Reloading port machine fault identification method based on homologous heterogeneous characteristic guide laminated pulse network
Through homogeneous heterogeneous trait characterization and optimization of stacked vein network model, the problem that traditional fault identification methods are difficult to identify complex faults is solved, and high-precision fault identification and quantitative characterization are achieved.
Patent Information
- Application Number
- CN202411736551.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-29
- Publication Date
- 2025-03-25
AI Technical Summary
Traditional reinstalled port fault identification methods are difficult to meet the needs of high-precision fault identification, especially when facing complex fault patterns and coupling and nonlinear features.
Through homogeneous heterogeneous characteristics characterization, complementary characteristics construction, fine-grained decomposition and reconstruction, and the establishment and optimization of stacked pulse network models, accurate identification of reinstalled port machine faults and quantitative characterization of fault degree are achieved.
It realizes accurate identification of complex fault types and quantitative characterization of fault degree, and provides a new high-precision reinstallation port machine structure fault identification method.
Smart Images
Figure CN119691545A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of high-precision heavy-duty port machinery fault identification, and in particular to a heavy-duty port machinery fault identification method based on homologous heterogeneous characteristic guided laminated vein network. Background Art
[0002] With the increasing complexity of industrial heavy-duty port machinery and engineering systems, traditional heavy-duty port machinery fault identification methods face severe challenges and can no longer meet the needs of high-precision fault identification. Traditional methods usually rely on a single data source for collection and analysis, and it is difficult to comprehensively and accurately identify the complex fault forms of the internal structure of heavy-duty port machinery, especially faults with coupling and nonlinear characteristics. This limitation poses a potential threat to the safety and performance stability of heavy-duty port machinery.
[0003] In this context, how to effectively characterize the diverse fault characteristics of heavy-duty port machinery structures is a problem currently faced. Therefore, it is necessary to propose a method that can accurately identify complex fault types and quantitatively characterize their fault degrees. Summary of the invention
[0004] Purpose of the invention: The purpose of the present invention is to provide a method for identifying heavy-duty port machinery faults based on homologous heterogeneous characteristic-guided laminated vein networks.
[0005] Technical solution: The present invention comprises the following steps:
[0006] (1) Characterization of the homogeneous and heterogeneous characteristics of heavy-duty port machinery;
[0007] (2) Construction of complementary homologous traits;
[0008] (3) fine-grained decomposition and reconstruction of fusion homologous and heterogeneous features;
[0009] (4) Establishment and optimization of the laminated vein network model;
[0010] (5) Fault identification of heavy-duty port machinery based on the stacked network model.
[0011] Furthermore, the step (1) comprises:
[0012] According to the operating amplitude state of the heavy-duty port machinery structure, the homologous vibration information of the heavy-duty port machinery is firstly acquired by collecting the target signal, and then the homologous and heterogeneous characteristics are characterized based on the time-frequency energy spectrum.
[0013] Further, the obtaining of homologous vibration information of the heavy-duty port machinery includes:
[0014] For the target signal The corresponding intrinsic mode function is:
[0015]
[0016] The time derivative of the phase modulation function of the target signal is solved, and the instantaneous frequency is:
[0017]
[0018] Among them, W n (t) represents the amplitude modulation function, φ n (t) represents the phase modulation function and ε(n) represents the noise term.
[0019] Furthermore, the isomeric characterization includes:
[0020] Based on the scale parameter and the time offset q, the wavelet coefficients W are completed. n Definition of (p, q):
[0021]
[0022] Wavelet coefficients W n (p, q) redistributes energy along the frequency axis to reduce blurring effects,
[0023] Map the time scale plane to [(q, p) → (q, W n (p, q))] to obtain the time-frequency energy spectrum E, the corresponding time-frequency energy spectrum E is:
[0024]
[0025] Among them, a n represents the discrete scale, W n (a n , q) represents the local energy distribution, Δω represents the frequency resolution, (Δp) n represents the scale interval, frequency ω l is the frequency center.
[0026] Furthermore, the step (2) comprises:
[0027] According to the heterogeneous characteristics of the heavy-duty port machinery extracted in step (1), the time-frequency energy spectrum E represents the local performance of the signal, and the phase space matrix constructs the complementary heterogeneous characteristics in the state space through its global characteristics.
[0028] For the target signal Normalized, the corresponding standard normal distribution Δ is expressed as:
[0029] Δ=υ1,υ2,υ2,…υ g , and
[0030] Among them, ri is the value of the timestamp at i, and the sequence length is g. is the average value, is the standard deviation;
[0031] The dimension of the time series is reduced to ξ by the piecewise aggregation approximation method, and the new smoothed time series S is expressed as:
[0032] S=s1,s2,s3,…s m , and
[0033] Among them, k is the reduction factor, and m is the length of the time series after dimensionality reduction;
[0034] The average value s of the corresponding interval in the original time series j It is expressed as:
[0035]
[0036] Among them, j is the jth data point in the original time series, ξ is the dimension of the time series;
[0037] Based on the smoothed time series i Construct the relative position matrix M, and the corresponding relative position matrix M is expressed as:
[0038]
[0039] Each row and column of the relative position matrix M corresponds to a specific timestamp;
[0040]
[0041] The matrix M is converted into a phase space matrix N on a uniform plane through Min-Max normalization, and the isomeric properties are constructed based on the phase space matrix N.
