A hybrid fault separation method for transmission device based on memory selection mechanism

Abstract

The present invention proposes a transmission device mixed fault separation method based on a memory selection mechanism, which separates the target transmission device fault sound from a variety of mixed transmission device fault sounds, and obtains experimental data without contact measurement of the transmission device. The specific steps are as follows: step 1 transmission device fault sound coding network, step 2 long-term memory storage module; step 3 transmission device fault sound separation network; wherein the long-term memory storage module connects the first part and the second part of the two networks in series, and is used to store and update the output sound signal of the transmission device fault sound coding network, and input the latest stored sound signal into the transmission device fault sound separation network. The present invention has a small dependence on the amount of training data, can learn and memorize and store the characteristics of multiple types of transmission device fault sound signals, and accurately separate the target type of transmission device fault sound based on the voiceprint, and the overall robustness and anti-interference ability of the system are strong.

Description

Technical Field

[0001] The present invention relates to the technical field of transmission device fault diagnosis, which is an essential component of high-end equipment, and in particular to a transmission device mixed fault separation method based on a memory selection mechanism. Background Art

[0002] In the field of high-end equipment manufacturing, transmission devices are essential components of most mechanical equipment and are widely used in engineering machinery, aerospace, shipbuilding and other systems due to their advantages such as high transmission power. Due to the complex structure of the transmission system and the harsh working environment, the transmission device is easily damaged, resulting in an increase in system failures. Therefore, it has become a top priority to seek effective transmission device fault diagnosis methods and effectively detect and identify the fault characteristic information of the transmission device.

[0003] At present, transmission fault diagnosis is mainly based on vibration signals, and there is less research on acoustic signal analysis. However, the acquisition of vibration signals is limited by contact measurement, and traditional transmission fault diagnosis based on acoustics relies heavily on signal processing technology and the prior knowledge of diagnostic experts. Neural networks have been widely used in many fields due to their unique properties. The characteristics of neural networks are that they can complete the modeling of systems with complex internal relationships and can solve traditional nonlinear problems. In addition, they also have good robustness and fault tolerance. In recent years, many scholars have gradually applied neural network methods to the field of transmission fault diagnosis and achieved good prediction results. For example, convolutional neural networks (CNNs) can input time domain and frequency domain signals into the model as raw signals without feature conversion. Compared with traditional transmission fault diagnosis methods based on feature engineering, this method has achieved better results in transmission fault diagnosis.

[0004] Common transmission faults often do not exist alone in the system. The mixed transmission sounds of multiple faults bring further challenges to fault diagnosis. The famous cocktail party problem proposed by British cognitive scientist Cherry in 1953 when studying the attention selection mechanism can be used to solve the complex mixed transmission fault sound separation problem. The core of the cocktail party problem is the sound separation method. The mainstream sound separation methods can be divided into multi-channel sound separation methods and single-channel sound separation methods according to the number of microphones used. Among them, the common multi-channel sound separation methods are beamforming methods based on microphone arrays and multi-channel blind signal separation methods. According to the different principles of the methods, traditional speech separation methods can be divided into signal processing-based methods, decomposition-based methods and rule-based methods, such as: spectral subtraction, Wiener filter, auditory computational scene analysis (CASA), etc. The above methods assume that the sound obeys a certain distribution, but the real complex environment is not so ideal. When it is difficult to meet its assumed conditions, such methods will find it difficult to play their role. Although deep learning has achieved great success nowadays, there are still many problems to be solved in developing a computational auditory model to solve the cocktail party problem. Existing methods such as deep clustering and deep attractor networks cannot perfectly solve the problems of label ambiguity or arrangement, output bit mismatch, etc. Based on the above analysis, the main purpose of this invention is to establish a memory selection neural network model that can accurately and efficiently separate the fault sound signal of the hybrid transmission device. Summary of the invention

[0005] The purpose of the present invention is to provide a transmission device mixed fault separation method based on a memory selection mechanism, which can effectively separate a certain target transmission device fault sound data by only inputting mixed data of a certain number of transmission device fault sounds into the network.

[0006] The technical solution adopted by the present invention is:

[0007] A transmission device mixed fault separation method based on memory selection mechanism, the specific steps are as follows:

[0008] Step 1: transmission fault sound coding network, step 2: long-term memory storage module; step 3: transmission fault sound separation network; wherein the long-term memory storage module connects the first part and the second part of the network in series, and is used to store and update the output sound signal of the transmission fault sound coding network, and input the latest stored sound signal into the transmission fault sound separation network.

