A method and system for predicting the remaining life of degraded equipment under zero-life label

Abstract

The present invention relates to a method and system for predicting the remaining life of a degraded device under a zero-life label. The method comprises: setting different network parameters to obtain different LSTM network models under zero-life labels; constructing a training data set by a sliding window method according to the time series under the zero-life label, training different bidirectional LSTM network models, and obtaining multiple trained bidirectional LSTM network models; obtaining the current degradation data of the degraded device to be tested, inputting it into the trained bidirectional LSTM network model, and outputting multiple current remaining lives; determining the network posterior probabilities of different trained bidirectional LSTM network models according to the training data set and the current data set using a Bayesian model; fusing multiple current remaining lives according to the network posterior probability, and predicting the final remaining life of the degraded device to be tested under the zero-life label. The present invention can reduce the remaining life prediction error.

Description

Technical Field

[0001] The present invention relates to the field of reliability engineering, and in particular to a method and system for predicting the remaining life of degraded equipment under a zero-life label. Background Art

[0002] With the continuous development of sensor technology and data storage technology, more and more degradation data are available in the current health management field. Since the degradation data itself contains rich system degradation information, how to establish a degradation model that can accurately describe the degradation law of the system based on the degradation data obtained by monitoring has become the research focus of the current data-driven remaining life prediction method. To accurately describe the degradation of complex systems containing a large number of physical and chemical processes, it is usually necessary to build a multi-degree-of-freedom degradation model. When using statistical data methods to model, the difficulty of implementing the model parameter estimation algorithm will increase exponentially with the increase in the number of parameters. Therefore, the remaining life prediction results can usually only be obtained through numerical simulation methods. In this case, the use of neural network methods to predict the remaining life of complex systems will have a greater advantage: first, the neural network method solves the problem of incorrect model selection when predicting the remaining life; second, the neural network method only needs to increase the model's degrees of freedom and fitting ability by stacking the number of network layers. Therefore, the neural network method has received more and more attention in the remaining life prediction of complex degraded systems.

[0003] Although the neural network method has achieved outstanding results in the prediction of the remaining life of the system, the selection and setting of hyperparameters in the neural network training has always been a key problem of the neural network method. Given a training data set, the weight parameters of the neural network obtained by training will be slightly different due to the differences in hyperparameters such as the number of training iterations, learning rate and sample batch size. For the life label network, since the training algorithm can ensure the optimality of the neural network on the training data set, the life label network obtained under different hyperparameter settings can ensure that the remaining life prediction results have similar optimality. However, for the zero-life label network whose training label is the degradation value, the training algorithm is difficult to ensure the consistency of local degradation prediction while ensuring the overall prediction accuracy of the degradation trajectory. Therefore, when the local degradation prediction error is relatively large, the remaining life of the system obtained by iterative calculation of the degradation prediction results will have a large deviation. Therefore, the uncertainty of the weight of the zero-life label neural network caused by different training conditions will cause uncertainty in the remaining life prediction. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for predicting the remaining life of degraded equipment under a zero-life label, so as to solve the problem of large error in predicting the remaining life of degraded equipment.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A method for predicting the remaining life of degraded equipment under a zero-life label, comprising:

[0007] Setting different network parameters to obtain bidirectional LSTM network models under different zero-lifetime labels; the network parameters include network structure, sample batch size, learning rate, and number of training iterations;

[0008] According to the time series under the zero-life label, a training data set is constructed by a sliding window method, and different bidirectional LSTM network models are trained to obtain multiple trained bidirectional LSTM network models; the training data set includes historical degradation data of the degraded equipment at different times and the historical remaining life of the degraded equipment corresponding to the historical degradation data at different times;

[0009] Obtaining current degradation data of the degraded device to be tested, inputting the data into the trained bidirectional LSTM network model, and outputting a plurality of current remaining lifespans;

[0010] Determining network posterior probabilities of different trained bidirectional LSTM network models using a Bayesian model according to the training data set and the current data set; the current data set includes the current degradation data and the current remaining life;

[0011] The multiple current remaining lifespans are fused according to the network posterior probability to predict the final remaining lifespan of the degraded device to be tested under the zero lifespan label.

[0012] Optionally, the LSTM neuron of the Lth layer at time t in the forward network layer of the bidirectional LSTM network model is:

[0013]

[0014] in, is the forward L-th layer forget gate; σ is the Sigmoid activation function; is the weight between the forward L-th layer forget gate and state h; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the forward network; is the weight between the forward L-th layer forget gate and the input x; is the output of the recurrent layer state of the memory unit at the L-1th layer of the forward network at time t; is the bias of the forward forget gate of the Lth layer; is the forward L-th layer input gate; is the weight between the forward L-th layer input gate and h; is the weight between the forward L-th layer input gate and x; is the bias of the forward L-th layer input gate; is the forward L-th layer output gate; is the weight between the forward L-th layer output gate and h; is the weight between the forward L-th layer output gate and x; is the bias of the forward L-th layer output gate; To filter the input information before t time in the forward Lth layer, the hyperbolic tangent here can realize the scaling of input information and can be regarded as a kind of short-term memory; is the weight between the forward L-th layer short-term memory and the state h; is the weight between the forward L-th layer short-term memory and the input x; is the bias in the forward L-th layer short-term memory; is the state of the memory unit in the forward Lth layer at time t; is the state of the memory unit in the forward Lth layer at time t-1; is the output of the recurrent layer state of the forward memory unit at the Lth layer at time t.

