Failure prediction model construction method, system and device and storage medium

Abstract

The invention discloses a failure prediction model construction method, system and device and a storage medium, and the key points of the technical scheme are that the method comprises the steps: simulating the degradation data of a mechanical part based on a degradation law, and obtaining a training set according to the degradation data; constructing a failure prediction model based on a cumulative neural network, and determining an initial weight, an error function and an activation function of the failure prediction model; and training a failure prediction model by adopting the training set and an NBN algorithm, calculating weighted input of each neuron in the training process, calculating output of each neuron according to the activation function and the weighted input of each neuron, updating weight according to the error function, and obtaining the failure prediction model after training is completed. According to the method, the training speed is higher, memory consumption is reduced, and high precision and high prediction success rate are achieved.

Description

Technical Field

[0001] The present invention belongs to the field of failure prediction model construction, and specifically relates to a failure prediction model construction method, system, device and storage medium. Background Art

[0002] Artificial neural networks (ANN) are the most commonly used AI algorithm for predicting the remaining useful life (RUL) of machinery. The prediction results are a key source of data that can be incorporated into maintenance management systems.

[0003] Artificial neural networks may perform well in RUL prediction of complex systems, and various types of artificial neural networks, such as recurrent neural networks (RNNs), fuzzy neural networks (FNNs), and deep convolutional neural networks, have been used for machinery RUL estimation and health prediction.

[0004] When training models composed of various neural networks, the algorithms used are usually the Error Back Propagation (EBP) algorithm and the Levenberg-Marquardt (LM) algorithm. However, the EBP algorithm has a low learning speed and low prediction accuracy. The LM algorithm is more efficient and faster than the EBP algorithm. However, the LM algorithm can only train the multi-layer perceptron (MLP) network architecture and solve a limited number of input patterns. Summary of the invention

[0005] The purpose of the present invention is to provide a failure prediction model construction method, system, device and storage medium, which can train faster, reduce memory consumption, and have high accuracy and high prediction success rate.

[0006] A first aspect of the present invention provides a method for constructing a failure prediction model, comprising:

[0007] Simulating degradation data of a mechanical component based on a degradation law, and obtaining a training set according to the degradation data;

[0008] Building a failure prediction model based on a cumulative neural network, and determining initial weights, error functions, and activation functions of the failure prediction model;

[0009] The failure prediction model is trained using the training set and the NBN algorithm. During the training process, the weighted input of each neuron is calculated, the output of each neuron is calculated based on the activation function and the weighted input of each neuron, the weight is updated based on the error function, and the failure prediction model after training is obtained.

[0010] In some embodiments, the calculating the weighted input of each neuron includes:

[0011] The weighted input calculation formula is used to calculate the weighted input of each neuron, and the weighted input calculation formula is:

[0012]

[0013] Among them, the net i is the weighted input of the ith neuron, is the number of internal inputs to the ith neuron, is the number of external inputs to the ith neuron, the w Yn(i,j) is the weight of the i-th neuron with the j-th internal input, y IY(i,j) is the jth internal input of the ith neuron, W Xn(i,j) is the weight of the i-th neuron with the j-th external input, w Bn(i) is the bias weight of the ith neuron, d n-Δ(j) is the degradation value of the ith neuron under the nth training sample, x n-Δ(j) is the jth external input under the nth training sample, no is the cardinality of o, and o is the index set of the output neuron.

[0014] In some embodiments, updating the weight according to the error function comprises:

[0015] The weight update formula is used to update the weight, and the weight update formula is:

[0016]

[0017] Among them, the w * is the updated weight, w is the weight before update, N is the number of training samples in the training set, I is the number of neuron outputs, H n,i is the quasi-Hessian matrix created by the nth training sample and the output of the i-th neuron, μ is the combination coefficient, E is the identity matrix, and G is the gradient vector of the error function.

[0018] In some embodiments, the calculating the output of each neuron according to the activation function and the weighted input of each neuron includes:

[0019] The weighted input of each neuron is substituted into the activation function to calculate the output of each neuron, and the activation function adopts the hyperbolic tangent function.

