A pulse neural network mechanical fault diagnosis method based on STFT dimension transformation

Abstract

The application is applied to the field of mechanical fault diagnosis signal processing, and particularly discloses a mechanical fault diagnosis method based on STFT dimension transformation, which comprises the following steps: collecting a one-dimensional mechanical vibration signal, and performing wavelet decomposition; performing wavelet reconstruction on low-frequency components and high-frequency components after denoising processing, so as to obtain a one-dimensional vibration signal after denoising; performing short-time Fourier transform, so as to convert the signal into a time-frequency two-dimensional matrix; inputting the time-frequency two-dimensional matrix into an improved HH threshold neuron model, performing Poisson sparse coding on the time-frequency two-dimensional matrix, and only performing pulse response on signal significant features; constructing a threshold coding convolution network with residual connection, inputting a sparse coding matrix, training by using an unsupervised learning rule based on STDP, and adaptively adjusting network synaptic weights; inputting into the trained threshold coding convolution network, and determining a fault diagnosis result by means of pulse firing activities of output layer neurons obtained through network forward propagation.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of signal processing, in particular to a method for mechanical fault diagnosis based on STFT dimension transformation of a supra-threshold encoding convolutional network. BACKGROUND

[0002] As one of the important parts of rotating machinery, gears are widely used in mechanical transmission, equipment manufacturing and other fields, which can directly affect the efficiency of industrial production, and gears usually need to work in various complex, unstable and high-strength conditions, which greatly increases the possibility of damage. Therefore, the fault diagnosis and detection of gears are important links to ensure safe production. The fault diagnosis of gears has experienced the development process of experience-based, sensor-based and data-driven, and with the development of artificial intelligence in various fields, the fault diagnosis of gears begins to use more and more data-driven and neural network-based methods, such as using CNN and LSTM for fault diagnosis.

[0003] In order to further improve the accuracy of fault diagnosis, methods combining signal processing methods with various neural networks are more widely used in practical applications, for example, wavelet-based convolutional attention neural network, discrete wavelet transform-based principal component analysis artificial neural network and genetic mutation particle swarm optimization probability neural network algorithm. There are some transfer learning methods based on deep convolutional neural network and methods based on deep convolutional generative adversarial network to deal with this problem.

[0004] Although the above methods have achieved good diagnostic results, there are still some shortcomings. Therefore, in order to reduce the influence of noise, effectively extract features and better process time series data, a method for mechanical fault diagnosis based on STFT dimension transformation of a supra-threshold encoding convolutional network is proposed. SUMMARY

[0005] The application proposes a method for mechanical fault diagnosis based on STFT dimension transformation of a pulse neural network.

[0006] The method for mechanical fault diagnosis based on STFT dimension transformation of a pulse neural network comprises the following steps:

[0007] A one-dimensional mechanical vibration signal is collected and wavelet decomposition is performed to obtain high-frequency components and low-frequency components. The low-frequency components and the denoised high-frequency components are reconstructed by wavelet to obtain a denoised one-dimensional vibration signal;

[0008] The denoised one-dimensional vibration signal is subjected to short-time Fourier transform to convert it into a time-frequency two-dimensional matrix;

[0009] The time-frequency two-dimensional matrix is input into an improved HH suprathreshold neuron model, a pulse firing mechanism with an adaptive threshold is constructed, Poisson sparse coding is performed on the time-frequency two-dimensional matrix, a binary pulse sequence is generated, only the signal significant features are responded to the pulse, and a sparse coding matrix is formed;

[0010] A threshold coding convolution network with residual connection is constructed, the sparse coding matrix is input into the network, the threshold coding convolution network is trained by using an unsupervised learning rule based on pulse time-dependent plasticity, and the network synaptic weight is adaptively adjusted; and the network synaptic weight is adaptively adjusted according to the time difference of the pulse firing of the presynaptic neuron and the postsynaptic neuron;

[0011] The mechanical vibration signal to be tested is input into the trained threshold coding convolution network, the pulse firing activity of the output layer neuron is obtained through network forward propagation, and the fault diagnosis result is determined based on the pulse firing activity.