[0042] Furthermore, the step (3) comprises:
[0043] According to the complementary isomeric traits determined in step (2), the phase space matrix N is combined with the time-frequency energy spectrum E to construct a fusion isomeric trait representation Θ = [E, N],
[0044] A two-dimensional convolutional layer is used to perform fine-grained decomposition of the time-frequency energy spectrum E and phase space matrix N in the fusion homologous heterogeneous trait representation Θ.
[0045] G E =BatchNorm(Conv2D(E, k E ))=BatchNorm(φ(W E *E+b E ))
[0046] G N =BatchNorm(Conv2D(N,k N ))=BatchNorm(φ(W N *N+b N ))
[0047] Among them, BatchNorm represents batch normalization, which is used to reduce internal covariate shift. Conv2D represents a two-dimensional convolution operation, k E and k N W denotes the convolution kernel size of the time-frequency energy spectrum E and the phase space matrix N, φ(·) is the nonlinear activation function of the element ReLU, and * denotes the convolution operator. E and W N denote the convolution kernel of the time-frequency energy spectrum E and the phase space matrix N, respectively, and b E and b N They represent the bias terms of the time-frequency energy spectrum E and the phase space matrix N respectively;
[0048] After fine-grained decomposition, G E and G N Reconstruction is performed and the time step information of the time-frequency energy spectrum E and the phase space matrix N is retained by position embedding.
[0049] Furthermore, the time step information of retaining the time-frequency energy spectrum E and the phase space matrix N by position embedding is expressed as:
[0050]
[0051] Among them, PE(T,d f ) represents mapping the time step T to the feature dimension d f The position embedding function. and represents the initial characteristic information of the time-frequency energy spectrum E and phase space matrix N before position embedding, F E and F N It represents the reconstructed characteristic information of the time-frequency energy spectrum E and the phase space matrix N after position embedding, LayerNorm represents layer normalization, and STencoder represents the reconstruction operation.
[0052] Furthermore, the step (4) comprises:
[0053] According to step (3), the fusion homologous heterogeneous trait representation Θ = [E, N] is fine-grained decomposition and reconstruction to obtain the reconstructed trait information F after position embedding E and F N , establish a layered vein network model:
[0054]
[0055] Among them, DWConv(·) represents the operation of depthwise separable convolution, and Conv(·) represents the convolution operation.
[0056] The trait information processed by convolution is linearly projected into a unified trait space, which promotes the establishment and optimization of the layered vein network model;
[0057]
[0058] Where L(·) represents the linear projection function, F add Represents the iterative output of the stacked vein network model.
[0059] Furthermore, the stacked vein network model is optimized by using the tanh activation function to add Perform nonlinear transformation and generate optimized output by softmax function.
[0060] ψ s =softmax(tanh(F add ))
[0061] Among them, ψ s Represents the final output of the optimized stacked vein network model.
[0062] Furthermore, the step (5) comprises:
[0063] According to the stacked network model established in step (4), the fault type of the heavy-duty port machinery structure is identified. The input characteristics are linearly transformed through the first fully connected layer of the stacked network model, and nonlinearly transformed using the ReLU activation function. The characteristics of the intermediate hidden layer are defined as:
[0064] H fc =ReLU(W1·ψ s +b1)
[0065] Among them, W1 and b1 are the weight matrix and bias vector of the first fully connected layer of the stacked vein network model respectively;
[0066] The characteristics of the middle hidden layer are passed through the second fully connected layer of the stacked neural network model and a softmax activation function to obtain the final result for heavy-duty port machinery fault identification:
[0067]
[0068] Among them, W κ and b κ are the weight matrix and bias vector of the softmax layer respectively.
[0069] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: through in-depth mining of homologous and heterogeneous characteristics, the diversified fault characteristics of heavy-duty port machinery structures are effectively characterized. Under the guidance of homologous and heterogeneous characteristics, a unique laminated vein network model is constructed, which can realize the accurate identification of complex fault types and quantitative characterization of their fault degrees, and provide a new high-precision heavy-duty port machinery structure fault identification method, which has important practical research significance. BRIEF DESCRIPTION OF THE DRAWINGS
[0070] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0071] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.