[0009] A further preferred solution is that step 1 of the transmission device fault sound coding network comprises the following steps:

[0010] Step 1-1: collecting sound signal data of a low-speed shaft transmission device, the sound signal data including sound signals of a normal transmission device, a broken tooth transmission device, a pitting transmission device and a worn transmission device;

[0011] Step 1-2: Divide the sound signal data in step 1-1 into three parts: training set, validation set and test set;

[0012] Step 1-3: Make the sound signal data in the training set into several tuples {X j～ ,(X k1 , ..., X kM )}, where X is the eigenvalue obtained after the fixed-length sound data is preprocessed by the logarithmic Mel filter bank, j and k represent the transmission device fault type corresponding to the sound signal; when j and k are from the same transmission device fault type, this tuple is a positive tuple, otherwise it is a negative tuple;

[0013] Step 1-4: Input the tuple obtained in step 1-3 into the LSTM network and calculate the L2 normalized response to obtain: {e, (e k1 , ..., e kM )}, where e is the fixed-dimensional embedding vector generated by LSTM; the loss function of the transmission fault sound coding network is:

[0014] L T (e j～ , c k )=δ(j,k)(1-σ(s))+(1-δ(j,k))σ(s); where: σ(x): 1 / (1+e -x ) is the standard sigmoid activation function; δ(j, k) is 1 when j = k, otherwise it is 0; s = ω·cose j～ , c k )+b is the cosine similarity function, where ω and b are the parameters updated by network learning;

[0015] Step 1-5: Calculate the tuple (e k1 , ..., e kM ) is the sound pattern C of type k of transmission fault k ; C k Multiple sound signals generated by k faults:

[0016] Step 1-6: After network training is completed, take out the final training result C in step 1-5 k As the sound print encoding of K-type transmission fault, it prepares for the subsequent long-term memory storage module operation.

[0017] A further preferred solution is that step 2 of the long-term memory storage module comprises the following steps:

[0018] Step 2-1: Collect the fault sound coding voiceprint and type label of the low-speed shaft transmission device to build and update the long-term memory module M=(K memory , V memory×vector , A memory ); wherein K stores the type of transmission fault sound, V stores the voiceprint encoding vector corresponding to the relevant fault sound type, and A stores the unupdated duration of the voiceprint encoding vector stored in the long-term memory module;

[0019] Step 2-2: Each time the transmission fault sound coding network training is completed, the value of K is updated:

[0020] The K value is updated in such a way that when the transmission device fault sound being trained is of an existing type in the memory module, the K value corresponding to the type remains unchanged; if it is a new transmission device fault sound, a new space is opened in the memory module to store the K value;

[0021] Step 2-3: Each time the transmission fault sound coding network training is completed, update the value of A:

[0022] The A value is updated in such a way that when the transmission fault sound being trained is of an existing type in the memory module, the A corresponding to the type is reset to zero; if it is a new transmission fault sound, a new space is opened in the memory module and the corresponding A is reset to zero; this update does not involve the addition of 1 to the A corresponding to the transmission fault sound type;

[0023] Step 2-4: Each time the transmission fault sound coding network training is completed, the value of V is updated:

[0024] The V value is updated in such a way that the transmission fault sound is trained to the existing type in the memory module.

[0025] Where V is the voiceprint encoding value stored in the previous memory module, C k is the voiceprint encoding value obtained in this training; if it is a new transmission failure sound, C k Store directly into the memory module.

[0026] A further preferred solution is that step 3 of separating the transmission device fault sound network comprises the following steps:

[0027] Step 3-1: collecting sound signal data of a low-speed shaft transmission device, the sound signal data including sound signals of a normal transmission device, a broken tooth transmission device, a pitting transmission device and a worn transmission device;

[0028] Step 3-2: Randomly select two of the four transmission device fault sound signal data obtained in step 3-1, and perform linear mixing to obtain mixed input sound signal data x;

[0029] Step 3-3: For the mixed input x obtained in step 3-2, perform STFT transformation to obtain the time-frequency spectrum X t，f ;

[0030] Step 3-4: For the mixed input time-frequency spectrum X obtained in step 3-3, apply BiLSTM to calculate the hidden state along the time dimension: The detailed calculation process of LSTM is as follows:

[0031]

[0032] Among them, x t is the input feature point of the network, C t is the new storage content of the memory unit, C t-1 is the storage content of the memory unit before updating, h t is the new output of the neuron, h t-1 is the output of the neuron before updating; σ is the sigmoid activation function. b f The weights and biases needed to be learned for the forget gate, b g is the weight and bias to be learned for the input gate tanh activation function, b i is the weight and bias to be learned for the input gate σ activation function, b o are the weights and biases that the output gate needs to learn. f is the input of the forget gate, g is the input of the input gate tanh activation function, i is the input of the input gate σ activation function, and o is the input of the output gate; ⊙ is the dot product of the corresponding position elements in the matrix.