[0012] Preferably, the denoising processing of the high-frequency component includes an adaptive threshold function, the adaptive threshold is determined by calculating a noise standard deviation of the high-frequency component, and the noise standard deviation is calculated and obtained by using a median absolute deviation method.

[0013] Preferably, the short-time Fourier transform of the denoised one-dimensional vibration signal includes local spectrum analysis of the signal by using a window function, the window length of the window function is adaptively determined according to the signal characteristics of the denoised one-dimensional vibration signal, and the signal characteristics include a sampling frequency and a highest frequency component of the denoised one-dimensional vibration signal.

[0014] After the time-frequency two-dimensional matrix obtained by conversion is processed by taking a modulus, logarithmic scaling is performed, so as to avoid taking logarithm of zero and enhance the feature visualization effect.

[0015] Preferably, the improved HH suprathreshold neuron model includes an adaptive dynamic threshold and a noise robustness mechanism introduced on the basis of a classical HH model.

[0016] The adaptive dynamic threshold includes a basic threshold, a threshold adjustment coefficient and a threshold decay time constant, when the neuron membrane potential reaches the adaptive dynamic threshold, the neuron fires a pulse, the membrane potential is reset at the same time, and the adaptive dynamic threshold is increased instantaneously and then decays exponentially; and the noise robustness mechanism is realized by introducing Gaussian white noise, so as to improve the noise tolerance of the model.

[0017] As preferred, when inputting the time-frequency two-dimensional matrix into the improved HH suprathreshold neuron model, each element of the time-frequency two-dimensional matrix is converted into an input current of a neuron, and the input current is normalized by a scaling factor, a mean value and a standard deviation of the time-frequency two-dimensional matrix, and a bias current constant.

[0018] As preferred, the constructing the suprathreshold coding convolutional network with residual connection comprises the following: obtaining the residual connection of the suprathreshold coding convolutional network through a weight matrix of a short circuit connection and a nonlinear transformation result in a residual block, and the nonlinear transformation comprises a convolution operation and an improved suprathreshold neuron spiking process.

[0019] As preferred, the unsupervised learning rule based on the pulse time-dependent plasticity is used to train the suprathreshold coding convolutional network, and the adaptive adjustment of the network synaptic weight comprises the following: the weight of the pre-synaptic neuron pulse sequence and the post-synaptic neuron pulse sequence with a positive causal relationship is enhanced, and the weight of the pre-synaptic neuron pulse sequence and the post-synaptic neuron pulse sequence without correlation is inhibited; and the weight update only depends on the local spiking time of the pre-synaptic neuron and the post-synaptic neuron.

[0020] As preferred, the determination of the fault diagnosis result based on the spiking activity comprises the following: calculating the average spiking rate of each neuron of the output layer within a set time window, adopting an all-win strategy, and taking the fault type corresponding to the neuron with the highest average spiking rate as the final fault diagnosis result; and simultaneously calculating the relative spiking rate difference of the output layer neurons, and when the relative spiking rate difference is lower than a set threshold, it is determined that the fault diagnosis result is unreliable, and the mechanical vibration signal to be measured needs to be re-acquired, processed and diagnosed.

[0021] The application further discloses a suprathreshold coding convolutional network mechanical fault diagnosis device based on STFT dimension transformation, comprising a memory and one or more processors, the memory stores executable code, and the one or more processors execute the executable code to implement the suprathreshold coding convolutional network mechanical fault diagnosis method based on STFT dimension transformation.

[0022] The application further discloses a computer readable storage medium, which stores a program, and the program is executed by a processor to implement the suprathreshold coding convolutional network mechanical fault diagnosis method based on STFT dimension transformation.

[0023] The application has the following beneficial effects:

[0024] 1. The application only processes key fault features through wavelet denoising, STFT dimension transformation and Poisson sparse coding, does not need to process all signals, significantly reduces data volume and calculation amount, adapts to edge computing demand, and greatly improves calculation and storage efficiency.

