Gearbox life prediction method and system based on NACMD-LSTM

Abstract

The present application provides a gearbox life prediction method and system based on NACMD-LSTM, which includes: S1, adding pairs of positive and negative white noises that are opposite to each other to the original signal to construct a new time series; S2, based on the NACMD algorithm, performing modal decomposition on the test sequence to obtain the first sub-mode of the test sequence, and calculating the mean value to record it as the first sub-mode of the original signal; S3, taking the residual as the new original signal, and obtaining the residual signal; S4, judging whether the residual signal can be further decomposed, if so, looping step S3 until the obtained residual signal is a monotonic function and cannot be further decomposed; if not, outputting the k sub-modes of the original signal to be decomposed; S5, obtaining all the sub-modes, superimposing and reconstructing to obtain the processed sequence to be input; S6, using the CSSA optimization algorithm to optimize the parameters of the network structure, training and constructing the CSSA-LSTM model; S7, inputting the processed sequence to be input into the CSSA-LSTM network to obtain the prediction result of the gearbox life.

Description

Technical Field

[0001] The present application relates to the technical field of fault diagnosis of rotating machinery, and in particular to a gearbox life prediction method and system based on NACMD-LSTM. Background Art

[0002] The manufacturing industry and various aircraft are increasingly required to produce with high reliability, low environmental risk and personal safety, while running their processes at maximum output. Despite considerable progress in sensing instruments, data acquisition hardware and information technology infrastructure, condition monitoring and diagnostics are still mainly limited to the most critical system components. Gearboxes are one of the basic and most important parts of rotating machinery used in industry and helicopters. Their function is to transmit torque and power from one shaft to another. Typical applications include helicopters, automobiles, metal cutting machine tools and power systems. These machines, especially helicopters, are in an operating state, and any gear failure may have serious consequences. Therefore, accurate fault diagnosis of gearboxes is key to preventing mechanical system failures or even paralysis of the entire system.

[0003] There are many types of gear failures, including wear, misalignment, cracks, fatigue, and chipping. When any of these failures develop into severe degradation, it can cause a catastrophic accident in a machine such as a helicopter. Gear wear is a slow process. When wear reaches a severe level, the gear will fall off. Ultimately, it can cause machine failure, so the diagnosis of the degree of gear wear is even more important, and it is extremely important to conduct life prediction research on gearboxes.

[0004] At present, the prediction of remaining service life of gearboxes at home and abroad is becoming more and more perfect, and various prediction methods have been proposed, which usually include physical methods, statistical methods, machine learning methods and hybrid prediction methods.

[0005] Physical methods need to take into account the impact of temperature, environmental factors, etc. on the model, and also need to accurately describe the mechanical physical properties. The calculation difficulty and cost are very high, and it is not easy to get accurate predictions. Statistical methods use the linear relationship between historical time series variables to establish models, which overcomes the shortcomings of traditional physical methods, but there are still problems in dealing with nonlinear relationships between several elements. Machine learning methods can better handle nonlinear time series data, mainly including artificial neural networks, support vector machines, and extreme learning machines. The disadvantage is that a single machine learning algorithm is prone to overfitting and local optimization problems. With the continuous development of prediction technology, the requirements for prediction accuracy are getting higher and higher. A single model has the advantages of being simple and easy to use, but its prediction ability is limited, which often leads to low robustness and weak applicability. Therefore, the hybrid prediction model gradually shows its advantages.

[0006] There are two main types of hybrid models: data decomposition and weight coefficients. The former first decomposes the data into different components, and then predicts each component separately according to the different characteristics of the algorithm. The latter assigns certain weight coefficients to several single models and combines them. The weight coefficients are determined by entropy method, optimization algorithm, etc.

[0007] The existing hybrid models using modal decomposition and deep time series prediction models have some defects. For the original samples, the collected fault signals will be affected by many complex conditions in a complex working environment, so it is impossible to accurately extract the characteristic signals for network training. Therefore, it is necessary to use the modal decomposition algorithm to remove noise and enhance the signal credibility. In the process of hybridization, grid search or gradient descent are mostly used to optimize parameters. However, in the process of using these methods to optimize parameters, there are certain limitations and it is complex and time-consuming. Summary of the invention

[0008] The present application provides a gearbox life prediction method and system based on NACMD-LSTM to solve the technical problem that the model is easily affected by data quality, resulting in reduced computing efficiency and accuracy.