[0072] like Figure 1 As shown, the present invention comprises the following steps:
[0073] (1) Characterization of heterogeneous properties of heavy-duty port machinery
[0074] According to the operating amplitude state of the heavy-duty port machinery structure, the homologous vibration information of the heavy-duty port machinery is firstly acquired by collecting the target signal, and then the homologous heterogeneous characteristics are characterized based on the time-frequency energy spectrum:
[0075] For the target signal The corresponding intrinsic mode function is
[0076]
[0077] Among them, W n (t) represents the amplitude modulation function, φ n (t) represents the phase modulation function and ε(n) represents the noise term.
[0078] The time derivative of the phase modulation function of the target signal is solved, and the instantaneous frequency is
[0079]
[0080] Based on the scale parameter and the time offset q, the wavelet coefficients W are completed. n Definition of (p, q):
[0081]
[0082] Wavelet coefficients W n (p, q) is redistributed along the frequency axis to alleviate the blurring effect.
[0083] Map the time scale plane to [(q, p) → (q, W n (p, q))] to obtain the time-frequency energy spectrum E, the corresponding time-frequency energy spectrum E is
[0084]
[0085] Among them, a n represents the discrete scale, W n (a n , q) represents the local energy distribution, Δω represents the frequency resolution, (Δp) n represents the scale interval, frequency ω l is the frequency center.
[0086] (2) Construction of complementary homologous isomers
[0087] According to the heterogeneous characteristics of the heavy-duty port machinery extracted in the above step (1), considering that the time-frequency energy spectrum E mainly represents the local performance of the signal, the phase space matrix constructs complementary heterogeneous characteristics in the state space through its global characteristics.
[0088] For the target signal After normalization, the corresponding standard normal distribution Δ is expressed as
[0089] Δ=υ1,υ2,υ2,…υ g , and
[0090] Among them, ri is the value of the timestamp at i, and the sequence length is g. is the average value, is the standard deviation.
[0091] The dimension of the time series is reduced to ξ by the piecewise aggregation approximation method, and the new smoothed time series S is expressed as
[0092] S=s1,s2,s3,…s m , and
[0093] Among them, k is the reduction factor and m is the length of the time series after dimensionality reduction.
[0094] The average value s of the corresponding interval in the original time series i Expressed as
[0095]
[0096] Among them, j is the jth data point in the original time series, and ξ is the dimension of the time series.
[0097] Based on the smoothed time series j Construct the relative position matrix M, the corresponding relative position matrix M is expressed as
[0098]
[0099] Each row and column of the relative position matrix M corresponds to a specific timestamp. The matrix M is converted into a phase space matrix N on a uniform plane through Min-Max normalization, and the homologous heterogeneous characteristics are constructed based on the phase space matrix N:
[0100]
[0101] (3) Fine-grained decomposition and reconstruction of homologous and heterogeneous features
[0102] According to the complementary isomeric traits determined in the above step (2), the phase space matrix N is combined with the time-frequency energy spectrum E to construct a fused isomeric trait representation Θ = [E, N].
[0103] A two-dimensional convolutional layer is used to perform fine-grained decomposition of the time-frequency energy spectrum E and phase space matrix N in the fused homologous and heterogeneous trait representation Θ.
[0104] G E =BatchNorm(Conv2D(E, k E ))=BatchNorm(φ(W E *E+b E ))
[0105] G N =BatchNorm(Conv2D(N,k N ))=BatchNorm(φ(W N *N+b N ))
[0106] Among them, BatchNorm represents batch normalization, which is used to reduce internal covariate shift. Conv2D represents a two-dimensional convolution operation, k E and k N denote the convolution kernel size of the time-frequency energy spectrum E and the phase space matrix N respectively. φ(·) is the nonlinear activation function of the element ReLU, and * denotes the convolution operator. W E and W N denote the convolution kernel of the time-frequency energy spectrum E and the phase space matrix N, respectively, and b E and b N They represent the bias terms of the time-frequency energy spectrum E and the phase space matrix N respectively.
[0107] After fine-grained decomposition, G E and G N Reconstruction is performed and the time step information of the time-frequency energy spectrum E and the phase space matrix N is retained by position embedding.
[0108]
[0109] Among them, PE(T,d f ) represents mapping the time step T to the feature dimension d f The position embedding function. and represents the initial characteristic information of the time-frequency energy spectrum E and phase space matrix N before position embedding, F E and F N It represents the reconstructed characteristic information of the time-frequency energy spectrum E and the phase space matrix N after position embedding, LayerNorm represents layer normalization, and STencoder represents the reconstruction operation.
[0110] (4) Establishment and optimization of the laminated vein network model
[0111] According to the above step (3), the fusion homologous heterogeneous trait representation Θ = [E, N] is fine-grained decomposition and reconstruction to obtain the reconstructed trait information F after position embedding E and F N In order to deeply explore the deep homologous and heterogeneous characteristic information, a layered vein network model was established.