[0033] Step 3-5: The hidden state h obtained in step 3-4 t Input to the shaping layer, for each unit X in the time spectrum t，f Generate embedding vector h t，f =o⊙tanh(c t ), where h t，f is a d-dimensional vector;

[0034] Step 3-6: Use the time-frequency unit hidden state h obtained in step 3-5 t，f The target transmission fault type voiceprint code v that needs to be separated from the long-term memory storage module is calculated, and the time-frequency unit mask α is calculated. t，f =sigmod(g T tanh(W·v+U·ht，f )), where the d-dimensional vector g, the d×d-dimensional matrix, W and U are the parameters that the neural network needs to learn through training, α t，f are the parameters of the time-frequency selective attention filter;

[0035] Step 3-7: Minimize the target Loss function during training Thereby updating the network parameters U and W;

[0036] Step 3-8: After the transmission device fault sound separation network training is completed, the pure time spectrum of a certain transmission device fault separated from the multiple mixed transmission device fault sound signals is X t，f ×α t，f ;

[0037] Step 3-9: For a certain transmission device fault, the pure spectrum is X t，f ×α t，f Perform iSTFT to obtain a pure time domain signal x' of a certain transmission device fault;

[0038] Step 3-10: Compare the obtained pure time domain signal x' of a certain transmission device fault with the true value in the testdata and calculate GNSDR to verify the effectiveness of the network.

[0039] A further preferred solution is to divide the sound signal data in step 1-2 into three parts: a training set, a validation set and a test set according to a ratio of 7:2:1.

[0040] The present invention has the advantages that:

[0041] 1. The present invention adds a transmission device fault sound voiceprint coding network based on LSTM to generate a voiceprint vector of a certain transmission device fault sound. Compared with the traditional sound separation neural network, the network model after adding the voiceprint coding network solves the problems of uncertain output quantity and disordered order, and the subjective and objective effects of sound separation are improved, and the network is more stable.

[0042] 2. The present invention introduces a long-term memory storage module for storing and updating the voiceprint coding of the transmission device fault sound after each training, so that the network has long-term memory and can better record the acoustic characteristics of various types of transmission device fault sound signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of the transmission device mixed fault separation method based on the memory selection mechanism of the present invention.

[0044] Figure 2 It is a diagram of the transmission device fault sound encoding network and the transmission device fault sound separation network architecture of the present invention.

[0045] Figure 3 This is a diagram of the double-layer LSTM network architecture used in the present invention.

[0046] Figure 4a It is a schematic diagram of GNSDR comparison between the method of the present invention and the traditional method when the fault sounds of two transmission devices are mixed and there is no background noise.

[0047] Figure 4b It is a schematic diagram of GNSDR comparison between the method of the present invention and the traditional method when the fault sounds of three transmission devices are mixed and there is no background noise.

[0048] Figure 4c It is a schematic diagram of GNSDR comparison between the method of the present invention and the traditional method when the fault sounds of two transmission devices are mixed and there is background noise.

[0049] Figure 5a It is a time domain diagram of the mixed sound signal of the broken tooth transmission fault and the worn transmission fault.

[0050] Figure 5b It is the time domain diagram of the broken tooth transmission fault sound signal predicted by the transmission fault sound separation network.

[0051] Figure 5c It is the time domain diagram of the original broken tooth transmission fault sound signal.

[0052] Figure 5d It is the time domain diagram of the worn transmission fault sound signal predicted by the transmission fault sound separation network.

[0053] Figure 5e It is the time domain diagram of the mixed sound signal of the worn transmission fault.

[0054] Figure 6a It is a time-frequency diagram of the mixed sound signal of the broken tooth transmission device fault and the worn transmission device fault.