[0009] The present application provides a gearbox life prediction method based on NACMD-LSTM, which is characterized by comprising the following steps:

[0010] S1, adding pairs of positive and negative white noises with opposite numbers to each other to the original signal to construct a new time series as the test sequence;

[0011] S2. Based on the NACMD algorithm, the test sequence is modally decomposed to obtain the first sub-mode of the test sequence, the mean is recorded as the first sub-mode A1(t) of the original signal, and the corresponding residual is recorded;

[0012] S3, take the residual as the new original signal, and obtain the kth residual signal r k (t);

[0013] S4. Determine the kth residual signal r k (t) Whether it can be further decomposed, if so, loop through step S3 until the kth residual signal r is obtained. k (t) is a monotonic function and cannot be further decomposed; if not, the k sub-modes of the original signal to be decomposed x(t) are output, that is:

[0014]

[0015] Where x(t) represents the original signal to be decomposed, i = 1, 2...k, k represents the number of sub-modes, A i (t) represents the i-th extracted sub-modality, rk (t) represents the kth residual signal, t represents time;

[0016] S5, obtaining all sub-modes, superimposing and reconstructing to obtain the processed sequence to be input;

[0017] S6. Use the CSSA optimization algorithm to optimize the parameters of the network structure, train and build the CSSA-LSTM model;

[0018] S7. Input the processed input sequence into the CSSA-LSTM network to obtain the prediction result of the gearbox life.

[0019] Optionally, step S2 specifically includes the following steps:

[0020] S21, obtaining the original signal to be decomposed x(t) and the number of white noises N;

[0021] S22. Obtain the test signal after adding noise Test signal Determined by the following formula:

[0022]

[0023] Where x(t) represents the original signal to be decomposed, & = 1, 2, ε represents the final standard deviation of the error, and v j (t) represents a Gaussian white noise signal that satisfies the standard normal distribution, j=1, 2...N, N represents the number of white noises, and t represents time;

[0024] S23. Test signal To break it down:

[0025]

[0026] in, Indicates the test signal, represents the first sub-mode obtained, r1(t) represents the first residual signal, ACMD(·) represents the decomposition process after ACMD algorithm, and t represents time;

[0027] S24, after processing N test signals, N different Then the first extracted sub-mode is:

[0028]

[0029] Where A1(t) represents the first extracted sub-mode, j = 1, 2...N, N represents the number of white noises, that is, the number of test signals, represents the first submodal component obtained by decomposing the jth test signal, and t represents time;

[0030] S25. Calculate the residual after removing the first extracted submode A1(t):

[0031] r1(t)=x(t)-A1(t)

[0032] Wherein, r1(t) represents the first residual signal, that is, the residual after removing the first extracted sub-mode A1(t), x(t) represents the original signal to be decomposed, A1(t) represents the first extracted sub-mode, and t represents time.

[0033] Optionally, step S3 specifically includes the following steps:

[0034] S31, adding positive and negative paired Gaussian white noise to the first residual signal r1(t) to obtain a new signal, and then using the new signal as a carrier to repeat steps S23-S24 again, thereby obtaining the kth extracted sub-mode A k (t):

[0035]

[0036] Among them, A k (t) represents the kth extracted sub-mode, j = 1, 2...N, N represents the number of white noise, that is, the number of test signals, represents the first sub-mode obtained by decomposing the j-th new signal, and t represents time;

[0037] S32, calculate and remove the k-th extracted sub-mode A k The residual after (t):

[0038] r k (t) = r k-1 (t)-A k (t)

[0039] Among them, r k (t) represents the kth residual signal, that is, removing the kth extracted sub-mode A k (t) The residual after k-1 (t) represents the k-1th residual signal, that is, removing the k-1th extracted sub-mode A k-1 (t) The residual after, A k (t) represents the kth extracted sub-modality,

[0040] Optionally, step S21 specifically includes the following steps:

[0041] S211, obtaining an original signal to be decomposed, wherein the signal model of the original signal to be decomposed is expressed as:

[0042]

[0043] Among them, x(t) represents the original signal to be decomposed, α m (t) and β m (t) represents two non-stationary signals, Represents the frequency function of the non-stationary signal factor, m = 1, 2…M, x(t) is considered to be composed of M sub-modes, s represents an integral variable, and t represents time.

[0044] Optionally, step S21 specifically further includes the following steps:

[0045] S212, obtaining the number N of white noise, which is obtained by the following formula:

[0046]

[0047] Where N represents the number of white noises, ε represents the final standard deviation of the error, and A represents the amplitude of the noise.

[0048] Optionally, step S6 specifically includes the following steps:

[0049] S61, randomly generate chaotic vector y0=[y 0，1 ,y 0,2 , …, y 0,D ], where y 0,w ∈[0, 1], D represents the dimension of the search space, w = 1, 2, ..., D;

[0050] S62, generate chaotic sequence y1, y2, ..., y G ,

[0051] y v，w =4(y v-1，w ) 3 -3y v-1，w

[0052] Among them, y v，w Represents chaotic sequences in different dimensions, y v =[y v,1 ,y v,2 , …, y v,D ], v = 1, 2, ..., G, where G represents the population size, w = 1, 2, ..., D, where D represents the dimension of the search space;

[0053] S63, map the chaotic sequence to the original search space, and obtain the initialization population X = [x1, x2, ..., x G ], where x p =[x p,1 , x p,2 , …, x p,D ],

[0054] x p,q =x min，q +y p，q ·(x max，q -x min，q )