[0015] The LSTM neuron of the Lth layer at time t in the backward network layer of the bidirectional LSTM network model is:

[0016]

[0017] in, It is the backward forget gate of the Lth layer; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the backward network; is the weight between the backward L-th layer forget gate and state h; is the weight between the backward L-th layer forget gate and the input x; is the output of the recurrent layer state of the memory unit at the L-1th layer of the backward network at time t; is the bias of the backward forget gate of the Lth layer; is the input gate of the backward Lth layer; is the weight between the input gate of the backward Lth layer and h; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the backward network; is the weight between the input gate of the backward Lth layer and x; is the bias of the backward input gate; is the backward output gate; is the weight between the backward output gate and h; is the weight between the output gate of the backward Lth layer and x; is the bias of the output gate of the backward Lth layer; The input information of the Lth layer of the backward network is filtered out before t moments, which can be regarded as a kind of short-term memory; is the weight between the backward L-th layer short-term memory and the state h; is the weight between the backward L-th layer short-term memory and the input x; is the bias in the Lth layer of short-term memory; is the state of the memory unit at the Lth layer at time t; is the state of the memory unit in the backward Lth layer at time t-1; is the output of the recurrent layer state of the memory unit at time t in the Lth layer of the backward network;

[0018] The output of the bidirectional LSTM network model is:

[0019]

[0020] Among them, y t is the output of the bidirectional LSTM network model; is the weight between the last layer of the forward hidden layer and the model output; is the output of the last layer of the forward hidden layer; is the weight between the last layer of the backward hidden layer and the model output; is the output of the last layer of the backward hidden layer; b y is the output bias of the bidirectional LSTM network model.

[0021] Optionally, determining the network posterior probabilities of different trained bidirectional LSTM network models using a Bayesian model according to the training data set and the current data set specifically includes:

[0022] Determine the training test accuracy of the trained bidirectional LSTM network model according to the training data set;

[0023] Determine the network prior probability of the trained bidirectional LSTM network model according to the training test accuracy; the network prior probability is used as the prior probability of the Bayesian model;

[0024] Determine the overall prediction accuracy of the trained bidirectional LSTM network model according to the current data set;

[0025] determining a prediction degradation probability according to the overall prediction accuracy;

[0026] Based on the Bayesian model, the network posterior probability of the trained bidirectional LSTM network model is determined according to the network prior probability and the predicted degradation probability.

[0027] Optionally, the network prior probability is:

[0028]

[0029] Among them, P TA (Mc ) is the network prior probability; M C is the trained bidirectional LSTM network model; M C The training and testing accuracy of ; S is the number of candidate models; i is the index of the sample label; j is the index of the candidate model; N is the number of test samples; is the historical degradation data; Represents M c Predicted historical remaining life.

[0030] Optionally, the predicted degradation probability is:

[0031]

[0032] Among them, P PA (X m |M c ) is the network posterior probability; X m is the sequence of the current remaining lifespan; is the overall prediction accuracy of the cth network model; is the overall prediction accuracy of the i-th network model; m is any moment; W is the length of the input sequence; k is the number of prediction steps; x j is the degradation label of the jth sample; y c，j The predicted value obtained by inputting the j-th sample into the c-th network model.

[0033] Optionally, the network posterior probability is:

[0034]

[0035] Among them, P A (M c |X m ) is the network posterior probability.

[0036] A system for predicting the remaining life of degraded equipment under a zero-life label, comprising:

[0037] The bidirectional LSTM network model determination module is used to set different network parameters to obtain bidirectional LSTM network models under different zero-lifetime labels; the network parameters include network structure, sample batch size, learning rate and number of training iterations. ;

[0038] A training module is used to construct a training data set by a sliding window method according to a time series under a zero-lifetime label, train different bidirectional LSTM network models, and obtain multiple trained bidirectional LSTM network models; the training data set includes historical degradation data of the degraded equipment at different times and historical remaining life of the degraded equipment corresponding to the historical degradation data at different times;

[0039] A current remaining life output module is used to obtain current degradation data of the degraded equipment to be tested, input it into the trained bidirectional LSTM network model, and output multiple current remaining lives;

[0040] A network posterior probability determination module, used to determine the network posterior probabilities of different trained bidirectional LSTM network models using a Bayesian model according to the training data set and the current data set; the current data set includes the current degradation data and the current remaining life;

[0041] The final remaining life prediction module is used to fuse multiple current remaining lifespans according to the network posterior probability to predict the final remaining lifespan of the degraded device to be tested under the zero lifespan label.

[0042] Optionally, the LSTM neuron of the Lth layer at time t in the forward network layer of the bidirectional LSTM network model is:

[0043]

[0044] in, is the forward L-th layer forget gate; σ is the Sigmoid activation function; is the weight between the forward L-th layer forget gate and state h; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the forward network; is the weight between the forward L-th layer forget gate and the input x; is the output of the recurrent layer state of the memory unit at the L-1th layer of the forward network at time t; is the bias of the forward forget gate of the Lth layer; is the forward L-th layer input gate; is the weight between the forward L-th layer input gate and h; is the weight between the forward L-th layer input gate and x; is the bias of the forward L-th layer input gate; is the forward L-th layer output gate; is the weight between the forward L-th layer output gate and h; is the weight between the forward L-th layer output gate and x; is the bias of the forward L-th layer output gate; To filter the input information before t time in the forward Lth layer, the hyperbolic tangent here can realize the scaling of input information and can be regarded as a kind of short-term memory; is the weight between the forward L-th layer short-term memory and the state h; is the weight between the forward L-th layer short-term memory and the input x; is the bias in the forward L-th layer short-term memory; is the state of the memory unit in the forward Lth layer at time t; is the state of the memory unit in the forward Lth layer at time t-1; is the output of the recurrent layer state of the forward memory unit at the Lth layer at time t.

[0045] The LSTM neuron of the Lth layer at time t in the backward network layer of the bidirectional LSTM network model is:

[0046]

[0047] in, It is the backward forget gate of the Lth layer; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the backward network; is the weight between the backward L-th layer forget gate and state h; is the weight between the backward L-th layer forget gate and the input x; is the output of the recurrent layer state of the memory unit at the L-1th layer of the backward network at time t; is the bias of the backward forget gate of the Lth layer; is the input gate of the backward Lth layer; is the weight between the input gate of the backward Lth layer and h; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the backward network; is the weight between the input gate of the backward Lth layer and x; is the bias of the backward input gate; is the backward output gate; is the weight between the backward output gate and h; is the weight between the output gate of the backward Lth layer and x; is the bias of the output gate of the backward Lth layer; The input information of the Lth layer of the backward network is filtered out before t moments, which can be regarded as a kind of short-term memory; is the weight between the backward L-th layer short-term memory and the state h; is the weight between the backward L-th layer short-term memory and the input x; is the bias in the Lth layer of short-term memory; is the state of the memory unit at the Lth layer at time t; is the state of the memory unit in the backward Lth layer at time t-1; is the output of the recurrent layer state of the memory unit at time t in the Lth layer of the backward network;

[0048] The output of the bidirectional LSTM network model is:

[0049]

[0050] Among them, y t is the output of the bidirectional LSTM network model; is the weight between the last layer of the forward hidden layer and the model output; is the output of the last layer of the forward hidden layer; is the weight between the last layer of the backward hidden layer and the model output; is the output of the last layer of the backward hidden layer; b y is the output bias of the bidirectional LSTM network model.