[0055] Figure 6b It is the time-frequency diagram of the broken tooth transmission fault sound signal predicted by the transmission fault sound separation network.

[0056] Figure 6c It is the time-frequency diagram of the original broken tooth transmission fault sound signal.

[0057] Figure 6d It is the time-frequency diagram of the worn transmission fault sound signal predicted by the transmission fault sound separation network.

[0058] Figure 6e It is the time-frequency diagram of the mixed sound signal of the worn transmission fault. DETAILED DESCRIPTION

[0059] Below is the attached figure Figure 1-3The present invention is described in further detail.

[0060] A transmission device mixed fault separation method based on memory selection mechanism, the specific steps are as follows:

[0061] Step 1: transmission fault sound coding network, step 2: long-term memory storage module; step 3: transmission fault sound separation network; wherein the long-term memory storage module connects the first part and the second part of the network in series, and is used to store and update the output sound signal of the transmission fault sound coding network, and input the latest stored sound signal into the transmission fault sound separation network.

[0062] A further preferred solution is that step 1 of the transmission device fault sound coding network comprises the following steps:

[0063] Step 1-1: collecting sound signal data of a low-speed shaft transmission device, the sound signal data including sound signals of a normal transmission device, a broken tooth transmission device, a pitting transmission device and a worn transmission device;

[0064] Step 1-2: Divide the sound signal data in step 1-1 into three parts: training set, validation set and test set;

[0065] Step 1-3: Make the sound signal data in the training set into several tuples {X j～ ,(X k1 , ..., X kM )}, where X is the eigenvalue obtained after the fixed-length sound data is preprocessed by the logarithmic Mel filter bank, j and k represent the transmission device fault type corresponding to the sound signal; when j and k are from the same transmission device fault type, this tuple is a positive tuple, otherwise it is a negative tuple;

[0066] Step 1-4: Input the tuple obtained in step 1-3 into the LSTM network and calculate the L2 normalized response to obtain: {e, (e k1 , ..., e kM )}, where e is the fixed-dimensional embedding vector generated by LSTM; the loss function of the transmission fault sound coding network is:

[0067] L T (e j～ , c k )=δ(j,k)(1-σ(s))+(1-δ(j,k))σ(s); where: σ(x)=1 / (1+e -x ) is the standard sigmoid activation function; δ(j, k) is 1 when j = k, otherwise it is 0; s = ω·cos(e j～ , c k)+b is the cosine similarity function, where ω and b are the parameters updated by network learning;

[0068] Step 1-5: Calculate the tuple (e k1 , ..., e kM ) is the sound pattern C of type k of transmission fault k ; C k Multiple sound signals generated by k faults:

[0069] Step 1-6: After network training is completed, take out the final training result C in step 1-5 k As the sound print encoding of K-type transmission fault, it prepares for the subsequent long-term memory storage module operation.

[0070] A further preferred solution is that step 2 of the long-term memory storage module comprises the following steps:

[0071] Step 2-1: Collect the fault sound coding voiceprint and type label of the low-speed shaft transmission device to build and update the long-term memory module M=(K memory , V memory×vector , A memory ); wherein K stores the type of transmission fault sound, V stores the voiceprint encoding vector corresponding to the relevant fault sound type, and A stores the unupdated duration of the voiceprint encoding vector stored in the long-term memory module;

[0072] Step 2-2: Each time the transmission fault sound coding network training is completed, the value of K is updated:

[0073] The K value is updated in such a way that when the transmission device fault sound being trained is of an existing type in the memory module, the K value corresponding to the type remains unchanged; if it is a new transmission device fault sound, a new space is opened in the memory module to store the K value;

[0074] Step 2-3: Each time the transmission fault sound coding network training is completed, update the value of A:

[0075] The A value is updated in such a way that when the transmission fault sound being trained is of an existing type in the memory module, the A corresponding to the type is reset to zero; if it is a new transmission fault sound, a new space is opened in the memory module and the corresponding A is reset to zero; this update does not involve the addition of 1 to the A corresponding to the transmission fault sound type;

[0076] Step 2-4: Each time the transmission fault sound coding network training is completed, the value of V is updated:

[0077] The V value is updated in such a way that the transmission fault sound is trained to the existing type in the memory module.

[0078] Where V is the voiceprint encoding value stored in the previous memory module, C k is the voiceprint encoding value obtained in this training; if it is a new transmission failure sound, C k Store directly into the memory module.