[0055] Among them, x p，q represents the pth individual in the qth dimension in the initialized population, x min，q and x max，q They represent the minimum and maximum values ​​of the qth dimension of the search space, q = 1, 2, ..., D, where D represents the dimension of the search space;

[0056] S64, calculating the fitness value of the initial population and marking the best individual;

[0057] S65. Update current finder location

[0058]

[0059] in, Indicates the current finder location, represents the position of the discoverer in the previous iteration, T represents the current iteration number, p represents the number of individuals, q represents the dimension, iter max represents the maximum number of iterations, ∝ is a random number in [0, 1], Q is a random number that follows a normal distribution, L represents a single-row d-column matrix whose internal elements are all 1, ST represents the safety value, and R2 represents the warning value;

[0060] S66. Update the joiner's position:

[0061]

[0062] in, Indicates the current joiner position, represents the position of the joiner in the previous iteration, Indicates the optimal position of the finder at present. represents the worst position in the current population; A represents a 1×d matrix, and each element is randomly assigned 1 or -1, L is a single-row d-column matrix whose internal elements are all 1, T represents the current iteration number, p represents the number of individuals, q represents the dimension, Q is a random number that obeys the normal distribution, n represents the number of participants, A + Indicates A T (AA T ) -1 ;

[0063] S67, Update predator position:

[0064]

[0065] in, represents the updated predator position, represents the position of the predator in the current iteration, p represents the number of individuals, q represents the dimension, β represents the step size control parameter, a random number that obeys a normal distribution with a mean of 0 and a variance of 1, and K represents a random number between -1 and 1. Indicates that the sparrow is currently in the safest position. represents the worst position in the current population, f p represents the fitness value of the current sparrow individual, f great and f worst Represent the current best and worst fitness values ​​respectively, γ represents the smallest constant to avoid the denominator being 0, and T represents the current number of iterations;

[0066] S68, determine the fitness value f of the current sparrow individual p Whether the preset value is met, if so, output the optimal sparrow position, otherwise return to step S65.

[0067] Optionally, before step 61, the following steps are also included:

[0068] S601, setting the learning rate, number of iterations, and number of neurons in the second layer of the LSTM network within a certain range, and initializing parameters related to the sparrow algorithm;

[0069] S602, taking the mean square error of the test set as the sparrow fitness value and sorting them;

[0070] The CSSA-LSTM model includes an input layer, three LSTM hidden layers and an output layer.

[0071] Correspondingly, the present application also provides an electronic device, which includes a memory and a processor, the memory being used to store executable program code; the processor being connected to the memory, and running a computer program corresponding to the executable program code by reading the executable program code, so as to execute any step in the aforementioned gearbox life prediction method based on NACMD-LSTM.

[0072] Correspondingly, the present application also provides a gearbox life prediction system based on NACMD-LSTM, which includes the aforementioned electronic device.

[0073] The present application provides a method and system for predicting the life of a gearbox based on NACMD-LSTM. The NACMD signal decomposition method adds multiple groups of positive and negative Gaussian white noise signals to the original signal to form a group of test signals and extract the first sub-mode; then, the residual signal of each decomposition is used as a new input signal for noise processing again to extract the second sub-mode; the loop is iterated until the residual signal can no longer be decomposed, and all modal extractions are completed. This method has better noise resistance in a noisy environment, overcomes the problem of modal aliasing to a certain extent, has better signal decomposition capabilities, and improves the accuracy of subsequent predictions.

[0074] The CSSA optimization algorithm replaces the conventional random uniform distribution with cubic chaotic mapping for population initialization of the SSA algorithm, which enhances the diversity of the population to improve the performance of the algorithm, further avoids falling into local optimality, quickly searches for the global optimal solution, and can more efficiently solve some optimal problems.

[0075] Using the CSSA optimization algorithm in the CSSA-LSTM model to optimize the LSTM model parameters can help improve the performance and generalization ability of the model, accelerate the model convergence speed, and improve the prediction accuracy. At the same time, CSSA can also help find better parameter combinations, effectively search the parameter space, and improve the generalization performance of the model.

[0076] The hybrid model of NACMD and CSSA-LSTM decomposes the original signal into simpler sub-modes, reduces the complexity of the input signal, and then predicts them separately through the network model, and reconstructs the output results on average to obtain the final prediction value, which improves the network's computing efficiency and enhances the robustness and applicability of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0078] Figure 1 It is a flow chart of the gearbox life prediction method based on NACMD-LSTM provided in this application;

[0079] Figure 2 It is an overall flow chart of the gearbox life prediction method based on NACMD-LSTM provided in this application;

[0080] Figure 3 It is a NACMD flow chart in the gearbox life prediction method based on NACMD-LSTM provided in this application;

[0081] Figure 4 It is the CSSA algorithm framework in the gearbox life prediction method based on NACMD-LSTM provided in this application;

[0082] Figure 5 It is a schematic diagram of the LSTM structure framework in the gearbox life prediction method based on NACMD-LSTM provided in this application;

[0083] Figure 6 It is the CSSA-LSTM prediction model in the gearbox life prediction method based on NACMD-LSTM provided in this application. DETAILED DESCRIPTION

[0084] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application. In addition, it should be understood that the specific implementation methods described herein are only used to illustrate and explain the present application and are not used to limit the present application. In the present application, unless otherwise stated, the directional words such as "up", "down", "left" and "right" used generally refer to the upper, lower, left and right of the device in actual use or working state, specifically the drawing direction in the accompanying drawings. In this article, mode, sub-mode, modal component and sub-modal component have the same or similar meanings.