[0020] In some embodiments, simulating degradation data of a mechanical component based on a degradation law includes:

[0021] According to the average degradation value of the mechanical parts, the corresponding degradation law is selected to calculate the degradation data;

[0022] In the case where the average degradation value of a mechanical component decreases over time and becomes a constant value after a preset time, the degradation law is:

[0023]

[0024] In the case where the average degradation value of a mechanical component is a nonlinear function, the degradation mean law is:

[0025]

[0026] Among them, the represents the average degradation value, t represents the time, d0 represents the final degradation value, α i represents the degradation parameter.

[0027] In some embodiments, simulating degradation data of a mechanical component based on a degradation law includes:

[0028] In the case where the average degradation value of a mechanical component is uncertain, the degradation law is:

[0029] D(t)=σ 2 γ(t),

[0030]

[0031] Wherein, t represents time, D(t) represents the simulated degradation value at time t, represents the average degradation value, γ(t) represents the value with a scale parameter of 1 and a shape parameter of The gamma distribution, σ 2 represents variance;

[0032] The average degradation value and the simulated degradation value are calculated according to the degradation law.

[0033] In some embodiments, obtaining a training set according to the degradation data comprises:

[0034] Filtering each degradation cycle in the degradation data to obtain degradation valid data, wherein the degradation valid data includes: discontinuous constant data points and a preset number of constant data points;

[0035] The degraded valid data is used as a training set.

[0036] A second aspect of the present invention provides a failure prediction model construction system, comprising:

[0037] A data simulation module, used for simulating degradation data of mechanical components based on degradation laws, and obtaining a training set according to the degradation data;

[0038] A model building module, used to build a failure prediction model based on a cumulative neural network, and determine the initial weight, error function and activation function of the failure prediction model;

[0039] The model training module is used to train the failure prediction model using the training set and the NBN algorithm. During the training process, the weighted input of each neuron is calculated, the output of each neuron is calculated according to the activation function and the weighted input of each neuron, and the weight is updated according to the error function to obtain the failure prediction model after the training is completed.

[0040] A third aspect of the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0041] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the above method when executed by a processor.

[0042] The technical solution provided by the present invention has the following advantages and effects: the training set is obtained by using the degradation data of simulated mechanical parts, which can avoid long-term and high-cost degradation tests on actual mechanical parts, reduce installation costs and maintenance costs, and has higher flexibility and applicability. The failure prediction model is trained using the NBN algorithm, which can process any connected neural network, has a faster training speed, can reduce memory consumption, has high accuracy and high prediction success rate, and can maintain the monotonic trend of the degradation path. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the process of the failure prediction model construction method provided by the present invention;

[0044] Figure 2 is a schematic diagram of simulated degradation data provided by the present invention;

[0045] Figure 3 is a schematic diagram of data points and degradation values ​​provided by the present invention;

[0046] Figure 4 is a schematic diagram of the failure point and the remaining service life provided by the present invention;

[0047] Figure 5 It is a schematic diagram of the network architecture of the original ANN-NBN model provided by the present invention;

[0048] Figure 6 It is a schematic diagram of the network architecture of the cumulative ANN-NBN model provided by the present invention;

[0049] Figure 7It is a structural block diagram of the failure prediction model building system provided by the present invention;

[0050] Figure 8 It is a diagram of the internal structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to facilitate the understanding of the present invention, specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings.

[0052] Unless otherwise specified or defined, the "first, second..." used in this article is merely used to distinguish names and does not represent a specific quantity or order.

[0053] Unless specifically stated or defined otherwise, the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0054] It should be noted that, in this document, “fixed to” or “connected to” may mean directly fixing or connecting to an element, or indirectly fixing or connecting to an element.

[0055] like Figure 1 As shown, this embodiment provides a method for constructing a failure prediction model, including the following steps S1 to S3:

[0056] Step S1, simulating degradation data of a mechanical component based on a degradation law, and obtaining a training set according to the degradation data.

[0057] In practical applications, this application can avoid long-term and high-cost degradation tests on actual mechanical components by simulating the degradation data of mechanical components, compared with real-time collection of degradation data of mechanical components. There is no need to install sensors, build data acquisition systems, etc., thereby reducing installation and maintenance costs. The simulation parameters can be adjusted according to the actual mechanical components, which has higher flexibility and applicability.