[0051] Optionally, the network posterior probability determination module specifically includes:

[0052] A training and testing accuracy determination unit, used to determine the training and testing accuracy of the trained bidirectional LSTM network model according to the training data set;

[0053] A network prior probability determination unit, used to determine the network prior probability of the trained bidirectional LSTM network model according to the training test accuracy; the network prior probability is used as the prior probability of the Bayesian model;

[0054] An overall prediction accuracy determination unit, used to determine the overall prediction accuracy of the trained bidirectional LSTM network model according to the current data set;

[0055] A prediction degradation probability determination unit, configured to determine the prediction degradation probability according to the overall prediction accuracy;

[0056] A network posterior probability determination unit is used to determine the network posterior probability of the trained bidirectional LSTM network model based on the Bayesian model, according to the network prior probability and the predicted degradation probability.

[0057] Optionally, the network prior probability is:

[0058]

[0059] Among them, P TA (M c ) is the network prior probability; M C is the trained bidirectional LSTM network model; M C The training and testing accuracy of ; S is the number of candidate models; i is the index of the sample label; j is the index of the candidate model; N is the number of test samples; is the historical degradation data; Represents M c Predicted degradation data.

[0060] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the present invention provides a method and system for predicting the remaining life of degraded equipment under a zero-life label, which overcomes the influence of the setting of the training label interval step number in the network training data set, and realizes the prediction of the remaining life of the equipment at any monitoring point. By reconstructing the predicted current degradation data and historical degradation data as the input of the zero-life label network, the influence of the network training test accuracy and the historical prediction accuracy under the zero-life label on the current remaining life prediction accuracy is considered, which can effectively reduce the system remaining life prediction error caused by the uncertainty of the network model, provide a reference for the maintenance arrangement of the equipment, and has a broad application space. 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 will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0062] Figure 1 A flow chart of the method for predicting the remaining life of degraded equipment under a zero-life label provided by the present invention;

[0063] Figure 2 This is a structural diagram of the system for predicting the remaining life of degraded equipment under the zero-life label provided by the present invention;

[0064] Figure 3 This is a diagram of the degradation trajectory of the battery discharge capacitance provided by the present invention. DETAILED DESCRIPTION

[0065] 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.

[0066] The purpose of the present invention is to provide a method and system for predicting the remaining life of degraded equipment under a zero-life label, which can reduce the remaining life prediction error.

[0067] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0068] Figure 1 The flowchart of the method for predicting the remaining life of degraded equipment under the zero-life label provided by the present invention is as follows: Figure 1As shown, a method for predicting the remaining life of a degraded device under a zero-life label includes:

[0069] Step 101: setting different network parameters to obtain a bidirectional long-short term memory (LSTM) network model under different zero-lifetime labels; the network parameters include network structure, sample batch size, learning rate and number of training iterations.

[0070] Step 102: According to the time series under the zero-lifetime label, a training data set is constructed by a sliding window method, and different bidirectional LSTM network models are trained to obtain multiple trained bidirectional LSTM network models; the training data set includes historical degradation data of the degraded device at different times and the historical remaining life of the degraded device corresponding to the historical degradation data at different times. When the degraded device is a battery, the historical degradation data can be the charge and discharge capacity; when the degraded device is a bearing, the historical degradation data can be the vibration amplitude; when the degraded device is an inertial device, the historical degradation data can be drift data, etc.; different degraded devices correspond to different historical degradation data.

[0071] The LSTM neurons in the Lth layer at time t in the forward network layer of the bidirectional LSTM network model are:

[0072]

[0073] in, is the forward L-th layer forget gate; σ is the Sigmoid activation function; is the weight between the forward L-th layer forget gate and state h; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the forward network; is the weight between the forward L-th layer forget gate and the input x; is the output of the recurrent layer state of the memory unit at the L-1th layer of the forward network at time t; is the bias of the forward L-th layer forget gate; is the forward L-th layer input gate; is the weight between the forward L-th layer input gate and h; is the weight between the forward L-th layer input gate and x; is the bias of the forward L-th layer input gate; is the forward L-th layer output gate; is the weight between the forward L-th layer output gate and h; is the weight between the forward L-th layer output gate and x; is the bias of the forward L-th layer output gate; To filter the input information before t time in the forward Lth layer, the hyperbolic tangent here can realize the scaling of input information and can be regarded as a kind of short-term memory; is the weight between the forward L-th layer short-term memory and the state h; is the weight between the forward L-th layer short-term memory and the input x; is the bias in the forward L-th layer short-term memory; is the state of the memory unit in the forward Lth layer at time t; is the state of the memory unit in the forward Lth layer at time t-1; is the output of the recurrent layer state of the forward memory unit at the Lth layer at time t.

[0074] The LSTM neuron of the Lth layer at time t in the backward network layer of the bidirectional LSTM network model is:

[0075]

[0076] in, It is the backward forget gate of the Lth layer; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the backward network; is the weight between the backward L-th layer forget gate and state h; is the weight between the backward L-th layer forget gate and the input x; is the output of the recurrent layer state of the memory unit at the L-1th layer of the backward network at time t; is the bias of the backward forget gate of the Lth layer; is the input gate of the backward Lth layer; is the weight between the input gate of the backward Lth layer and h; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the backward network; is the weight between the input gate of the backward Lth layer and x; is the bias of the backward input gate; is the backward output gate; is the weight between the backward output gate and h; is the weight between the output gate of the backward Lth layer and x; is the bias of the output gate of the backward Lth layer; The input information of the Lth layer of the backward network is filtered out before t moments, which can be regarded as a kind of short-term memory; is the weight between the backward L-th layer short-term memory and the state h; is the weight between the backward L-th layer short-term memory and the input x; is the bias in the Lth layer of short-term memory; is the state of the memory unit at the Lth layer at time t; is the state of the memory unit in the backward Lth layer at time t-1; is the output of the recurrent layer state of the memory unit at time t in the Lth layer of the backward network;

[0077] The output of the bidirectional LSTM network model is:

[0078]

[0079] Among them, y t is the output of the bidirectional LSTM network model; is the weight between the last layer of the forward hidden layer and the model output; is the output of the last layer of the forward hidden layer; is the weight between the last layer of the backward hidden layer and the model output; is the output of the last layer of the backward hidden layer; b y is the output bias of the bidirectional LSTM network model.