[0085] The present application provides a gearbox life prediction method and system based on NACMD-LSTM, which are described in detail below. It should be noted that the description order of the following embodiments does not limit the preferred order of the embodiments of the present application. In the following embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0086] See also Figure 1, the present application provides a gearbox life prediction method based on NACMD-LSTM. Aiming at the problem that the gearbox life prediction model is easily affected by data quality, resulting in reduced computing efficiency and accuracy, and the problem that the sparrow algorithm is prone to fall into the local optimal solution when optimizing LSTM parameters. A life prediction method based on NACMD-LSTM is proposed. Among them, based on the noise-assisted chirp mode decomposition (NACMD) algorithm, multiple groups of test sequences are constructed using white noise with positive and negative cancellation, and the anti-noise ability of signal decomposition is enhanced by average reconstruction, the robustness of the algorithm is improved, and more effective prediction data is provided for the life prediction model. Then, the sparrow optimization algorithm is improved by using cubic chaos mapping, which is used to optimize the parameters such as the learning rate and the number of nodes of the LSTM network. Finally, a hybrid model of mode decomposition and network structure is used to make a more accurate prediction of the remaining life of the gearbox.

[0087] See also Figure 1-Figure 6 , the NACMD algorithm is used to realize the denoising decomposition of the gearbox acceleration vibration signal. Specifically, pairs of positive and negative white noises with opposite numbers are added as auxiliary noises to the original signal to form N groups of test sequences. Since the noise added in each test is different, the sub-modes obtained in each test do not show any correlation with the corresponding sub-modes of other tests, and the redundant auxiliary white noise remaining in the reconstructed signal can be eliminated. Then, the N groups of test sequences are iteratively solved using the NACMD algorithm to obtain several sub-modes, and finally the time series after denoising is obtained by average reconstruction. Next, the cubic chaos map is used to optimize the initial distribution of the sparrow optimization algorithm to improve the algorithm's optimization ability, and then it is used to optimize the main parameters of the LSTM to establish the optimal CSSA-LSTM model. The processed prediction data is input into the CSSA-LSTM network for prediction to obtain the final prediction result. The following is the detailed process:

[0088] S1, adding pairs of positive and negative white noises with opposite numbers to each other to the original signal to construct a new time series as the test sequence;

[0089] S2. Based on the NACMD algorithm, the test sequence is modally decomposed to obtain the first sub-mode of the test sequence, the mean is recorded as the first sub-mode A1(t) of the original signal, and the corresponding residual is recorded;

[0090] In practical applications, non-stationary signals often contain multiple sub-signals, which are called Chirp signals in this paper. Each linear frequency modulation pattern can be modeled as an AM-FM signal.

[0091] Step S2 specifically includes the following steps:

[0092] S21, obtaining the original signal to be decomposed x(t) and the number of white noises N;

[0093] Step S21 specifically includes the following steps:

[0094] S211, obtaining an original signal to be decomposed, wherein the signal model of the original signal to be decomposed is expressed as:

[0095]

[0096] Among them, x(t) represents the original signal to be decomposed, α m (t) and β m (t) represents two non-stationary signals, Represents the frequency function of the non-stationary signal factor, m = 1, 2…M, x(t) is considered to be composed of M sub-modes, s represents an integral variable, and t represents time.

[0097] As mentioned above, estimating all signal modes simultaneously is difficult and may be affected by stability issues. This scheme uses a recursive framework to extract signal modes one by one. Let A i (·) is the i-th submode obtained after iterative decomposition of the original signal x(t). The i-th submode obtained by NACMD decomposition is E i (·), v j For a Gaussian white noise signal that satisfies the standard normal distribution, j=1, 2...N, N is the number of white noises, ε is the weight parameter of the white noise, and x(t) is the original signal to be decomposed.

[0098] S212, obtaining the number N of white noise, which is obtained by the following formula:

[0099]

[0100] Where N represents the number of white noise, ε represents the final standard deviation of the error, and A represents the amplitude of the noise. Therefore, the number of integrations N can be obtained by setting the value of the standard deviation of the expected error and the value of the noise amplitude A.