[0058] Specifically, simulating degradation data of mechanical components based on degradation laws includes:

[0059] According to the average degradation value of the mechanical parts, the corresponding degradation law is selected to calculate the degradation data.

[0060] In the case where the average degradation value of a mechanical component decreases over time and becomes a constant value after a preset time, the degradation law is:

[0061]

[0062] In the case where the average degradation value of a mechanical component is a nonlinear function, the degradation mean law is:

[0063]

[0064] Among them, the represents the average degradation value, t represents the time, d0 represents the final degradation value, α i represents the degradation parameter.

[0065] In the case where the average degradation value of a mechanical component is uncertain, the degradation law is:

[0066] D(t)=σ 2 γ(t),

[0067]

[0068] Wherein, t represents time, D(t) represents the simulated degradation value at time t, represents the average degradation value, γ(t) represents the value with a scale parameter of 1 and a shape parameter of The gamma distribution, σ 2 represents variance;

[0069] The average degradation value and the simulated degradation value are calculated according to the degradation law.

[0070] In practical applications, different degradation laws are selected according to different average degradation values. That is to say, by simulating the degradation data of mechanical components in this application, the degradation law can be adjusted according to the degradation process of the actual mechanical components, which is more flexible and applicable. The final degradation value is the degradation value when the degradation of the mechanical component approaches the saturation point. At this time, the degradation ends, and the degradation parameter α i Adjusted according to the average degradation value of the actual mechanical component to be simulated. For the case where the average degradation value is uncertain, γ(t) is a gamma distribution with a right continuous trajectory, which is an independent increment of D(t), Δγ(t)=γ(t+Δt)-γ(t), the degradation process can be simulated. Assuming that the simulation time range is T=4 and the number of degradation points is 1000, Figure 2 As shown, Figure 2 The red curve in the graph represents the average degradation value, and the blue curve represents the simulated degradation value. In practical applications, the required failure threshold can be identified from the available information through a predefined threshold, for example, the temperature of liquid A should not exceed B°C. When the simulated degradation value reaches the final degradation value, the mechanical component will fail due to degradation. Therefore, after obtaining the degradation data, the failure point needs to be determined, the final degradation value is determined according to the failure threshold, and the failure point of the degradation data is determined according to the final degradation value.

[0071] Specifically, obtaining a training set according to the degradation data includes:

[0072] Filtering each degradation cycle in the degradation data to obtain degradation valid data, wherein the degradation valid data includes: discontinuous constant data points and a preset number of constant data points;

[0073] The degraded valid data is used as a training set.

[0074] In practical applications, after obtaining the degradation data, in order to reduce the invalid data for the training of the failure prediction model, the degradation data is first filtered to obtain the degradation valid data, which helps to remove a large number of continuous constant values. These constant values ​​are useless for the training process. For example, in a set of degradation data of {0, 0.05, 0.05, 0.05, 0.07, 0.07, 0.07, 0.07, 0.07, 0.07, 0.1, 0.1, 0.1, 0.15}, after removal, it is: {0, 0.05, 0.07, 0.07, 0.1, 0.1, 0.15}, and some continuous constants in the early stage of the degradation cycle are filtered to prevent the deletion of the highest point. The degradation level of the mechanical component is:

[0075]

[0076] When the simulated degradation value of the mechanical component reaches the failure threshold, the mechanical component is basically failed. The degradation data of the mechanical component after reaching the failure threshold can be changed to zero, and then the training of the next degradation cycle can be started. In this embodiment, the predefined failure threshold is 0.3. Figure 3 As shown in FIG. 1 , when the degradation value reaches 0.3, the mechanical component basically fails, and the degradation value 0.3 is the final degradation value.

[0077] In practical applications, degradation data can be divided into training set and test set. The training set is used to train the failure prediction model, and the test set is used to test the performance of the trained failure prediction model. The training set accounts for 80% to 90% of the degradation data. The data points that reach the failure threshold in the degradation data are taken as failure points. The failure points corresponding to the degradation data in the training set are taken as training labels, and the failure points corresponding to the degradation data in the test set are taken as test labels.

[0078] Step S2: construct a failure prediction model based on the cumulative neural network, and determine the initial weight, error function and activation function of the failure prediction model.