[0079] A further preferred solution is that step 3 of separating the transmission device fault sound network comprises the following steps:

[0080] Step 3-1: collecting sound signal data of a low-speed shaft transmission device, the sound signal data including sound signals of a normal transmission device, a broken tooth transmission device, a pitting transmission device and a worn transmission device;

[0081] Step 3-2: Randomly select two of the four transmission device fault sound signal data obtained in step 3-1, and perform linear mixing to obtain mixed input sound signal data x;

[0082] Step 3-3: For the mixed input x obtained in step 3-2, perform STFT transformation to obtain the time-frequency spectrum X t，f ;

[0083] Step 3-4: For the mixed input time-frequency spectrum X obtained in step 3-3, apply BiLSTM to calculate the hidden state along the time dimension: The detailed calculation process of LSTM is as follows:

[0084]

[0085] Among them, x t is the input feature point of the network, C t is the new storage content of the memory unit, C t-1 is the storage content of the memory unit before updating, h t is the new output of the neuron, h t-1 is the output of the neuron before updating. σ is the sigmoid activation function. b f The weights and biases needed to be learned for the forget gate, b g is the weight and bias to be learned for the input gate tanh activation function, b i is the weight and bias to be learned for the input gate σ activation function, b o are the weights and biases that the output gate needs to learn. f is the input of the forget gate, g is the input of the input gate tanh activation function, i is the input of the input gate σ activation function, and o is the input of the output gate. ⊙ is the dot product of the corresponding position elements in the matrix.

[0086] Step 3-5: The hidden state h obtained in step 3-4 t Input to the shaping layer, for each unit X in the time spectrum t，f Generate embedding vector h t，f =o⊙tanh(c t ), where h t，f is a d-dimensional vector;

[0087] Step 3-6: Use the time-frequency unit hidden state h obtained in step 3-5 t，f The target transmission fault type voiceprint code v that needs to be separated from the long-term memory storage module is calculated, and the time-frequency unit mask α is calculated. t，f =sigmod(g T tanh(W·v+U·h t，f )), where the d-dimensional vector g, the d×d-dimensional matrix, W and U are the parameters that the neural network needs to learn through training, α t，f are the parameters of the time-frequency selective attention filter;

[0088] Step 3-7: Minimize the target Loss function during training Thereby updating the network parameters U and W.

[0089] Step 3-8: After the transmission device fault sound separation network training is completed, the pure time spectrum of a certain transmission device fault separated from the multiple mixed transmission device fault sound signals is X t，f ×α t，f ;

[0090] Step 3-9: For a certain transmission device fault, the pure spectrum is X t，f ×α t，f Perform iSTFT to obtain a pure time domain signal x' of a certain transmission device fault;

[0091] Step 3-10: Compare the obtained pure time domain signal x' of a certain transmission device fault with the true value in the testdata and calculate GNSDR to verify the effectiveness of the network.

[0092] A further preferred solution is to divide the sound signal data in step 1-2 into three parts: a training set, a validation set and a test set according to a ratio of 7:2:1.

[0093] According to the continuity and time series characteristics of the transmission fault sound, the present invention uses the LSTM (Long Short-Term Memory) neural network model to identify and separate the fault sound. LSTM is a long short-term memory network, a time recursive neural network (RNN), which is mainly used to solve the gradient vanishing and gradient explosion problems in the long sequence training process. In short, compared with ordinary RNN, LSTM can perform better in longer sequences and can separate a variety of mixed transmission fault sounds into a plurality of single-type transmission fault sounds.

[0094] The present invention utilizes the reference signal of the target transmission fault sound to separate the target transmission fault sound from multiple mixed transmission fault sounds. The present invention achieves this by training two independent neural networks and a long-term memory storage module:

[0095] (1) Transmission fault sound encoding network: A fault type encoding network that generates a target transmission fault sound embedding vector (voiceprint). The input is a pure single-type transmission fault sound, and the output is the voiceprint code.

[0096] (2) Long-term memory storage module: used to update and store the single-type transmission fault sound voiceprint codes produced by the transmission fault sound coding network, and store them by type. The memory module is updated after each network update.

[0097] (3) Transmission fault sound separation network: A sound separation network that generates a spectrum mask for a pure single-type transmission fault sound. The input is a mixed signal of multiple transmission fault sounds and the voiceprint encoding of the target transmission fault sound, and the output is a pure target transmission fault sound. The network uses a two-layer BiLSTM neural network.