[0101] S22. Obtain the test signal after adding noise Test signal Determined by the following formula:

[0102]

[0103] Where x(t) represents the original signal to be decomposed, & = 1, 2, ε represents the final standard deviation of the error, and v j (t) represents a Gaussian white noise signal that satisfies the standard normal distribution, j=1, 2...N, N represents the number of white noises, and t represents time;

[0104] S23. Test signal To break it down:

[0105]

[0106] in, Indicates the test signal, represents the first sub-mode obtained, r1(t) represents the first residual signal, ACMD(·) represents the decomposition process after ACMD algorithm, and t represents time;

[0107] S24, after processing N test signals, N different Then the first extracted sub-mode is:

[0108]

[0109] Where A1(t) represents the first extracted sub-mode, j = 1, 2...N, N represents the number of white noises, that is, the number of test signals, represents the first sub-mode obtained by decomposing the j-th test signal, and t represents time;

[0110] S25. Calculate the residual after removing the first submode:

[0111] r1(t)=x(t)-A1(t)

[0112] Wherein, r1(t) represents the first residual signal, that is, the residual after removing the first extracted sub-mode A1(t), x(t) represents the original signal to be decomposed, A1(t) represents the first extracted sub-mode, and t represents time.

[0113] S26. Add positive and negative paired Gaussian white noise to the first residual signal r1(t) to obtain a new signal, and then perform ACMD decomposition again with the new signal as the carrier to obtain the first sub-mode From this, we can get the second extracted sub-mode as:

[0114]

[0115] Where A2(t) represents the second extracted sub-mode, j = 1, 2...N, N represents the number of white noises, that is, the number of test signals, represents the first sub-mode obtained by decomposing the j-th new signal, and t represents time;

[0116] S27. Calculate the residual after removing the second submode:

[0117] r2(t)=r1(t)-A2(t)

[0118] Where r2(t) represents the residual after removing the second sub-mode, r1(t) represents the residual after removing the first sub-mode, A2(t) represents the second extracted sub-mode, and t represents time;

[0119] S3, take the residual component as the new original signal, and obtain the kth residual signal r k (t);

[0120] Step S3 specifically includes the following steps:

[0121] S31, adding positive and negative paired Gaussian white noise to r1(t) to obtain a new signal, and then looping through steps S23-S24 again with the new signal as the carrier, and thereby obtaining the kth extracted sub-mode as:

[0122]

[0123] Among them, A k (t) represents the kth extracted sub-mode, j = 1, 2...N, N represents the number of white noise, that is, the number of test signals, represents the first sub-mode obtained by decomposing the j-th new signal, and t represents time;

[0124] S32. Calculate the residual after removing the kth sub-mode:

[0125] r k (t) = r k-1 (t)-A k (t)

[0126] Among them, r k (t) represents the kth residual signal, that is, removing the kth extracted sub-mode A k (t) The residual after k-1 (t) represents the k-1th residual signal, that is, removing the k-1th extracted sub-mode A k-1 (t) The residual after, A k (t) represents the kth extracted sub-modality, and t represents time.

[0127] S4. Determine the kth residual signal r k (t) Whether it can be further decomposed, if so, loop through steps S23-S24 until the kth residual signal r is obtained. k (t) is a monotonic function and cannot be further decomposed; if not, the k sub-modes of the original signal to be decomposed x(t) are output, that is:

[0128]

[0129] Where x(t) represents the original signal to be decomposed, i = 1, 2...k, k represents the number of sub-modes, A i (t) represents the i-th extracted sub-mode, r k (t) represents the kth residual signal, t represents time;

[0130] S5, obtaining all sub-modes, superimposing and reconstructing to obtain the processed sequence to be input;

[0131] S6. Use the CSSA optimization algorithm to optimize the parameters of the network structure, train and build the CSSA-LSTM model;

[0132] SSA is a new type of swarm intelligence optimization algorithm proposed by Xue Jiankai et al. in 2019, inspired by the foraging and anti-predation behaviors of sparrows. This paper applies the sparrow optimization algorithm to the deep extreme learning machine to optimize its parameters. In the SSA algorithm, sparrows in the population are usually divided into discoverers and joiners, and their identities change dynamically. Whenever a sparrow becomes a discoverer, another sparrow must become a joiner. In addition, each sparrow will be given an initial position and a fitness determined by the fitness function.

[0133] The specific steps of the improved sparrow algorithm are as follows:

[0134] In the sparrow optimization algorithm, the initialized population is randomly generated in the search space. However, the quality of the initialized population has a great influence on the efficiency of the optimization algorithm. Generally, a population evenly distributed in the search space is conducive to expanding the search range, thereby improving the convergence speed and accuracy of the algorithm. Due to the randomness and regularity of chaotic mapping, and the ability to not repeatedly traverse all states within a certain range, chaotic mapping is widely used in the initialization of population. Therefore, the present invention adopts cubic chaotic mapping to improve the sparrow optimization algorithm, and the population initialization process is as follows:

[0135] S601, setting the learning rate, number of iterations, and number of neurons in the second layer of the LSTM network within a certain range, and initializing parameters related to the sparrow algorithm;