[0079] In practical applications, failure prediction models can be constructed based on cumulative ANN (Artificial Neural Network), such as Figure 5The figure shows the original ANN network architecture, which mainly includes: input layer, hidden layer and output layer. The input layer is the starting point of the neural network, which is responsible for receiving input information from the outside world and converting it into a format that can be processed by the neural network. The hidden layer is the core part of the neural network and can be composed of multiple hidden layers. It is used for information transmission and calculation in the neural network. Each hidden layer is composed of multiple neurons. Each neuron receives input from the previous layer, performs nonlinear transformation on it, and passes the transformation result as output to the next layer. Through such a calculation process, the neural network can learn the complex features of the input data and perform pattern recognition and prediction based on these features. Figure 5 Based on the original ANN network architecture shown in the figure, we make improvements to obtain Figure 6 The improved cumulative ANN network architecture shown in this application is based on Figure 6 The network architecture shown builds a failure prediction model, X 1(p) , X 1(p-1) Indicates external input, Y 1(p-1) , Y 1(p-2) , Y 1(p-3) , Y 2(p-1) , Y 2(p-2) Indicates internal input, Y 1(p) , Y 2(p) Indicates output, Y 1(p) , Y 2(p) Indicates a degraded value.

[0080] For the setting of initial weights, Xavier initialization is used to set the initial weights. For example, if the input dimension of a layer is a and the output dimension is b, the weight parameter of this layer should be randomly selected from a Gaussian distribution with a mean of 0 and a variance of 2 / (a+b). This can keep the variance of the activation value of each layer relatively stable, which is beneficial to the training and convergence of the model.

[0081] For the error function, the calculation formula is:

[0082]

[0083] Among them, the ε n,m represents the error between the actual output and the predicted output at the mth output port, Y n,m represents the actual output of the nth training sample at the mth output port, d n,m Represents the predicted output of the nth training sample at the mth output port. The error function is used to calculate the difference between the actual output and the predicted output, so that the weight of the model can be adjusted according to the difference.

[0084] For the activation function, by introducing nonlinearity, simulating the behavior of biological neurons, helping gradient propagation, feature detection, and stabilizing and accelerating training, the failure prediction model can learn and represent complex functional relationships, making it easier to solve complex problems.

[0085] Step S3, using the training set and the NBN algorithm to train the failure prediction model, calculating the weighted input of each neuron during the training process, calculating the output of each neuron according to the activation function and the weighted input of each neuron, updating the weight according to the error function, and obtaining the failure prediction model after training.

[0086] In practical applications, the NBN (Neuron-by-Neuron) algorithm is used to train the failure prediction model. The NBN algorithm can process arbitrarily connected neural networks, has a faster training speed, and can reduce memory consumption.

[0087] Specifically, the calculation of the weighted input of each neuron includes:

[0088] The weighted input calculation formula is used to calculate the weighted input of each neuron, and the weighted input calculation formula is:

[0089]

[0090] Among them, the net i is the weighted input of the ith neuron, is the number of internal inputs to the ith neuron, is the number of external inputs to the ith neuron, the w Yn(i,j) is the weight of the i-th neuron with the j-th internal input, y IY(i,j) is the jth internal input of the ith neuron, W Xn(i,j) is the weight of the i-th neuron with the j-th external input, w Bn(i) is the bias weight of the ith neuron, d n-Δ(j) is the degradation value of the ith neuron under the nth training sample, x n-Δ(j) is the jth external input under the nth training sample, no is the cardinality of o, and o is the index set of the output neuron.

[0091] The weighted input of each neuron is calculated through the weighted input calculation formula, and then the output of each neuron is calculated according to the weighted input and the activation function. The error between the output of the neuron and the actual output is calculated according to the error function. The weight is updated according to the error to realize the prediction of the failure prediction model. After the convergence condition is reached, the failure prediction model after training is obtained.

[0092] Specifically, calculating the output of each neuron according to the activation function and the weighted input of each neuron includes:

[0093] The weighted input of each neuron is substituted into the activation function to calculate the output of each neuron, and the activation function adopts the hyperbolic tangent function.

[0094] In practical applications, the activation function of the hidden layer adopts the hyperbolic tangent function, and the activation function of the output layer adopts a nonlinear function, such as the hyperbolic tangent function. The use of the hyperbolic tangent function can force the output of the network of the failure prediction model to be monotonically increasing, which matches the nature of the degradation process of the mechanical parts and is suitable for predicting the failure point of the mechanical parts, that is, predicting the remaining service life of the mechanical parts.