[0025] 2. The application expresses the time-frequency feature by STFT enhancement, prevents gradient disappearance by residual network, adaptively learns the time sequence feature by STDP rule, effectively resists noise interference, accurately identifies mechanical faults, and guarantees fault diagnosis accuracy.

[0026] 3. The application can work in a high-noise industrial environment, can identify gear normal, inner ring, outer ring, and rolling body faults, supports real-time diagnosis, provides a feasible solution for mechanical fault monitoring, and is suitable for complex practical scenes. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 a schematic diagram of the mechanical fault diagnosis method based on the STFT dimension transformation threshold coding convolution network of the application;

[0028] Figure 2 a mechanical fault diagnosis overall algorithm flowchart of the STFT dimension transformation threshold coding convolution network of the application;

[0029] Figure 3 a wavelet threshold denoising algorithm diagram of the application;

[0030] Figure 4 a SHHResNet neural pulse network structure diagram of the application;

[0031] Figure 5 a schematic diagram of the mechanical fault diagnosis device based on the STFT dimension transformation threshold coding convolution network of the application. DETAILED DESCRIPTION

[0032] In order to make the person in the art better understand the technical solutions in the application, the technical solutions in the embodiments of the application will be clearly described below in combination with the embodiments.

[0033] As Figure 1 , 2As shown, the present application proposes a pulse neural network mechanical fault diagnosis method based on STFT dimension transformation, and the present application is aimed at the problem of large storage and calculation caused by the need to process all signals in the traditional artificial neural network and deep learning method, and proposes a mechanical fault diagnosis method based on STFT dimension transformation threshold coding convolution network, which can convert one-dimensional signal into two-dimensional signal, and then map the two-dimensional signal into neuron poisson sparse coding, so that the original all signals do not need to be processed, greatly reducing the calculation and storage workload of the original signal. After poisson coding of the two-dimensional signal, the neuron is equivalent to feature sampling, and only binary data is stored in the sampling two-dimensional matrix, and only the neuron pulse response of the value 1 is processed in the subsequent threshold coding convolution network training process. The present application denoises the collected mixed noise signal, and then performs short-time Fourier transform (STFT) to convert it into a time-frequency two-dimensional matrix, which is trained based on the unsupervised learning rule of spike-timing-dependent plasticity (STDP) after neuron sampling. Finally, the test signal is identified and classified. This method can greatly improve the calculation and storage efficiency while ensuring the recognition accuracy, and is especially suitable for real-time diagnosis of mechanical faults in edge computing environment. Therefore, the present application proposes a mechanical fault diagnosis method based on STFT dimension transformation threshold coding convolution network.

[0034] The present application first decomposes the collected one-dimensional mechanical signal into high-frequency and low-frequency signals by wavelet decomposition, performs threshold processing on the high-frequency signal, and then inversely transforms the high-frequency and low-frequency signals by wavelet to obtain a denoised one-dimensional signal. Then, the denoised one-dimensional signal is converted into a time-frequency two-dimensional signal by short-time Fourier transform, and the time-frequency two-dimensional signal is sampled by neuron unit time threshold. The input data is converted into a pulse sequence using an encoder to form a poisson coding sparse matrix. Then, the sampled sparse matrix is processed by a threshold coding convolution network, and the constructed pulse neural network includes two residual blocks to prevent gradient disappearance. The signal is trained to output a weight file, and finally the test signal is identified and classified. The present application can effectively convert the signal to form a sparse sampling, reduce the problem of large amount of calculation and time-consuming in the training process, and use a residual method in the construction of the threshold coding convolution network to prevent gradient disappearance in the training process.

[0035] The present application is mainly based on the following three points:

[0036] (1) The mechanical vibration signal is prone to noise interference in the collection process, which may affect the diagnostic accuracy. Therefore, the wavelet decomposition and reconstruction technology is used to pretreat the original signal, the high-frequency noise is filtered out and the low-frequency main body characteristics are retained, so that the signal quality is effectively improved.