[0136] S602, taking the mean square error of the test set as the sparrow fitness value and sorting them;

[0137] The CSSA-LSTM model includes an input layer, three LSTM hidden layers and an output layer;

[0138] S61, randomly generate chaotic vector y0=[y 0，1 ,y 0,2 , …, y 0,D ], where y 0,w ∈[0, 1], D represents the dimension of the search space, w = 1, 2, ..., D;

[0139] S62, generate chaotic sequence y1, y2, ..., y G ,

[0140] y v,w =4(y v-1，w ) 3 -3y v-1，w

[0141] Among them, y v,w Represents chaotic sequences in different dimensions, y v =[y v,1 ,y v,2 , …, y v,D ], v = 1, 2, ..., G, where G represents the population size, w = 1, 2, ..., D, where D represents the dimension of the search space;

[0142] S63, map the chaotic sequence to the original search space, and obtain the initialization population X = [x1, x2, ..., x G ], where x p =[x p,1 ,x p,2 ,…,x p,D ],

[0143] x p,q =x min,q +y p,q ·(x max,q -x min,q )

[0144] Among them, x p,q represents the pth individual in the qth dimension in the initialized population, x min,q and x max,q They represent the minimum and maximum values ​​of the qth dimension of the search space, q = 1, 2, …, D, where D represents the dimension of the search space;

[0145] S64, calculating the fitness value of the initial population and marking the best individual;

[0146] S65. Update current finder location

[0147]

[0148] in, Indicates the current finder location, represents the position of the discoverer in the previous iteration, T represents the current iteration number, p represents the number of individuals, q represents the dimension, iter maxrepresents the maximum number of iterations, ∝ is a random number in [0, 1], Q is a random number that follows a normal distribution, L represents a single-row d-column matrix whose internal elements are all 1, ST represents the safety value, and R2 represents the warning value;

[0149] S66. Update the joiner's position:

[0150]

[0151] in, Indicates the current joiner position, represents the position of the joiner in the previous iteration, Indicates the optimal position of the finder at present. represents the worst position in the current population; A represents a 1×d matrix, and each element is randomly assigned 1 or -1, L is a single-row d-column matrix whose internal elements are all 1, T represents the current iteration number, p represents the number of individuals, q represents the dimension, Q is a random number that obeys the normal distribution, n represents the number of participants, A + Indicates A T (AA T ) -1 ;

[0152] S67, Update predator position:

[0153]

[0154] in, represents the updated predator position, represents the position of the predator in the current iteration, p represents the number of individuals, q represents the dimension, β represents the step size control parameter, a random number that obeys a normal distribution with a mean of 0 and a variance of 1, and K represents a random number between -1 and 1. Indicates that the sparrow is currently in the safest position. represents the worst position in the current population, f i represents the fitness value of the current sparrow individual, f great and f worst Represent the current best and worst fitness values ​​respectively, γ represents the smallest constant to avoid the denominator being 0, and T represents the current number of iterations;

[0155] S68, determine the fitness value f of the current sparrow individual p Whether the preset value is met, if so, output the optimal sparrow position, otherwise return to step S65.

[0156] S7. Input the processed input sequence into the CSSA-LSTM network to obtain the prediction result of the gearbox life.

[0157] The original data sequence {X1(t), X2(t)...X n (t)} is put into the CSSA-LSTM network, and the predicted sequence {Y1(t), Y2(t)...Y n (t)}, to realize the prediction of gearbox life.

[0158] This embodiment uses a gearbox life prediction method based on NACMD-LSTM, which uses the input layer-three-layer LSTM hidden layer-output layer as the structural basis of the prediction model. The stacking of three layers of LSTM hidden layers not only improves the prediction model's ability to solve complex nonlinear regression problems in production scenarios, but also does not affect the model training efficiency due to too many layers, and reduces the real-time performance of the prediction model update and gearbox life prediction during the production process.

[0159] LSTM is the basis of the model, and its structure is as follows Figure 5 shown.

[0160] exist Figure 5 Medium t is the current input, h t is the hidden layer vector at time t, C t is the cell state at time t, Y t is the final output, with W and b representing the weight matrix and bias vector respectively.

[0161] The gearbox life prediction model structure in this embodiment is as follows: Figure 6 As shown in the figure, the model structure represents the number of LSTM layers and the number of neurons in each layer. In order to alleviate the overfitting problem, a Dropout layer is added after the LSTM hidden layer, and the attention mechanism is introduced to improve the model's attention to key manufacturing features. The fully connected layer is used as the output layer to implement the regression task, and the improved SSA is used to optimize the hyperparameters of the prediction model.