[0095] Specifically, updating the weight according to the error function includes:

[0096] The weight update formula is used to update the weight, and the weight update formula is:

[0097]

[0098] Among them, the w * is the updated weight, w is the weight before update, N is the number of training samples in the training set, I is the number of neuron outputs, H n,i is the quasi-Hessian matrix created by the nth training sample and the output of the i-th neuron, μ is the combination coefficient, E is the identity matrix, and G is the gradient vector of the error function.

[0099] In practical applications, through the weight update formula in this application, that is, the training method of the NBN algorithm, the quasi-Hessian matrix created by the output of the nth training sample and the i-th neuron can be directly calculated without calculating and storing the Jacobian matrix, thereby reducing the amount of calculation and memory usage. The gradient vector calculation formula of the error function is:

[0100]

[0101] Among them, the J N,m represents the row vector of the Jacobian matrix, K represents the number of weights, It can be calculated by the following expression:

[0102]

[0103] Among them, j N,n,l The value can be calculated by the following formula:

[0104] If l = Yn(i,j), then

[0105] If l = Xn(i,j), then

[0106] If l = Bn(i), then

[0107] Among them, no is the cardinality of the index set of output neurons, and the parameter It can be calculated by the following expression:

[0108] δ i,i =f′ i (net i )

[0109]

[0110] Among them, f′ i () represents the derivative of the activation function. The matrix δ is triangular. The qth neuron must be located before the ith neuron. Therefore, i ≥ q, δ i,i is the slope of the activation function of the ith neuron, w Wn(i,k) is the connection weight between the ith neuron and the qth neuron under the nth training sample, δ k,q It is the signal gain obtained by the offset weight and other parts of the network connected to the offset weight. The gradient vector of the error function is calculated according to the gradient vector calculation formula. In the present invention, the quasi-Hessian matrix and the gradient vector are calculated, and then the quasi-Hessian matrix and the gradient vector are substituted into the weight update formula to calculate the updated weight, thereby adjusting the model and avoiding the direct calculation of the Jacobian matrix, so as to facilitate faster acquisition of a failure prediction model with higher accuracy.

[0111] After completing the training of the failure prediction model, the training results of the failure prediction model were obtained. The number of iterations was 578 times, and the SSE (Sum of square error) of the training set was 0.005247. Then the test set was used to test the failure prediction model to obtain the test results. The SSE of the test set was 0.006587. The slight difference between the two SSE values ​​indicated that the training error was very small and the correlation between the training data and the test data was very high. Therefore, the performance of the failure prediction model was acceptable.

[0112] To calculate the RUL (remaining useful life), subtract the failure point (ft) from the last training point, e.g. Figure 4As shown, two RULs can be defined: RUL1 is the RUL predicted by the failure prediction model based on the test set, and RUL2 is the expected RUL based on the expected degradation path of the test set. RUL1 can be calculated by subtracting the predicted failure point of the test set (ft1) from the last training point, and then RUL2 can be calculated by subtracting the expected failure point of the test set (ft2) from the last training point. The calculated RUL1 of the mechanical component is approximately equal to 197 cycles, while RUL2 is approximately equal to 192 cycles.

[0113] In order to compare the RUL prediction accuracy of the cumulative ANN-NBN model and the original ANN-NBN model, a 1000-cycle running experiment was conducted on both. The network architecture diagram of the original ANN-NBN model is shown in Figure 5 As shown, the network architecture diagram of the improved cumulative ANN-NBN model is as follows Figure 6 shown.

[0114] Table 1 lists the experimental parameters. Except for the maximum training error, all parameters are fixed in all runs. Then the RUL1 and RUL2 of the mechanical parts of each run are estimated, and then the prediction error is calculated using the prediction error formula, which is:

[0115] prediction error=|RUL2-RUL1|,

[0116] The prediction error represents a prediction error.