[0037] (2) The one-dimensional vibration signal has limited feature expression ability and low information dimension. On the basis of signal denoising, the signal is converted into a time-frequency two-dimensional matrix through short-time Fourier transform (STFT), the feature dimension is upgraded, and the time-frequency feature expression ability of the non-stationary signal is enhanced.

[0038] (3) In order to reduce the storage and calculation overhead in the training process, the HH threshold neuron is further used to sample the pulse emission of the time-frequency two-dimensional matrix, a sparse coding matrix with Poisson distribution is generated, the original feature is converted into a binary pulse sequence, the data size is significantly compressed while the key features are retained, and the foundation is laid for efficient convolution processing of the super-threshold coding convolution network. In summary, by fusing signal denoising, time-frequency transformation and sparse pulse coding, the recognition accuracy and efficiency of the fault mode in the one-dimensional mechanical vibration signal are improved while the calculation complexity is controlled.

[0039] The mechanical fault diagnosis method based on the STFT dimension transformation of the super-threshold coding convolution network comprises the following steps:

[0040] Step (1) refers to Figure 3 The one-dimensional mechanical vibration signal collected is subjected to wavelet decomposition to obtain high-frequency and low-frequency components; the high-frequency component is subjected to threshold denoising treatment, and then the low-frequency component is subjected to wavelet reconstruction to obtain the denoised one-dimensional vibration signal;

[0041] Step (2) refers to

[0042] Step (2) refers to

[0043] Step (4) refers to Figure 4, a suprathreshold encoding convolutional network structure with residual connection is constructed, including a plurality of SHH residual blocks, so as to alleviate the gradient vanishing problem and enhance the feature extraction capability of the deep layer of the network; the sparse coding matrix is input into the network, and an unsupervised learning rule based on spike time-dependent plasticity (STDP) is used for training, the synaptic weight is adaptively adjusted according to the accurate time difference of the pulse emission of the presynaptic neuron and the postsynaptic neuron, the weight of the pulse sequence with positive causal relationship is enhanced, and the weight of the irrelevant pulse sequence is inhibited;

[0044] Step (5) utilizes the trained suprathreshold encoding convolutional network to identify and classify the test signal, and outputs a fault diagnosis result.

[0045] In a feasible embodiment, the structure of the SHH residual block includes: 1 layer of convolution kernel 3*3, space-time convolution with a step of 1, 1 layer of improved HH neuron layer, 1 layer of batch normalization, short circuit connection adopts 1*1 convolution to match the dimension, and the number of neurons is consistent with the feature dimension of the input sparse matrix.

[0046] In a feasible embodiment, the attributes of the HH model include: the recommended value of the basic threshold Vth0 is-50~-40mV, the adjustment coefficient β is 0.1~0.3, and the decay time constant τth is 5~10ms; the intensity of the Gaussian white noise is set to 5%~10% of the input current, so as to avoid excessive interference with the signal characteristics.

[0047] The application has the following characteristics:

[0048] 1. The wavelet decomposition and reconstruction technology is used for pretreating the original one-dimensional mechanical vibration signal, high-frequency noise and low-frequency useful components are effectively separated, noise interference is significantly inhibited through threshold processing, and a high-quality signal basis is provided for subsequent feature extraction.

[0049] 2. The wavelet decomposition and reconstruction technology is used for pretreating the original one-dimensional mechanical vibration signal, high-frequency noise and low-frequency useful components are effectively separated, noise interference is significantly inhibited through adaptive threshold processing, and a high-quality signal basis is provided for subsequent feature extraction.

[0050] 3. The one-dimensional vibration signal is converted into a time-frequency two-dimensional matrix through short-time Fourier transform (STFT), dimension expansion from time domain to time-frequency domain is realized, the feature expression capability of the non-stationary vibration signal is enhanced, and a good foundation is provided for fully mining fault features.

