, For diversity, For novelty.

8. The method for predicting the remaining useful life of rotating machinery based on long-term evolution learning according to claim 1, characterized in that: The representative samples selected from all previous learning stages accumulated in the knowledge cache are used to participate in the training of the prediction model in the subsequent stage, so that the prediction model can gradually accumulate the knowledge learned in the previous stage.

Images