[0117] Table 1 Experimental parameters - Comparison between cumulative ANN-NBN and original ANN-NBN

[0118] Total number of runs / cycles 1000 Filter Points 11 Predefined failure thresholds 0.3 Maximum training error (cumulative ANN-NBN) 0.00001 Maximum training error (original ANN-NBN) 0.0007 Last training point 6323

[0119] The number of filter points in Table 1 is the number of data points of the continuous constant filtered at the beginning of each degradation cycle of the degradation data. Table 2 is the experimental results obtained after experiments on the cumulative ANN-NBN and the original ANN-NBN according to the experimental parameters listed in Table 1.

[0120] Table 2 Experimental results - Comparison between cumulative ANN-NBN and original ANN-NBN

[0121] Cumulative ANN-NBN Original ANN-NBN Run time (1000 cycles) About 10 hours About 3.5 hours SSE Prediction 78449 (number of times within the sampling period) 143802 (number of times within the sampling period) Prediction success rate 100％ 97.6％

[0122] The prediction success rate in Table 2 refers to the number of runs where the prediction algorithm successfully predicts a fault exceeding the predefined threshold of 0.3, and then estimates the RUL. The experimental results in Table 2 show that the cumulative ANN-NBN outperforms the original ANN-NBN in terms of prediction error (the smaller the prediction error, the higher the prediction accuracy), and the cumulative ANN-NBN prediction success rate is 100%, while the prediction success rate of the original ANN-BNN is 97.6%.

[0123] The present invention conducts a performance comparison analysis on the developed cumulative ANN-NBN model and other ANN models.

[0124] As shown in Table 3, we designed six ANN models (ANN 1-ANN 6) with different key parameters such as network structure, training algorithm, activation function and calculation method. From the perspective of relevant performance factors such as prediction error and prediction success rate, we studied the impact of changing these parameters on performance factors.

[0125] Table 3 Comparison between cumulative ANN-NBN and original ANN-NBN

[0126]

[0127] Note: MLP 1-10-1 represents a multilayer perceptron (MLP), with neural nodes 1-10-1 and 10 neurons in the hidden layer. Similarly, MLP 1-2-1 represents a multilayer perceptron with neural nodes 1-2-1 and 2 neurons in the hidden layer, and MLP 1-3-1 represents a multilayer perceptron with neural nodes 1-3-1 and 3 neurons in the hidden layer.

[0128] The prediction success rate in Table 3 is defined as the number of successful predictions of the failure point when the degradation exceeds a predefined threshold. Except for the first model, the training and test errors of all models are relatively small, and the prediction success rate of all models is relatively high. The best model with a success rate of 100% is ANN6 (cumulative ANN-NBN). Overall, the cumulative ANN-NBN model proposed in the present invention outperforms other comparative models in terms of fault prediction success rate and produces very small prediction errors. Comparing the ANN 4, ANN 5 and ANN 6 models, although ANN 4 has the smallest SSE and the best training performance, ANN 6 has a clear advantage in terms of prediction success rate.

[0129] The cumulative ANN-NBN network design, thanks to the improved network design and the output activation function used, guarantees a 100% prediction success rate. In summary, all NBN models performed well under different key parameters and achieved high accuracy and prediction success rate. Similar parameter changes were applied to models ANN 2 and ANN 3 in terms of the number of neurons, training algorithm, and output activation function. The results show that ANN 3 with the LM training algorithm and the “tanh” output activation function achieved the best training and test errors compared to ANN 2 with the “Identity” output activation function and the gradient descent training algorithm. Therefore, model ANN 3 outperforms model ANN 2 in terms of prediction accuracy and prediction success rate.

[0130] In order to verify the effectiveness of the developed cumulative ANN-NBN model, the prediction results of the cumulative ANN-NBN model are compared with other commonly used data-driven models, such as nonlinear regression, random forest (RF), support vector machine (SVM), and time series models, such as autoregressive integrated moving average (ARIMA) and its special case "exponential smoothing". Table 4 shows the prediction results of each model.

[0131] Table 4 Experimental results of other data-driven models

[0132]

[0133]

[0134] Table 4 shows that the RF, SVM and other models have much higher errors than the cumulative ANN-NBN model. Therefore, they are not suitable for predicting such degraded data.