[0098] In order to prove the effectiveness of the transmission fault sound separation method proposed in the present invention (a system consisting of a transmission fault sound encoding network, a long-term memory storage module, and a transmission fault sound separation network), sound signal data of a low-speed shaft transmission device was collected, and a certain number of sound signals of normal transmission devices, broken tooth transmission devices, pitting transmission devices, and worn transmission devices were collected, and divided into three different data sets (training set traindata, validation set validdata, and test set testdata).

[0099] The DC (deep clustering) method, DC (-40), DC (-60), DC (-80) method and the transmission device fault sound separation method designed by the present invention are used to train the network model based on the training set and the verification set, and the network effectiveness is verified by the test set after the network training is completed. Among them, -40dB means that the background noise threshold is set to 40dB of the maximum input amplitude, which is an adjustable parameter of the deep clustering method.

[0100] according to Figure 4a , Figure 4b and Figure 4c It can be seen that the method proposed in the present invention has excellent performance in GNSDR on the three test sets, and is much greater than the four DC-based methods when there is background noise. After adding the transmission fault sound encoding network, the model can better learn the acoustic characteristics of the transmission fault sound; after adding the long-term memory module, the model can better store the learned features into the memory module, so that the overall performance of the transmission fault sound separation network is further improved. Combining the three situations, the transmission fault sound separation method designed by the present invention has the largest average GNSDR and the best robustness among the five methods.

[0101] The network training is performed on the sound signal data sets of normal transmission, broken tooth transmission, pitting transmission and worn transmission respectively, and the trained transmission fault sound separation network is used to predict and separate the sound signal of the mixed sound of the broken tooth transmission fault sound and the worn transmission sound, and the predicted broken tooth transmission fault sound and the worn transmission fault sound are obtained. The separation time domain results are as follows Figure 5a , Figure 5b , Figure 5c , Figure 5d , Figure 5e The time-frequency domain results are shown in Figure 6a , Figure 6b , Figure 6c , Figure 6d , Figure 6e shown.

[0102] The present invention can effectively separate the hybrid transmission fault sound into several pure transmission fault sounds, has good separation capabilities in both time domain and frequency domain, and the average GNSDR of experimental data is as high as 4.39, which has good robustness.

Claims

1. A gear fault sound signal separation algorithm based on memory selection neural network, characterized in that: The specific steps are as follows: Step 1: gear fault sound encoding network, step 2: long-term memory storage module; step 3: gear fault sound separation network; wherein the long-term memory storage module connects the first part and the second part of the network in series, and is used to store and update the output sound signal of the gear fault sound encoding network, and input the latest stored sound signal into the gear fault sound separation network; Step 2 The long-term memory storage module includes the following steps: Step 2-1: Collect the fault sound coding voiceprint and type label of the low-speed shaft gear to build and update the long-term memory module M=(K memory V memory×vector , A memory ); wherein K stores the type of gear fault sound, V stores the voiceprint encoding vector corresponding to the relevant fault sound type, and A stores the unupdated duration of the voiceprint encoding vector stored in the long-term memory module; Step 2-2: Each time the gear fault sound coding network training is completed, update the value of K: The K value is updated in such a way that when the gear fault sound being trained is of an existing type in the memory module, the K value corresponding to the type remains unchanged; if it is a new gear fault sound, a new space is opened in the memory module to store the K value; Step 2-3: Each time the gear fault sound encoding network training is completed, update the value of A: The way to update the A value is that when the trained gear fault sound is an existing type in the memory module, the A corresponding to the type is reset to zero; if it is a new gear fault sound, a new space is opened in the memory module and the corresponding A is reset to zero; this update does not involve adding 1 to the A corresponding to the gear fault sound type; Step 2-4: Each time the gear fault sound encoding network training is completed, update the value of V: The V value is updated by training the gear failure sound to the existing type in the memory module. Where V is the voiceprint encoding value stored in the previous memory module, C k is the voiceprint encoding value obtained in this training; if it is a new gear failure sound, C k Store directly into the memory module.