[0080] Step 103: obtaining current degradation data of the degraded device to be tested, inputting the data into the trained bidirectional LSTM network model, and outputting a plurality of current remaining lifespans.

[0081] Step 104: Determine the network posterior probabilities of different trained bidirectional LSTM network models using a Bayesian model according to the training data set and the current data set; the current data set includes the current degradation data and the current remaining life.

[0082] The step 104 specifically includes: determining the training test accuracy of the trained bidirectional LSTM network model according to the training data set; determining the network prior probability of the trained bidirectional LSTM network model according to the training test accuracy; using the network prior probability as the prior probability of the Bayesian model; determining the overall prediction accuracy of the trained bidirectional LSTM network model according to the current data set; determining the prediction degradation probability according to the overall prediction accuracy; based on the Bayesian model, determining the network posterior probability of the trained bidirectional LSTM network model according to the network prior probability and the prediction degradation probability.

[0083] Among them, the network prior probability is:

[0084]

[0085] Among them, P TA (M c ) is the network prior probability; M C is the trained bidirectional LSTM network model; M CThe training and testing accuracy of ; S is the number of candidate models; i is the index of the sample label; j is the index of the candidate model; N is the number of test samples; is the degradation label; Represents M c Predicted degradation data.

[0086] The predicted degradation probability is:

[0087]

[0088] Among them, P PA (X m |M c ) is the network posterior probability; X m is the sequence of the current remaining lifespan; is the overall prediction accuracy of the cth network model; is the overall prediction accuracy of the i-th network model; m is any moment; W is the length of the input sequence; k is the number of prediction steps; x j is the degradation label of the jth sample; y c，j , is the predicted value obtained by inputting the j-th sample into the c-th network model.

[0089] The network posterior probability is:

[0090]

[0091] Among them, P A (M c |X m ) is the network posterior probability.

[0092] Step 105: A plurality of the current remaining lifetimes are fused according to the network posterior probability to predict the final remaining lifetime of the degraded device to be tested under the zero lifetime label.

[0093] Figure 2 This is a structural diagram of the system for predicting the remaining life of degraded equipment under the zero-life label provided by the present invention, such as Figure 2 As shown, a system for predicting the remaining life of degraded equipment under a zero-life label includes:

[0094] The bidirectional LSTM network model determination module 201 is used to set different network parameters to obtain bidirectional LSTM network models under different zero-lifetime labels; the network parameters include network structure, sample batch size, learning rate and number of training iterations.

[0095] The training module 202 is used to construct a training data set by a sliding window method according to the time series under the zero-life label, train different bidirectional LSTM network models, and obtain multiple trained bidirectional LSTM network models; the training data set includes historical degradation data of the degraded equipment at different times and the historical remaining life of the degraded equipment corresponding to the historical degradation data at different times.

[0096] The current remaining life output module 203 is used to obtain the current degradation data of the degraded equipment to be tested, input it into the trained bidirectional LSTM network model, and output multiple current remaining lives.

[0097] The network posterior probability determination module 204 is used to determine the network posterior probabilities of different trained bidirectional LSTM network models using a Bayesian model based on the training data set and the current data set; the current data set includes the current degradation data and the current remaining life.

[0098] The final remaining life prediction module 205 is used to fuse multiple current remaining lives according to the network posterior probability to predict the final remaining life of the degraded device to be tested under the zero life label.

[0099] The LSTM neurons in the Lth layer at time t in the forward network layer of the bidirectional LSTM network model are:

[0100]

[0101] in, is the forward L-th layer forget gate; σ is the Sigmoid activation function; is the weight between the forward L-th layer forget gate and state h; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the forward network; is the weight between the forward L-th layer forget gate and the input x; is the output of the recurrent layer state of the memory unit at the L-1th layer of the forward network at time t; is the bias of the forward L-th layer forget gate; is the forward L-th layer input gate; is the weight between the forward L-th layer input gate and h; is the weight between the forward L-th layer input gate and x; is the bias of the forward L-th layer input gate; is the forward L-th layer output gate; is the weight between the forward L-th layer output gate and h; is the weight between the forward L-th layer output gate and x; is the bias of the forward L-th layer output gate; To filter the input information before t time in the forward Lth layer, the hyperbolic tangent here can realize the scaling of input information and can be regarded as a kind of short-term memory; is the weight between the forward L-th layer short-term memory and the state h; is the weight between the forward L-th layer short-term memory and the input x; is the bias in the forward L-th layer short-term memory; is the state of the memory unit in the forward Lth layer at time t; is the state of the memory unit in the forward Lth layer at time t-1; is the output of the recurrent layer state of the forward memory unit at the Lth layer at time t.

[0102] The LSTM neuron of the Lth layer at time t in the backward network layer of the bidirectional LSTM network model is:

[0103]

[0104] in, It is the backward forget gate of the Lth layer; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the backward network; is the weight between the backward L-th layer forget gate and state h; is the weight between the backward L-th layer forget gate and the input x; is the output of the recurrent layer state of the memory unit at the L-1th layer of the backward network at time t; is the bias of the backward forget gate of the Lth layer; is the input gate of the backward Lth layer; is the weight between the input gate of the backward Lth layer and h; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the backward network; is the weight between the input gate of the backward Lth layer and x; is the bias of the backward input gate; is the backward output gate; is the weight between the backward output gate and h; is the weight between the output gate of the backward Lth layer and x; is the bias of the output gate of the backward Lth layer; The input information of the Lth layer of the backward network is filtered out before t moments, which can be regarded as a kind of short-term memory; is the weight between the backward L-th layer short-term memory and the state h; is the weight between the backward L-th layer short-term memory and the input x; is the bias in the Lth layer of short-term memory; is the state of the memory unit at the Lth layer at time t; is the state of the memory unit in the backward Lth layer at time t-1; is the output of the recurrent layer state of the memory unit at time t in the Lth layer of the backward network;

[0105] The output of the bidirectional LSTM network model is:

[0106]

[0107] Among them, y t is the output of the bidirectional LSTM network model; is the weight between the last layer of the forward hidden layer and the model output; is the output of the last layer of the forward hidden layer; is the weight between the last layer of the backward hidden layer and the model output; is the output of the last layer of the backward hidden layer; b y is the output bias of the bidirectional LSTM network model.