[0135] The failure prediction model construction method of the present invention uses the degradation data of simulated mechanical parts to obtain a training set, which can avoid long-term and high-cost degradation tests on actual mechanical parts, reduce installation costs and maintenance costs, and has higher flexibility and applicability. The failure prediction model is trained using the NBN algorithm, which can process any connected neural network, has a faster training speed, can reduce memory consumption, has high accuracy and high prediction success rate, and can maintain the monotonic trend of the degradation path.

[0136] like Figure 7 As shown, an embodiment of the present invention further provides a failure prediction model construction system, including:

[0137] A data simulation module, used for simulating degradation data of mechanical components based on degradation laws, and obtaining a training set according to the degradation data;

[0138] A model building module, used to build a failure prediction model based on a cumulative neural network, and determine the initial weight, error function and activation function of the failure prediction model;

[0139] The model training module is used to train the failure prediction model using the training set and the NBN algorithm. During the training process, the weighted input of each neuron is calculated, the output of each neuron is calculated according to the activation function and the weighted input of each neuron, and the weight is updated according to the error function to obtain the failure prediction model after the training is completed.

[0140] In some embodiments, the model building module includes the following units (not shown):

[0141] The first calculation unit is used to calculate the weighted input of each neuron using a weighted input calculation formula, wherein the weighted input calculation formula is:

[0142]

[0143] Among them, the net i is the weighted input of the ith neuron, is the number of internal inputs to the ith neuron, is the number of external inputs to the ith neuron, the w Yn(i,j) is the weight of the i-th neuron with the j-th internal input, y IY(i,j) is the jth internal input of the ith neuron, W Xn(i,j) is the weight of the i-th neuron with the j-th external input, w Bn(i) is the bias weight of the ith neuron, d n-Δ(j),IX(i,j) is the degradation value of the ith neuron under the nth training sample, x n-Δ(j),IX(i,j) is the jth external input under the nth training sample, no is the cardinality of o, and o is the index set of the output neuron.

[0144] In some embodiments, the model building module includes the following units (not shown):

[0145] The second calculation unit is used to update the weight using a weight update formula, where the weight update formula is:

[0146]

[0147] Among them, the w * is the updated weight, w is the weight before update, N is the number of training samples in the training set, I is the number of neuron outputs, H n,i is the quasi-Hessian matrix created by the nth training sample and the output of the i-th neuron, μ is the combination coefficient, E is the identity matrix, and G is the gradient vector of the error function.

[0148] In some embodiments, the model building module includes the following units (not shown):

[0149] The third computing unit is used to substitute the weighted input of each neuron into the activation function to calculate the output of each neuron, and the activation function adopts the hyperbolic tangent function.

[0150] In some embodiments, the data simulation module includes the following units (not shown):

[0151] A law selection unit, used for selecting a corresponding degradation law according to an average degradation value of a mechanical component to calculate and obtain degradation data;

[0152] The first degradation unit is used for the case where the average degradation value of the mechanical component decreases over time and becomes a constant value after a preset time, and the degradation law is:

[0153]

[0154] The second degradation unit is used when the average degradation value of the mechanical component is a nonlinear function, and the degradation mean value law is:

[0155]

[0156] Among them, the represents the average degradation value, t represents the time, d0 represents the final degradation value, α i represents the degradation parameter.

[0157] In some embodiments, the data simulation module includes the following units (not shown):

[0158] The third degradation unit is used when the average degradation value of the mechanical component is uncertain, and the degradation law is:

[0159] D(t)=σ 2 γ(t),

[0160]

[0161] Wherein, t represents time, D(t) represents the simulated degradation value at time t, represents the average degradation value, γ(t) represents the value with a scale parameter of 1 and a shape parameter of The gamma distribution, σ 2 represents variance;

[0162] The degradation calculation unit is used to calculate the average degradation value and the simulated degradation value according to the degradation law.

[0163] In some embodiments, the data simulation module includes the following units (not shown):

[0164] Filtering each degradation cycle in the degradation data to obtain degradation valid data, wherein the degradation valid data includes: discontinuous constant data points and a preset number of constant data points;

[0165] The degraded valid data is used as a training set.

[0166] Each module of the above failure prediction model construction system can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above modules and units can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each of the above modules.

[0167] like Figure 8 As shown, an embodiment of the present invention discloses a computer device, including a memory and a processor, wherein the memory stores a computer program;

[0168] The computer device may be a server, and its internal structure diagram may be as follows: Figure 8 As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the failure prediction model construction method described in the above embodiments is implemented.