[0051] 4. The Poisson sparse coding mechanism based on HH suprathreshold neurons is innovatively introduced, the pulse emission characteristics of biological neurons are simulated, the time-frequency two-dimensional matrix is converted into a binary pulse sequence, the data scale is significantly compressed, the calculation and storage overhead is reduced, and key features are highlighted.

[0052] 5、Design a kind of threshold encoding convolution network structure containing residual connection, effectively alleviate the gradient vanishing problem in deep network training, enhance the model feature learning and generalization ability, and utilize the event-driven characteristics to reduce system power consumption.

[0053] 6、Adopt the pulse time-dependent plasticity (STDP) rule for network training, so that the model can adaptively learn the timing characteristics and fault patterns in the vibration signal. This mechanism has event-driven and unsupervised learning characteristics, and can efficiently capture timing characteristics and fault modes by adaptively adjusting weights through precise timing of pulses, reduce the dependence on labeled data, and improve the generalization ability in noisy environments. The STDP mechanism weight update only depends on the local pulse time, and does not need global error backpropagation, which significantly reduces the computational energy consumption and hardware implementation complexity, and is especially suitable for online real-time fault diagnosis on edge computing devices.

[0054] 7、This method has the advantages of efficient signal processing and low power consumption calculation, and is not only suitable for mechanical fault diagnosis in high-noise industrial environments, but also provides a practical solution for real-time fault monitoring on edge computing devices.

[0055] Embodiment one

[0056] Step (1) Collect one-dimensional vibration signals of mechanical equipment , the sampling frequency is , and the signal length is N. In order to suppress noise interference and retain fault features, first, the original signal is decomposed by discrete wavelet transform (DWT), and the expression is as follows:

[0057] (1)

[0058] Among them, is the Jth layer approximation coefficient (low-frequency component), is the jth layer detail coefficient (high-frequency component), and J is the decomposition layer number, which is selected as J=4 according to the signal characteristics. The high-frequency detail coefficient is denoised by using an improved threshold function:

[0059] (2)

[0060] The threshold is determined adaptively:

[0061] (3)

[0062] Among them, is the noise standard deviation estimate of the jth layer detail coefficient, which is estimated by using the median absolute deviation (MAD):

[0063] (4)

[0064] de-noised high frequency component and the reserved low frequency component de-noised one-dimensional vibration signal is obtained by wavelet reconstruction :

[0065] (5)

[0066] wherein, denotes the wavelet inverse transform operator. The signal-to-noise ratio (SNR) of the reconstructed signal is significantly improved, laying a foundation for subsequent feature extraction.

[0067] wherein, when the gear rotational speed < 1000 r / min, J = 3; 1000 ~ 3000 r / min, J = 4; > 3000 r / min, J = 5. The higher the rotational speed, the more complex the high frequency component of the signal, and more layers are needed to separate the noise.

[0068] Step (2) is to enhance the feature expression ability of the non-stationary vibration signal, and the de-noised one-dimensional signal is converted into a time-frequency two-dimensional matrix by short-time Fourier transform (STFT). STFT realizes local spectral analysis of the signal by sliding the window function on the time axis, and its mathematical expression is:

[0069] (6)

[0070] wherein, is a window function, and a Hamming window (Hamming Window) is selected in this paper, and its expression is:

[0071] (7)

[0072] The window length N is adaptively determined according to the signal characteristics:

[0073] (8)

[0074] wherein, is the sampling frequency, is the highest frequency component of the signal. For a discrete signal , the discrete form of STFT is:

[0075] (9)

[0076] In a feasible embodiment, when the sampling frequency is 2000 Hz and the highest frequency is 1000 Hz, the window length = 2000 / 1000 = 200 points, and the frame shift = 50 points.

[0077] where m is the time frame index, k is the frequency bin index, R is the frame shift, R = N / 4 is taken, and N is the FFT point number, N = 1024 is taken. In order to enhance feature visualization and subsequent processing effect, the time-frequency two-dimensional matrix is taken modulo and logarithmically scaled:

[0078] (10)

[0079] where, is a small constant to avoid taking the logarithm of zero.