[0108] The network posterior probability determination module 204 specifically includes: a training test accuracy determination unit, which is used to determine the training test accuracy of the trained bidirectional LSTM network model according to the training data set; a network prior probability determination unit, which is used to determine the network prior probability of the trained bidirectional LSTM network model according to the training test accuracy; the network prior probability is used as the prior probability of the Bayesian model; an overall prediction accuracy determination unit, which is used to determine the overall prediction accuracy of the trained bidirectional LSTM network model according to the current data set; a prediction degradation probability determination unit, which is used to determine the prediction degradation probability according to the overall prediction accuracy; a network posterior probability determination unit, which is used to determine the network posterior probability of the trained bidirectional LSTM network model based on the Bayesian model, according to the network prior probability and the prediction degradation probability.

[0109] The network prior probability is:

[0110]

[0111] Among them, P TA (M c ) is the network prior probability; M C is the trained bidirectional LSTM network model; M C The training and testing accuracy of ; S is the number of candidate models; i is the index of the sample label; j is the index of the candidate model; N is the number of test samples; is the historical degradation label; Represents M c Historical degradation data for predictions.

[0112] In practical applications, the present invention replaces the Sigma or Tanh neurons in the traditional RNN with LSTM neurons to construct an LSTM neural network. The output of the L-1 layer at time t in the recursive layer of the constructed bidirectional LSTM neural network is And this output will be assigned to the neuron input in layer L at time t Therefore, the bidirectional LSTM network has only the most basic input-output connection relationship in the depth dimension; in the time dimension, the recursive connection relationship only occurs between neurons in the same layer. The structure of the forward network layer in the bidirectional LSTM network is exactly the same as that of the stacked LSTM network. The calculation of the network layer is from t = 1 to t = T, but the calculation sequence of the backward network layer is the opposite. It is calculated from t = T to t = 1. Based on the above analysis, the mathematical expression of the LSTM neuron in the Lth layer at time t in the forward network layer of the bidirectional LSTM network model is:

[0113]

[0114] in, is the forward L-th layer forget gate; σ is the Sigmoid activation function; is the weight between the forward L-th layer forget gate and state h; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the forward network; is the weight between the forward L-th layer forget gate and the input x; is the output of the recurrent layer state of the memory unit at the L-1th layer of the forward network at time t; is the bias of the forward L-th layer forget gate; is the forward L-th layer input gate; is the weight between the forward L-th layer input gate and h; is the weight between the forward L-th layer input gate and x; is the bias of the forward L-th layer input gate; is the forward L-th layer output gate; is the weight between the forward L-th layer output gate and h; is the weight between the forward L-th layer output gate and x; is the bias of the forward L-th layer output gate; To filter the input information before t time in the forward Lth layer, the hyperbolic tangent here can realize the scaling of input information and can be regarded as a kind of short-term memory; is the weight between the forward L-th layer short-term memory and the state h; is the weight between the forward L-th layer short-term memory and the input x; is the bias in the forward L-th layer short-term memory; is the state of the memory unit in the forward Lth layer at time t; is the state of the memory unit in the forward Lth layer at time t-1; is the output of the recurrent layer state of the forward memory unit at the Lth layer at time t.

[0115] At time t, the mathematical expression of the LSTM neuron in the Lth layer of the backward network is

[0116]

[0117] in, It is the backward forget gate of the Lth layer; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the backward network; is the weight between the backward L-th layer forget gate and state h; is the weight between the backward L-th layer forget gate and the input x; is the output of the recurrent layer state of the memory unit at the L-1th layer of the backward network at time t; is the bias of the backward forget gate of the Lth layer; is the input gate of the backward Lth layer; is the weight between the input gate of the backward Lth layer and h; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the backward network; is the weight between the input gate of the backward Lth layer and x; is the bias of the backward input gate; is the backward output gate; is the weight between the backward output gate and h; is the weight between the output gate of the backward Lth layer and x; is the bias of the output gate of the backward Lth layer; The input information of the Lth layer of the backward network is filtered out before t moments, which can be regarded as a kind of short-term memory; is the weight between the backward L-th layer short-term memory and the state h; is the weight between the backward L-th layer short-term memory and the input x; is the bias in the Lth layer of short-term memory; is the state of the memory unit at the Lth layer at time t; is the state of the memory unit in the backward Lth layer at time t-1; is the output of the recurrent layer state of the memory unit at time t in the Lth layer of the backward network;

[0118] The output calculation of the bidirectional LSTM network model is as follows

[0119]

[0120] in, and They are the outputs of the last layer of the forward hidden layer and the backward hidden layer, respectively. The determination of the internal weight coefficients of the bidirectional LSTM network needs to be achieved through a scientific and reasonable network training process. Usually, in order to ensure that the neural network can fully learn the degradation law contained in the training data set, it is necessary to fully ensure that the constructed bidirectional LSTM network has sufficient degrees of freedom, and use the normalization technology Dropout method to avoid network overfitting.

[0121] The present invention uses the remaining life prediction algorithm of the zero-life label network, which means that the RNN model after training is

[0122] y c，t+k =f c (x t-W+1 , x t-W+2 , ..., x t-1 , x t ) (4)

[0123] Where W and k represent the length of the input sequence and the number of prediction steps, respectively, and x t is the true degradation value, which is the input of the model at time t, y c,t+k For model M c The predicted value at time t. Due to the limitation of the model input length, the model can only predict the system degradation trajectory from time t = W, when the corresponding input and output are [x1, x2, ..., x W ] and y c，W+k . Let RUL c，m represents the remaining life obtained by system degradation prediction at time t = m, X m =[x1, x2, …, x m ] represents the actual degradation value from t=1 to time m.