[0169] Those skilled in the art will understand that Figure 8 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0170] An embodiment of the present invention further discloses a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the failure prediction model construction method described in the above embodiments.

[0171] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0172] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for constructing a failure prediction model, characterized in that: The following steps are involved: Simulating degradation data of a mechanical component based on a degradation law, and obtaining a training set according to the degradation data; Building a failure prediction model based on a cumulative neural network, and determining initial weights, error functions, and activation functions of the failure prediction model; The failure prediction model is trained using the training set and the NBN algorithm. During the training process, the weighted input of each neuron is calculated, the output of each neuron is calculated based on the activation function and the weighted input of each neuron, the weight is updated based on the error function, and the failure prediction model after training is obtained.

2. The failure prediction model construction method according to claim 1, characterized in that: The calculation of the weighted input of each neuron includes: The weighted input calculation formula is used to calculate the weighted input of each neuron, and the weighted input calculation formula is: Among them, the net i is the weighted input of the ith neuron, is the number of internal inputs to the ith neuron, is the number of external inputs to the ith neuron, the w Yn(i,j) is the weight of the i-th neuron with the j-th internal input, y IY(i,j) is the jth internal input of the ith neuron, W Xn(i,j) is the weight of the i-th neuron with the j-th external input, w Bn(i) is the bias weight of the ith neuron, d n-Δ(j) is the degradation value of the ith neuron under the nth training sample, x n-Δ(j) is the jth external input under the nth training sample, no is the cardinality of o, and o is the index set of the output neuron.

3. The failure prediction model construction method according to claim 1, characterized in that: The updating of weights according to the error function comprises: The weight update formula is used to update the weight, and the weight update formula is: Among them, the w * is the updated weight, w is the weight before update, N is the number of training samples in the training set, I is the number of neuron outputs, H n,i is the quasi-Hessian matrix created by the nth training sample and the output of the i-th neuron, μ is the combination coefficient, E is the identity matrix, and G is the gradient vector of the error function.

4. The failure prediction model construction method according to claim 1, characterized in that: The step of calculating the output of each neuron according to the activation function and the weighted input of each neuron comprises: The weighted input of each neuron is substituted into the activation function to calculate the output of each neuron, and the activation function adopts the hyperbolic tangent function.

5. The failure prediction model construction method according to claim 1, characterized in that: The degradation data of the mechanical component simulated based on the degradation law includes: According to the average degradation value of the mechanical parts, the corresponding degradation law is selected to calculate the degradation data; In the case where the average degradation value of a mechanical component decreases over time and becomes a constant value after a preset time, the degradation law is: In the case where the average degradation value of a mechanical component is a nonlinear function, the degradation mean law is: Among them, the represents the average degradation value, t represents the time, d0 represents the final degradation value, α i represents the degradation parameter.

6. The failure prediction model construction method according to claim 1, characterized in that: The degradation data of the mechanical component simulated based on the degradation law includes: In the case where the average degradation value of a mechanical component is uncertain, the degradation law is: D(t)=σ 2 γ(t), Wherein, t represents time, D(t) represents the simulated degradation value at time t, represents the average degradation value, γ(t) represents the value with a scale parameter of 1 and a shape parameter of The gamma distribution, σ 2 represents variance; The average degradation value and the simulated degradation value are calculated according to the degradation law.

7. The failure prediction model construction method according to claim 1, characterized in that: The obtaining of a training set according to the degradation data comprises: Filtering each degradation cycle in the degradation data to obtain degradation valid data, wherein the degradation valid data includes: discontinuous constant data points and a preset number of constant data points; The degraded valid data is used as a training set.

8. A failure prediction model construction system, characterized in that: include: A data simulation module, used for simulating degradation data of mechanical components based on degradation laws, and obtaining a training set according to the degradation data; A model building module, used to build a failure prediction model based on a cumulative neural network, and determine the initial weight, error function and activation function of the failure prediction model; The model training module is used to train the failure prediction model using the training set and the NBN algorithm. During the training process, the weighted input of each neuron is calculated, the output of each neuron is calculated according to the activation function and the weighted input of each neuron, and the weight is updated according to the error function to obtain the failure prediction model after the training is completed.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images