[0080] Step (3) takes the time-frequency two-dimensional matrix S(m, k) obtained in step (2) as input, and performs pulse sparse coding through an improved Hodgkin-Huxley (HH) threshold neuron model to generate a binary pulse sequence P(t, n) ∈ {0, 1}T×N, where N is the number of neurons. Each element of the time-frequency two-dimensional matrix is converted into the input current of the neuron:

[0081] (11)

[0082] where g is a scaling factor, μ is the mean of S, σ is the standard deviation of S, and I bias is a bias current constant.

[0083] The improved HH threshold neuron model introduces an adaptive firing threshold and a noise robustness mechanism based on the classic HH model, and its mathematical expression is as follows:

[0084] (12)

[0085] where V is the cell membrane voltage; g Na, g K, and g L represent the maximum values of the sodium ion channel, potassium ion channel, and leakage current channel conductance, respectively; E Na, E K, and E L represent the corresponding channel reversal potentials; C m is the cell membrane capacitance; m and h are the gating variables of the sodium ion channel, and n is the gating variable of the potassium ion channel; α x (V) and β x (V) are the activation and inactivation functions of the gating variable x, respectively. The external input current I ext is obtained by the aforementioned conversion, and Inoise is the introduced Gaussian white noise. The improvement lies in the introduction of a dynamic threshold Vth(t), where Vth0 is the basic threshold, β is the threshold adjustment coefficient, ts is the pulse firing time, and τth is the threshold decay time constant. This mechanism enables the neuron to have adaptive firing characteristics and enhances the coding ability of the time-frequency features.

[0086] When the membrane potential V(t) ≥ Vth(t), the neuron fires a pulse, the membrane potential is reset to Vreset, and the threshold is temporarily increased and then exponentially decays, effectively preventing over-firing and achieving sparse coding. In order to simulate the inhibitory postsynaptic potential between neurons, the following mechanism is introduced:

[0087] (13)

[0088] where t, is the pre-neuron spike time, w, is the synaptic weight, and τsynis the synaptic time constant. The spike frequency is counted in a time window (100 ms) to obtain the binary spike sequence P(t, n), which realizes the Poisson sparse coding.

[0089] Step (4) The pulse sequence P(t, n) generated in step (3) is input into the SHHResNet network structure for feature learning and training. Preferably, the HH neuron is processed by a max-pooling layer, the network contains multiple SHHResidualBlock, and the pulse time-dependent plasticity (STDP) rule is used for unsupervised training. The input pulse sequence is first extracted by a spatio-temporal convolution layer:

[0090] (14)

[0091] where * represents the spatio-temporal convolution operation, W(l) is the convolution kernel weight, and b(l) is the bias. The convolution result is then processed by an improved HH supra-threshold neuron layer to generate a pulse:

[0092] (15)

[0093] The dynamic process of the gating variables m, h, and n is consistent with step (3). When V(l) ≥ Vth(l), a pulse V(l) = 1 is generated and the membrane potential is reset. The residual connection is represented as:

[0094] (16)

[0095] where F is a nonlinear transformation (including convolution and HH neuron spike generation) in the residual block, and Ws is the weight matrix of the short connection. The network training uses the STDP rule to adjust the convolution kernel weight, and the weight update formula is as follows:

[0096] (17)

[0097] where η+ and η− are the reinforcement and inhibition learning rates, respectively, τ+ and τ− are the corresponding time constants, t, and tj are the spike times of the pre-neuron and post-neuron, respectively. The global weight update formula is:

[0098] (18)

[0099] where α = 0.001 is the global learning rate, and λ = 0.0001 is the weight decay coefficient. The trained weights are saved.