[0124] In the remaining life prediction algorithm of the zero-life label network, when the interval between the failure time point and the prediction starting point exceeds the prediction step number k, the zero-life label network will use the predicted degradation value to continue iteratively predicting system degradation. Therefore, the remaining life prediction result RUL at t = m c，mIt is directly affected by the degradation prediction results at t = m and its subsequent points. The higher the accuracy of the degradation prediction results at t = m and its subsequent points, the more accurate the corresponding remaining life prediction results are. In general, when the network training parameters are set differently, the global prediction accuracy of the zero-life label network obtained by training is basically approximately equal. However, since the training algorithm usually only focuses on the global accuracy rather than the local accuracy, the degradation prediction values ​​obtained under different training conditions may have relatively large errors at some local prediction points. This shows that the uncertainty of the zero-life label network caused by different training conditions will cause uncertainty in local degradation prediction, which will further cause uncertainty in the remaining life prediction results. This means that when a network trained with degradation value labels is used to predict the remaining life of the system, a small degradation difference can cause a large life prediction error. In order to obtain a more accurate remaining life prediction result, the uncertainty of the zero-life label network is fully considered.

[0125] The present invention applies the zero-lifetime label network model averaging method to the remaining life prediction algorithm to deal with network model uncertainty. Due to the complexity of the mathematical expression of the network model, it is usually difficult to obtain the analytical form of its likelihood function, which makes it difficult to directly apply the Bayesian model averaging method to deal with the uncertainty of the zero-lifetime label network model. Therefore, excluding the case of network overfitting, the network that shows higher test accuracy on the training data set usually has better prediction ability, and then the model prior probability in the Bayesian model average is replaced by the network model training test accuracy. Let Represents the candidate network model M c The training test accuracy is expressed as

[0126]

[0127] Where N is the number of test samples, represents the real observation degradation, Representative network M c The degradation prediction of the observed degradation data in formula (5) is positive, then Considering that the model prior probability should be positively correlated with the test accuracy, the following model prior probability calculation formula is given:

[0128]

[0129] make Represents the candidate network model M c The overall prediction accuracy at time t = m can be expressed as

[0130]

[0131] Similar to the training and testing accuracy, the prediction accuracy always satisfies When the prediction accuracy Higher indicates Model M c The prediction performance from time t = 1 to t = m is better, which should correspond to a higher probability of prediction degradation, so the following expression is given

[0132]

[0133] Based on the probability functions obtained from the training test accuracy and prediction accuracy in formula (6) and formula (8), the following network posterior probability formula can be constructed according to the Bayesian model average theory:

[0134]

[0135] Due to the prediction accuracy will be degraded according to the true value of the new observation x m Continuously updated, so the posterior probability P of the network model defined A (M c |X m ) can also be updated in real time. Therefore, the model uncertainty of the zero-life label network in the remaining life prediction can be handled by the following formula

[0136]

[0137] Taking lithium batteries as an example, the present invention mainly uses the lithium battery experimental data published by the University of Maryland to verify the effectiveness and superiority of the proposed method. The discharge capacity of the batteries numbered CS2-35, CS2-36, CS2-37 and CS2-38 in the battery degradation data set is selected as the degradation feature. Table 2 shows the battery parameters in the degradation data set. The rated capacitance of the CS2 battery is 1.10Ah. It can be considered that the initial capacitance of all CS2 series batteries should not exceed 1.21Ah (110% rated capacitance).

[0138] Table 1 Degraded battery samples and their parameters

[0139]

[0140] After a series of charge and discharge test cycles, the above CS2 battery can be obtained as follows: Figure 3 The battery discharge capacity change trajectory shown in the figure is Figure 3 This is a diagram of the degradation trajectory of the battery discharge capacitance provided by the present invention.

[0141] In order to improve the accuracy of network training, Figure 3The battery degradation characteristics are normalized first. According to the rated capacity of the battery in Table 1, which is 1.10Ah, it can be considered that the initial capacity of all CS2 series batteries should not exceed 1.21Ah (110% rated capacity), otherwise the battery should be a substandard product. Therefore, the following normalization method is reasonable and effective.

[0142]

[0143] In the formula, is the degradation amount of the jth degraded battery in the ith test, is the corresponding normalized degradation amount. Further, according to Table 2, 4 different data sets are constructed from the normalized battery degradation data.

[0144] Table 2 Training data set table

[0145]

[0146]

[0147] The training sample CS2-35 (0~ω%) in Table 2 represents the CS2-35 battery in the test cycle 0>t≤T CS2-35 The monitored degradation in ω / 100, where T CS2-35 is the total number of tests for battery CS2-35; the corresponding predicted sample CS2-35 (w%~100%) is the test cycle T CS2-35 ·ω / 100<t≤T CS2-35 At the same time, according to the proportion of each battery sample data volume in the test data set, 200 samples in the training data set are randomly selected as the training test data set.

[0148] In order to verify the performance of the bidirectional LSTM recurrent network in predicting the degradation of complex systems, different types of recurrent neural networks in Table 3 were constructed, where the network training hyperparameters were set as follows: learning rate LR = 0.001, sample batch size BS = 200, number of training iterations ET = 1000, and the abandonment rate of the deep-dimensional Dropout method was 50%. The network training test accuracy and degradation prediction accuracy shown in Table 3 were obtained.

[0149] Table 3 Training test accuracy and degradation prediction accuracy of different types of recursive networks

[0150]

[0151]

[0152] In order to verify the effectiveness of the network model averaging method, the bidirectional LSTM network is used to further analyze the lithium battery discharge capacitance degradation trajectory results and remaining life results. Considering that the uncertainty of the zero-life label network model is mainly caused by the sample batch size, learning rate, number of training iterations and network structure, Table 4 sets nine different hyperparameter combinations, i.e., nine candidate bidirectional LSTM networks.

[0153] Table 4 Candidate network training parameter setting table

[0154]

[0155] According to the above parameter settings, the deep weighted Dropout method with a 50% abstention rate is adopted to obtain the training results under different data sets, as shown in Table 5.

[0156] Table 5 Training and test MSE of candidate networks under different data sets and the corresponding average prior probability of network models

[0157]

[0158]

[0159] Due to the particularity of the zero-life tag network, the remaining life prediction of lithium batteries needs to be carried out on the basis of the prediction of the discharge capacitor degradation trajectory. Therefore, the prediction results of the lithium battery discharge capacitor degradation trajectory are first given as shown in Table 6. At the same time, the arithmetic average method is used as a comparison algorithm for the average method of the network model in this chapter in Table 6.