2. The gear fault sound signal separation algorithm based on memory selection neural network according to claim 1 is characterized in that: Step 1 The gear fault sound encoding network includes the following steps: Step 1-1: Collecting sound signal data of a low-speed shaft gear, the sound signal data includes sound signals of a normal gear, a broken tooth gear, a pitting gear and a worn gear; Step 1-2: Divide the sound signal data in step 1-1 into three parts: training set, validation set and test set; Step 1-3: Make the sound signal data in the training set into several tuples {X j～ ,(X k1 , ..., X kM )}, where X is the eigenvalue obtained after the fixed-length sound data is preprocessed by the logarithmic Mel filter bank, and j and k represent the gear fault type corresponding to the sound signal; When j and k are from the same gear fault type, this tuple is a positive tuple, otherwise it is a negative tuple; Step 1-4: Input the tuple obtained in step 1-3 into the LSTM network and calculate the L2 normalized response to obtain: {e, (e k1 , ..., e kM )}, where e is an embedding vector of fixed dimension generated by LSTM; the loss function of the gear fault sound encoding network is: L T (e j ~c k )=δ(j,k)(1-σ(s))+(1-δ(j,k))+σ(s); where: σ(x)=1 / (1+e -x ) is the standard sigmoid activation function; δ(j, k) is 1 when j = k, otherwise it is 0; s=ω·cos(e j～ , c k )+b is the cosine similarity function, where ω and b are the parameters updated by network learning; Step 1-5: Calculate the tuple (e k1 , ..., e kM ) is the soundprint C of gear fault type k k ; C k Multiple sound signals generated by k faults: Step 1-6: After network training is completed, take out the final training result C in step 1-5 k As the sound print encoding of K-type gear failure, it prepares for the subsequent long-term memory storage module work.

3. The gear fault sound signal separation algorithm based on memory selection neural network according to claim 1 is characterized in that: Step 3 The gear fault sound separation network includes the following steps: Step 3-1: Collecting sound signal data of a low-speed shaft gear, the sound signal data includes sound signals of a normal gear, a broken tooth gear, a pitting gear and a worn gear; Step 3-2: Randomly select two of the four gear fault sound signal data obtained in step 3-1, and perform linear mixing to obtain mixed input sound signal data x; Step 3-3: For the mixed input x obtained in step 3-2, perform STFT transformation to obtain the time-frequency spectrum X t，f ; Step 3-4: For the mixed input time-frequency spectrum X obtained in step 3-3, apply BiLSTM to calculate the hidden state along the time dimension: The detailed calculation process of LSTM is as follows: Among them, x t is the input feature point of the network, C t is the new storage content of the memory unit, C t-1 is the storage content of the memory unit before updating, h t is the new output of the neuron, h t-1 is the output of the neuron before updating; σ is the sigmoid activation function; b f The weights and biases needed to be learned for the forget gate, b g is the weight and bias to be learned for the input gate tanh activation function, b i is the weight and bias to be learned for the input gate σ activation function, b o are the weights and biases that the output gate needs to learn; f is the input of the forget gate, g is the input of the input gate tanh activation function, i is the input of the input gate σ activation function, and o is the input of the output gate; ⊙ is the dot product of the elements at the corresponding position in the matrix; Step 3-5: The hidden state h′ obtained in step 3-4 t Input to the shaping layer, for each unit X in the time spectrum t,f Generate embedding vector h t，f =o⊙tanh(c t ), where h t，f is a d-dimensional vector; Step 3-6: Use the time-frequency unit hidden state h obtained in step 3-5 t，f The target gear fault type voiceprint encoding v that needs to be separated from the long-term memory storage module is calculated, and the time-frequency unit mask α is calculated. t，f =sigmod(g T tanh(W·v+U·h t，f )), where the d-dimensional vector g, the d×d-dimensional matrix, W and U are the parameters that the neural network needs to learn through training, α t，f are the parameters of the time-frequency selective attention filter; Step 3-7: Minimize the target Loss function during training Thereby updating the network parameters U and W; Step 3-8: After the gear fault sound separation network training is completed, the pure time spectrum of a certain gear fault separated from multiple mixed gear fault sound signals is X t，f ×α t，f ; Step 3-9: For a certain gear fault, the pure spectrum is X t，f ×α t，f Perform iSTFT to obtain a pure time domain signal x' of a certain gear fault; Step 3-10: Compare the obtained pure time domain signal x' of a certain gear fault with the true value in the testdata and calculate GNSDR to verify the effectiveness of the network.

4. The gear fault sound signal separation algorithm based on memory selection neural network according to claim 2 is characterized in that: The sound signal data in step 1-2 is divided into three parts: training set, validation set and test set according to 7:2:1.

Images