[0100] Step (5) uses the suprathreshold coding convolutional network trained in step (4) to identify and classify the test signal. The mechanical fault signal of the test signal is input into the trained network, and the pulse firing activity of the output layer is obtained through forward propagation. The pulse firing rate of the output layer neurons is calculated as follows:

[0101] (19)

[0102] in, This represents the average firing rate of the k-th output neuron within the time window T. ∈{0,1} represents the firing state of the neuron at time t. The fault type decision is based on the firing rate of the fault diagnosis output layer neurons, employing a winner-take-all strategy:

[0103] (20)

[0104] Where y represents the final fault diagnosis result, indicating that it is identified as a type k fault. To evaluate the reliability of the classification results, the relative pulse firing rate difference of the output layer is calculated:

[0105] (twenty one)

[0106] When the confidence level is below the threshold of 0.3, the classification result is deemed unreliable and the signal needs to be re-acquired for analysis.

[0107] There are four categories in this invention: normal signal, inner ring fault, outer ring fault, and rolling element fault.

[0108] In one feasible embodiment, the accuracy is 95.8% to 99.2% on a gear vibration dataset with a signal-to-noise ratio of 5 to 20 dB. Compared with traditional CNN, the computation time is reduced from 0.5 s to 0.35 s, and the storage requirement is reduced from 100 MB to 40 MB, ensuring accuracy and improving efficiency.

[0109] The embodiment of the mechanical fault diagnosis device based on STFT dimensional transformation of the present invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 2As shown in the figure, it is a hardware structure diagram of the apparatus for mechanical fault diagnosis based on the STFT dimension transformation threshold coding convolution network of the present application in any data processing capable device, in addition to Figure 2 In addition to the processor, the memory, the network interface, and the non-volatile memory shown in the figure, any data processing capable device in which the apparatus in the embodiment is generally also provided with other hardware according to the actual functions of the data processing capable device, and no further description is given here. The implementation process of the functions and roles of each unit in the apparatus is specifically described in the implementation process of the corresponding steps in the above method, and no further description is given here.

[0110] For the apparatus embodiment, since it basically corresponds to the method embodiment, the relevant part is described in the method embodiment. The apparatus embodiment described above is only schematic, and the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, that is, they can be located in one place, or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present application according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0111] The embodiment of the present application also provides a computer readable storage medium, which stores a program, and the program is executed by a processor to realize the apparatus for mechanical fault diagnosis based on the STFT dimension transformation threshold coding convolution network in the above embodiment.

[0112] The computer readable storage medium can be an internal storage unit of any data processing capable device, such as a hard disk or a memory. The computer readable storage medium can also be an external storage device of any data processing capable device, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. Further, the computer readable storage medium can include both the internal storage unit and the external storage device of any data processing capable device. The computer readable storage medium is used to store the computer program and other programs and data required by the data processing capable device, and can also be used to temporarily store data that has been output or will be output.

[0113] It should be noted that: the above description is only the preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. within the principle of the present application, should be included in the protection scope of the present application.

Claims

1. A mechanical fault diagnosis method based on STFT dimensional transformation spiking neural network, characterized in that, Includes the following steps: One-dimensional mechanical vibration signal is acquired and decomposed by wavelet to obtain high-frequency and low-frequency components. The low-frequency components and the denoised high-frequency components are then reconstructed by wavelet to obtain the denoised one-dimensional vibration signal. The denoised one-dimensional vibration signal is subjected to a short-time Fourier transform to convert it into a time-frequency two-dimensional matrix; The time-frequency two-dimensional matrix is ​​input into the improved HH suprathreshold neuron model to construct a pulse firing mechanism with an adaptive threshold. The time-frequency two-dimensional matrix is ​​then subjected to Poisson sparse coding to generate a binary pulse sequence. Only the salient features of the signal are processed by pulse response to form a sparse coding matrix. Construct an on-threshold coding convolutional network with residual connections, including several SHH residual blocks. Input a sparse coding matrix into the network and train the on-threshold coding convolutional network using an unsupervised learning rule based on pulse time-dependent plasticity. Adaptively adjust the network synaptic weights according to the time difference between the pulse firing of presynaptic neurons and postsynaptic neurons. The mechanical vibration signal to be tested is input into a trained suprathreshold coding convolutional network. The pulse firing activity of the output layer neurons is obtained through the forward propagation of the network, and the fault diagnosis result is determined based on the pulse firing activity. The structure of the SHH residual block includes: 1 layer of 3×3 convolutional kernel, spatiotemporal convolution with stride 1, 1 layer of improved HH neuron layer, 1 layer of batch normalization, and short-circuit connections using 1×1 convolutional matching dimension. The number of neurons is consistent with the feature dimension of the input sparse matrix. The improved HH suprathreshold neuron model includes the introduction of an adaptive dynamic threshold and noise robustness mechanism on the basis of the classic HH model. The adaptive dynamic threshold includes a base threshold, a threshold adjustment coefficient, and a threshold decay time constant. When the neuron membrane potential reaches the adaptive dynamic threshold, the neuron fires a pulse, and the membrane potential is reset. The adaptive dynamic threshold increases instantaneously and then decays exponentially. The noise robustness mechanism is achieved by introducing Gaussian white noise to improve the model's tolerance to noise.