[0160] Table 6 Overall MSE table of lithium battery discharge capacitance degradation trajectory prediction

[0161]

[0162] Order SP i，j represents the jth degradation observation point corresponding to data set i. According to the total length of degradation data of battery CS2-35, SP is selected i，1 =100, SP i，2 =300, SP i，3 =500 and SP i，4 =700, the corresponding remaining life prediction results at the four monitoring points are used to verify the effectiveness of the algorithm in this chapter. The failure threshold is set to 0.5, and the failure time of battery CS2-35 is t=802. Table 7 shows the remaining life prediction results under different data sets. The values ​​in brackets are the corresponding prediction errors. The optimal prediction results of a single network at each prediction point under each data set are marked in bold font.

[0163] Table 7 Remaining life prediction results and average prediction accuracy of each monitoring point under different data sets

[0164]

[0165]

[0166] The prediction accuracy calculation formula in Table 7 is as follows

[0167]

[0168] Where RUL Prediction is the predicted remaining useful life, RUL Real Represents the true remaining life. At the same time, the remaining life prediction results obtained at each monitoring point under different data sets all meet |RUL Prediction -RUL Real |<RUL Real , so the remaining life prediction accuracy satisfies 0≤RUL Accuracy ≤1, so the higher the prediction accuracy, the more accurate the prediction results.

[0169] 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 system 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.

[0170] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. 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 limiting the present invention.

Claims

1. A method for predicting the remaining life of degraded equipment under zero-life label, characterized in that: include: Setting different network parameters to obtain bidirectional LSTM network models under different zero-lifetime labels; the network parameters include network structure, sample batch size, learning rate, and number of training iterations; According to the time series under the zero-life label, a training data set is constructed by a sliding window method, and different bidirectional LSTM network models are trained to obtain multiple trained bidirectional LSTM network models; the training data set includes historical degradation data of the degraded equipment at different times and the historical remaining life of the degraded equipment corresponding to the historical degradation data at different times; Obtaining current degradation data of the degraded device to be tested, inputting the data into the trained bidirectional LSTM network model, and outputting a plurality of current remaining lifespans; According to the training data set and the current data set, the network posterior probabilities of different trained bidirectional LSTM network models are determined using a Bayesian model, specifically including: Determine the training test accuracy of the trained bidirectional LSTM network model according to the training data set; Determine the network prior probability of the trained bidirectional LSTM network model according to the training test accuracy; the network prior probability is used as the prior probability of the Bayesian model; Determine the overall prediction accuracy of the trained bidirectional LSTM network model according to the current data set; determining a prediction degradation probability according to the overall prediction accuracy; Based on the Bayesian model, determining the network posterior probability of the trained bidirectional LSTM network model according to the network prior probability and the predicted degradation probability; the current data set includes the current degradation data and the current remaining life; The multiple current remaining lifespans are fused according to the network posterior probability to predict the final remaining lifespan of the degraded device to be tested under the zero lifespan label.

2. The method for predicting the remaining life of degraded equipment under zero-lifetime label according to claim 1 is characterized in that: The LSTM neurons in the Lth layer at time t in the forward network layer of the bidirectional LSTM network model are: in, is the forward L-th layer forget gate; σ is the Sigmoid activation function; is the weight between the forward L-th layer forget gate and state h; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the forward network; is the weight between the forward L-th layer forget gate and the input x; is the output of the recurrent layer state of the memory unit at the L-1th layer of the forward network at time t; is the bias of the forward forget gate of the Lth layer; is the forward L-th layer input gate; is the weight between the forward L-th layer input gate and h; is the weight between the forward L-th layer input gate and x; is the bias of the forward L-th layer input gate; is the forward L-th layer output gate; is the weight between the forward L-th layer output gate and h; is the weight between the forward L-th layer output gate and x; is the bias of the forward L-th layer output gate; To filter the input information before t time in the forward Lth layer, the hyperbolic tangent here can realize the scaling of input information and can be regarded as a kind of short-term memory; is the weight between the forward L-th layer short-term memory and the state h; is the weight between the forward L-th layer short-term memory and the input x; is the bias in the forward L-th layer short-term memory; is the state of the memory unit in the forward Lth layer at time t; is the state of the memory unit in the forward Lth layer at time t-1; is the output of the recurrent layer state of the forward memory unit at the Lth layer at time t; The LSTM neurons of the Lth layer at time t in the backward network layer of the bidirectional LSTM network model are: in, It is the backward forget gate of the Lth layer; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the backward network; is the weight between the backward L-th layer forget gate and state h; is the weight between the backward L-th layer forget gate and the input x; is the output of the recurrent layer state of the memory unit at the L-1th layer of the backward network at time t; is the bias of the backward forget gate of the Lth layer; is the input gate of the backward Lth layer; is the weight between the input gate of the backward Lth layer and h; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the backward network; is the weight between the input gate of the backward Lth layer and x; is the bias of the backward input gate; is the backward output gate; is the weight between the backward output gate and h; is the weight between the output gate of the backward Lth layer and x; is the bias of the output gate of the backward Lth layer; The input information of the Lth layer of the backward network is filtered out before t moments, which can be regarded as a kind of short-term memory; is the weight between the backward L-th layer short-term memory and the state h; is the weight between the backward L-th layer short-term memory and the input x; is the bias in the Lth layer of short-term memory; is the state of the memory unit at the Lth layer at time t; is the state of the memory unit in the backward Lth layer at time t-1; is the output of the recurrent layer state of the memory unit at the Lth layer of the backward network at time t; the output of the bidirectional LSTM network model is: Among them, y t is the output of the bidirectional LSTM network model; is the weight between the last layer of the forward hidden layer and the model output; is the output of the last layer of the forward hidden layer; is the weight between the last layer of the backward hidden layer and the model output; is the output of the last layer of the backward hidden layer; b y is the output bias of the bidirectional LSTM network model.