[0162] Since the proposed prediction model involves multiple hyperparameters, and different hyperparameters have a great impact on the actual application effect of the model. In order to better find the hyperparameters of the prediction model to achieve the best prediction effect in the shortest time, the parameters of the LSTM network model are optimized through the CSSA algorithm. The main process is as follows:

[0163] 1. Set the learning rate, number of iterations, and number of neurons in the second layer of the LSTM network within a certain range, and initialize the parameters related to the sparrow algorithm;

[0164] 2. Take the mean square error (MSE) of the test set as the sparrow fitness value and sort it;

[0165] 3. Update the position of the sparrow population in iteration according to steps S65-S67 and calculate the fitness value;

[0166] 4. Determine whether the termination conditions are met, the algorithm ends and the optimal parameters of the network are obtained.

[0167] A NACMD signal decomposition method adds multiple groups of positive and negative Gaussian white noise signals to the original signal to form a set of test signals and extract the first sub-mode; then, the residual signal of each decomposition is used as a new input signal for noise processing again to extract the second sub-mode; iterates until the residual signal can no longer be decomposed and all modal extractions are completed. This method has better noise resistance in a noisy environment, overcomes the problem of modal aliasing to a certain extent, has better signal decomposition capabilities, and improves the accuracy of subsequent predictions.

[0168] The CSSA optimization algorithm replaces the conventional random uniform distribution with cubic chaotic mapping for population initialization of the SSA algorithm, which enhances the diversity of the population to improve the performance of the algorithm, further avoids falling into local optimality, quickly searches for the global optimal solution, and can more efficiently solve some optimal problems.

[0169] Using the CSSA optimization algorithm in the CSSA-LSTM model to optimize the LSTM model parameters can help improve the performance and generalization ability of the model, accelerate the model convergence speed, and improve the prediction accuracy. At the same time, CSSA can also help find better parameter combinations, effectively search the parameter space, and improve the generalization performance of the model.

[0170] This embodiment designs a hybrid model of NACMD and CSSA-LSTM, decomposes the original signal into simpler sub-modes, reduces the complexity of the input signal, and then predicts them respectively through the network model, and reconstructs the output results on average to obtain the final prediction value, which improves the computing efficiency of the network and enhances the robustness and applicability of the prediction.

[0171] The above is a detailed introduction to the gearbox life prediction method and system based on NACMD-LSTM provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technicians in this field, according to the ideas of the present application, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. The gearbox life prediction method based on NACMD-LSTM is characterized by: include: S1, adding pairs of positive and negative white noises with opposite numbers to each other to the original signal to construct a new time series as the test sequence; S2. Based on the NACMD algorithm, perform modal decomposition on the test sequence to obtain the first sub-mode of the test sequence, calculate the mean value and record it as the first sub-mode A1(t) of the original signal, and record the corresponding residual; S3, take the residual as the new original signal, and obtain the kth residual signal r k (t); S4. Determine the kth residual signal r k (t) Whether it can be further decomposed, if so, loop through step S3 until the kth residual signal r is obtained. k (t) is a monotonic function and cannot be further decomposed; if not, the k sub-modes of the original signal to be decomposed x(t) are output, that is: Where x(t) represents the original signal to be decomposed, i = 1, 2...k, k represents the number of sub-modes, A i (t) represents the i-th extracted sub-mode, r k (t) represents the kth residual signal, t represents time; S5, obtaining all sub-modes, superimposing and reconstructing to obtain the processed sequence to be input; S6. Use the CSSA optimization algorithm to optimize the parameters of the network structure, train and build the CSSA-LSTM model; S7. Input the processed sequence to be input into the CSSA-LSTM network to obtain the prediction result of the gearbox life. Wherein, step S2 comprises: S21, obtaining the original signal to be decomposed x(t) and the number of white noises N; S22. Obtain the test signal after adding noise Test signal Determined by the following formula: Where x(t) represents the original signal to be decomposed, & = 1, 2, ε represents the final standard deviation of the error, and v j (t) represents a Gaussian white noise signal that satisfies the standard normal distribution, j=1, 2...N, N represents the number of white noises, and t represents time; S23. Test signal To break it down: in, Indicates the test signal, represents the first sub-mode obtained, r1(t) represents the first residual signal, ACMD(·) represents the decomposition process after ACMD algorithm, and t represents time; S24, after processing N test signals, obtain N different Then the first extracted sub-mode is: Where A1(t) represents the first extracted sub-mode, j = 1, 2...N, N represents the number of white noises, that is, the number of test signals, represents the first sub-mode obtained by decomposing the j-th test signal, and t represents time; S25. Calculate the residual after removing the first extracted submode A1(t): r1(t)=x(t)-A1(t) Where r1(t) represents the first residual signal, that is, the residual after removing the first extracted sub-mode A1(t), x(t) represents the original signal to be decomposed, A1(t) represents the first extracted sub-mode, and t represents time. Wherein, step S6 comprises: S61, randomly generate chaotic vector y0=[y 0，1 ,y 0,2 , …, y 0,D ], where y 0,w ∈[0, 1], D represents the dimension of the search space, w = 1, 2, ..., D; S62, generate chaotic sequence y1, y2, ..., y G , and v，w =4(and v-1，w ) 3 -3y v-1，w Among them, y v，w Represents chaotic sequences in different dimensions, y v =[y v,1 ,y v,2 , …, y v,D ], v = 1, 2, ..., G, where G represents the population size, w = 1, 2, ..., D, where D represents the dimension of the search space; S63, map the chaotic sequence to the original search space, and obtain the initialization population X = [x1, x2, ..., x G ], where x p =[x p,1 , x p,2 , …, x p,D ], x p,q =x min，q +y p,q ·(x max，q -x min，q ) Among them, x p，q represents the pth individual in the qth dimension in the initialized population, x min，q and x max，q They represent the minimum and maximum values ​​of the qth dimension of the search space, q = 1, 2, ..., D, where D represents the dimension of the search space; S64, calculating the fitness value of the initial population and marking the best individual; S65. Update the finder's location in, Indicates the current finder location, represents the position of the discoverer in the previous iteration, T represents the current iteration number, p represents the number of individuals, q represents the dimension, iter max represents the maximum number of iterations, ∝ is a random number in [0, 1], Q is a random number that follows a normal distribution, L represents a single-row d-column matrix whose internal elements are all 1, ST represents the safety value, and R2 represents the warning value; S66. Update the joiner's position: in, Indicates the current joiner position, represents the position of the joiner in the previous iteration, Indicates the optimal position of the finder at present. represents the worst position in the current population; A represents a 1×d matrix, and each element is randomly assigned 1 or -1, L is a single-row d-column matrix whose internal elements are all 1, T represents the current iteration number, p represents the number of individuals, q represents the dimension, Q is a random number that obeys the normal distribution, n represents the number of participants, A + Indicates A T (AA T ) -1 ; S67, Update predator position: in, represents the updated predator position, represents the position of the predator in the current iteration, p represents the number of individuals, q represents the dimension, β represents the step size control parameter, a random number that obeys a normal distribution with a mean of 0 and a variance of 1, and K represents a random number between -1 and 1. Indicates that the sparrow is currently in the safest position. represents the worst position in the current population, f p represents the fitness value of the current sparrow individual, f great and f worst Represent the current best and worst fitness values ​​respectively, γ represents the smallest constant to avoid the denominator being 0, and T represents the current number of iterations; S68, determine the fitness value f of the current sparrow individual p Whether the preset value is met, if so, output the optimal sparrow position, otherwise return to step S65.