2. The method according to claim 1, characterized in that, The denoising process for the high-frequency components includes an adaptive threshold function. The adaptive threshold is determined by calculating the noise standard deviation of the high-frequency components, which is obtained by calculating the median absolute deviation method.

3. The method according to claim 1, characterized in that, The step of performing a short-time Fourier transform on the denoised one-dimensional vibration signal includes using a window function to perform local spectrum analysis of the signal. The window length of the window function is adaptively determined according to the signal characteristics of the denoised one-dimensional vibration signal. The signal characteristics include the sampling frequency of the denoised one-dimensional vibration signal and the highest frequency component of the signal. After performing modulo operation on the transformed time-frequency two-dimensional matrix, logarithmic scaling is then applied to avoid taking the logarithm of zero and to enhance the visualization of features.

4. The method according to claim 1, characterized in that, When the time-frequency two-dimensional matrix is ​​input into the improved HH suprathreshold neuron model, each element of the time-frequency two-dimensional matrix is ​​converted into the input current of the neuron, and the input current is standardized by scaling factor, mean and standard deviation of the time-frequency two-dimensional matrix, and bias current constant.

5. The method according to claim 1, characterized in that, The construction of the above-threshold coding convolutional network with residual connections includes the following: obtaining the residual connections of the above-threshold coding convolutional network by combining the nonlinear transformation results in the residual block with the weight matrix of the short-circuit connection, wherein the nonlinear transformation includes convolution operation and improved above-threshold neuron spiking process.

6. The method according to claim 1, characterized in that, The unsupervised learning rule based on pulse time-dependent plasticity trains the suprathreshold coding convolutional network and adaptively adjusts the network synaptic weights, including the following: weight enhancement is performed on the pulse sequences of anterior and posterior synaptic neurons with positive causal relationships, and weight suppression is performed on the pulse sequences of unrelated anterior and posterior synaptic neurons; the weight update depends only on the local pulse firing time of the anterior and posterior synaptic neurons.

7. The method according to claim 1, characterized in that, The process of determining the fault diagnosis result based on the pulse firing activity includes: calculating the average pulse firing rate of each neuron in the output layer within a set time window; adopting a winner-take-all strategy, taking the fault type corresponding to the neuron with the highest average pulse firing rate as the final fault diagnosis result; and simultaneously calculating the relative pulse firing rate difference of the neurons in the output layer. When the relative pulse firing rate difference is lower than a set threshold, the fault diagnosis result is determined to be unreliable, and the mechanical vibration signal to be tested needs to be re-acquired for processing and diagnosis.

8. A mechanical fault diagnosis device based on STFT dimensional transformation pulse neural network, characterized in that: The device includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the mechanical fault diagnosis method based on STFT dimension transformation according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that: It stores a program that, when executed by a processor, implements the mechanical fault diagnosis method based on STFT dimension transformation according to any one of claims 1-7.

Images