3. The method for predicting the remaining life of degraded equipment under zero-lifetime label according to claim 1 is characterized in that: The network prior probability is: Among them, P TA (M c ) is the network prior probability; M C is the trained bidirectional LSTM network model; M C The training and testing accuracy of ; S is the number of candidate models; i is the index of the sample label; j is the index of the candidate model; N is the number of test samples; is the historical degradation data; Represents M c Predicted degradation data.

4. The method for predicting the remaining life of degraded equipment under zero-lifetime label according to claim 3 is characterized in that: The predicted degradation probability is: Among them, P PA (X m |M c ) is the network posterior probability; X m is the sequence of the current remaining lifespan; is the overall prediction accuracy of the cth network model; is the overall prediction accuracy of the i-th network model; m is any moment; W is the length of the input sequence; k is the number of prediction steps; x j is the true degradation label of the jth sample; y c，j The predicted value obtained by inputting the j-th sample into the c-th network model.

5. The method for predicting the remaining life of degraded equipment under zero-lifetime label according to claim 4 is characterized in that: The network posterior probability is: Among them, P A (M c |X m ) is the network posterior probability.

6. A system for predicting the remaining life of degraded equipment under zero-life label, characterized in that: include: A bidirectional LSTM network model determination module is used to set different network parameters to obtain bidirectional LSTM network models under different zero-lifetime labels; the network parameters include network structure, sample batch size, learning rate and number of training iterations; A training module is used to construct a training data set by a sliding window method according to a time series under a zero-lifetime label, train different bidirectional LSTM network models, and obtain multiple trained bidirectional LSTM network models; the training data set includes historical degradation data of the degraded equipment at different times and historical remaining life of the degraded equipment corresponding to the historical degradation data at different times; A current remaining life output module is used to obtain current degradation data of the degraded equipment to be tested, input it into the trained bidirectional LSTM network model, and output multiple current remaining lives; A network posterior probability determination module, used to determine the network posterior probabilities of different trained bidirectional LSTM network models using a Bayesian model according to the training data set and the current data set; The current data set includes the current degradation data and the current remaining life; The network posterior probability determination module specifically includes: A training and testing accuracy determination unit, used to determine the training and testing accuracy of the trained bidirectional LSTM network model according to the training data set; A network prior probability determination unit, used to determine the network prior probability of the trained bidirectional LSTM network model according to the training test accuracy; the network prior probability is used as the prior probability of the Bayesian model; An overall prediction accuracy determination unit, used to determine the overall prediction accuracy of the trained bidirectional LSTM network model according to the current data set; A prediction degradation probability determination unit, configured to determine the prediction degradation probability according to the overall prediction accuracy; A network posterior probability determination unit, used to determine the network posterior probability of the trained bidirectional LSTM network model based on the Bayesian model, according to the network prior probability and the predicted degradation probability; The final remaining life prediction module is used to fuse multiple current remaining lifespans according to the network posterior probability to predict the final remaining lifespan of the degraded device to be tested under the zero lifespan label.

7. The system for predicting the remaining life of degraded equipment under zero-lifetime label according to claim 6, characterized in that: The LSTM neurons in the Lth layer at time t in the forward network layer of the bidirectional LSTM network model are: in, is the forward L-th layer forget gate; σ is the Sigmoid activation function; is the weight between the forward L-th layer forget gate and state h; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the forward network; is the weight between the forward L-th layer forget gate and the input x; is the output of the recurrent layer state of the memory unit at the L-1th layer of the forward network at time t; is the bias of the forward forget gate of the Lth layer; is the forward L-th layer input gate; is the weight between the forward L-th layer input gate and h; is the weight between the forward L-th layer input gate and x; is the bias of the forward L-th layer input gate; is the forward L-th layer output gate; is the weight between the forward L-th layer output gate and h; is the weight between the forward L-th layer output gate and x; is the bias of the forward L-th layer output gate; To filter the input information before t time in the forward Lth layer, the hyperbolic tangent here can realize the scaling of input information and can be regarded as a kind of short-term memory; is the weight between the forward L-th layer short-term memory and the state h; is the weight between the forward L-th layer short-term memory and the input x; is the bias in the forward L-th layer short-term memory; is the state of the memory unit in the forward Lth layer at time t; is the state of the memory unit in the forward Lth layer at time t-1; is the output of the recurrent layer state of the forward memory unit at the Lth layer at time t; The LSTM neurons of the Lth layer at time t in the backward network layer of the bidirectional LSTM network model are: in, It is the backward forget gate of the Lth layer; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the backward network; is the weight between the backward L-th layer forget gate and state h; is the weight between the backward L-th layer forget gate and the input x; is the output of the recurrent layer state of the memory unit at the L-1th layer of the backward network at time t; is the bias of the backward forget gate of the Lth layer; is the input gate of the backward Lth layer; is the weight between the input gate of the backward Lth layer and h; is the output of the recurrent layer state of the memory unit at time t-1 of the Lth layer of the backward network; is the weight between the input gate of the backward Lth layer and x; is the bias of the backward input gate; is the backward output gate; is the weight between the backward output gate and h; is the weight between the output gate of the backward Lth layer and x; is the bias of the output gate of the backward Lth layer; The input information of the Lth layer of the backward network is filtered out before t moments, which can be regarded as a kind of short-term memory; is the weight between the backward L-th layer short-term memory and the state h; is the weight between the backward L-th layer short-term memory and the input x; is the bias in the Lth layer of short-term memory; is the state of the memory unit at the Lth layer at time t; is the state of the memory unit in the backward Lth layer at time t-1; is the output of the recurrent layer state of the memory unit at time t in the Lth layer of the backward network; The output of the bidirectional LSTM network model is: Among them, y t is the output of the bidirectional LSTM network model; is the weight between the last layer of the forward hidden layer and the model output; is the output of the last layer of the forward hidden layer; is the weight between the last layer of the backward hidden layer and the model output; is the output of the last layer of the backward hidden layer; b y is the output bias of the bidirectional LSTM network model.

8. The system for predicting the remaining life of degraded equipment under zero-life label according to claim 7, characterized in that: The network prior probability is: Among them, P TA (M c ) is the network prior probability; M C is the trained bidirectional LSTM network model; M C The training and testing accuracy of ; S is the number of candidate models; i is the index of the sample label; j is the index of the candidate model; N is the number of test samples; is the historical degradation data; Represents M c Predicted degradation value.

Images