2. The gearbox life prediction method based on NACMD-LSTM according to claim 1 is characterized in that: Step S3 includes: S31, adding positive and negative paired Gaussian white noise to the first residual signal r1(t) to obtain a new signal, and then using the new signal as a carrier to repeat steps S23-S24 again, and thereby obtaining the kth extracted sub-mode A k (t): Among them, A k (t) represents the kth extracted sub-modality, N represents the number of white noise, that is, the number of test signals. Indicates The first sub-mode obtained by decomposing the new signal, t represents time; S32, calculate and remove the k-th extracted sub-mode A k The residual after (t): r k (t)=r k-1 (t)-A k (t) Among them, r k (t) represents the kth residual signal, that is, removing the kth extracted sub-mode A k (t) The residual after k-1 (t) represents the k-1th residual signal, that is, removing the k-1th extracted sub-mode A k-1 (t) The residual after, A k (t) represents the kth extracted sub-modality, and t represents time.

3. The gearbox life prediction method based on NACMD-LSTM according to claim 1 is characterized in that: Step S21 includes: S211, obtaining an original signal to be decomposed, wherein the signal model of the original signal to be decomposed is expressed as: Among them, x(t) represents the original signal to be decomposed, α m (t) and β m (t) represents two non-stationary signals, Represents the frequency function of the non-stationary signal factor, m=1,2…M, x(t) is composed of M sub-modes, s represents an integral variable, and t represents time.

4. The gearbox life prediction method based on NACMD-LSTM according to claim 3 is characterized in that: Step S21 also includes: S212, obtaining the number N of white noise, which is obtained by the following formula: Where N represents the number of white noises, ε represents the final standard deviation of the error, and A represents the amplitude of the noise.

5. The gearbox life prediction method based on NACMD-LSTM according to claim 1 is characterized in that: Before step 61, the gearbox life prediction method based on NACMD-LSTM further includes: S601, setting the learning rate, number of iterations, and number of neurons in the second layer of the LSTM network within a certain range, and initializing parameters related to the sparrow algorithm; S602, taking the mean square error of the test set as the sparrow fitness value and sorting them; The CSSA-LSTM model includes an input layer, three LSTM hidden layers and an output layer.

6. An electronic device, characterized in that include: A memory for storing executable program codes; as well as A processor is connected to the memory, and runs a computer program corresponding to the executable program code by reading the executable program code, so as to execute the gearbox life prediction method based on NACMD-LSTM according to any one of claims 1-5.

7. The gearbox life prediction system based on NACMD-LSTM is characterized by: Comprising an electronic device according to claim 6.

Images