[0112]
[0113] Among them, DWConv(·) represents the operation of depthwise separable convolution, and Conv(·) represents the convolution operation.
[0114] The trait information processed by convolution is linearly projected into a unified trait space to facilitate the establishment of the stacked vein network model.
[0115]
[0116] Where L(·) represents the linear projection function, F add Represents the iterative output of the stacked vein network model.
[0117] The stacked vein network model is optimized and the tanh activation function is used to optimize F add A nonlinear transformation is performed and the optimized output is generated by the softmax function.
[0118] ψ s =softmax(tanh(F add ))
[0119] Among them, ψ s Represents the final output of the optimized stacked vein network model.
[0120] (5) Fault identification of heavy-duty port machinery based on the stacked network model
[0121] According to the stacked network model established in step (4) above, the fault type of the heavy-duty port machinery structure is identified. The input characteristics are linearly transformed through the first fully connected layer of the stacked network model, and then nonlinearly transformed using the ReLU activation function. The characteristics of the intermediate hidden layer are defined as:
[0122] H fc =ReLU(W1·ψ s +b1)
[0123] Among them, W1 and b1 are the weight matrix and bias vector of the first fully connected layer of the stacked vein network model respectively.
[0124] The characteristics of the middle hidden layer are passed through the second fully connected layer of the stacked neural network model and a softmax activation function to obtain the final result for heavy-duty port machinery fault identification:
[0125]
[0126] Among them, W κ and b κ are the weight matrix and bias vector of the softmax layer respectively.
Claims
1. A method for identifying faults of heavy-duty port machinery based on homogeneous heterogeneous characteristics guided laminated vein network, characterized in that: The method comprises the following steps: (1) Characterization of the homogeneous and heterogeneous characteristics of heavy-duty port machinery; (2) Construction of complementary homologous traits; (3) fine-grained decomposition and reconstruction of fusion homologous and heterogeneous features; (4) Establishment and optimization of the laminated vein network model; (5) Fault identification of heavy-duty port machinery based on the stacked network model.
2. The method for identifying heavy-duty port machinery faults based on homogeneous heterogeneous characteristic guided laminated vein network according to claim 1 is characterized in that: The step (1) comprises: According to the operating amplitude state of the heavy-duty port machinery structure, the homologous vibration information of the heavy-duty port machinery is firstly acquired by collecting the target signal, and then the homologous and heterogeneous characteristics are characterized based on the time-frequency energy spectrum.
3. The method for identifying heavy-duty port machinery faults based on homogeneous heterogeneous characteristic guided laminated vein network according to claim 2 is characterized in that: The obtaining of homologous vibration information of the heavy-duty port machinery comprises: For the target signal The corresponding intrinsic mode function is: The time derivative of the phase modulation function of the target signal is solved, and the instantaneous frequency is: Among them, W n (t) represents the amplitude modulation function, φ n (t) represents the phase modulation function and ε(n) represents the noise term.
4. The method for identifying heavy-duty port machinery faults based on homogeneous heterogeneous characteristic guided laminated vein network according to claim 2 is characterized in that: The isomeric characterizations include: Based on the scale parameter and the time offset q, the wavelet coefficients W are completed n Definition of (p, q): Wavelet coefficients W n (p, q) redistributes energy along the frequency axis to reduce blurring effects, Map the time scale plane to [(q, p) → (q, W n (p, q))] to obtain the time-frequency energy spectrum E, the corresponding time-frequency energy spectrum E is: Among them, a n represents the discrete scale, W n (a n , q) represents the local energy distribution, Δω represents the frequency resolution, (Δp) n represents the scale interval, frequency ω l is the frequency center.
5. The method for identifying heavy-duty port machinery faults based on homogeneous heterogeneous characteristic guided laminated vein network according to claim 1 is characterized in that: The step (2) comprises: According to the heterogeneous characteristics of the heavy-duty port machinery extracted in step (1), the time-frequency energy spectrum E represents the local performance of the signal, and the phase space matrix constructs the complementary heterogeneous characteristics in the state space through its global characteristics. For the target signal After normalization, the corresponding standard normal distribution Δ is expressed as: Among them, r i is the value of the timestamp at i, the sequence length is g, is the average value, is the standard deviation; The dimension of the time series is reduced to ξ by the piecewise aggregation approximation method, and the new smoothed time series S is expressed as: Among them, k is the reduction factor, and m is the length of the time series after dimensionality reduction; The average value s of the corresponding interval in the original time series i It is expressed as: Among them, j is the jth data point in the original time series, ξ is the dimension of the time series; Based on the smoothed time series i Construct the relative position matrix M, and the corresponding relative position matrix M is expressed as: Each row and column of the relative position matrix M corresponds to a specific timestamp; The matrix M is converted into a phase space matrix N on a uniform plane through Min-Max normalization, and the isomeric properties are constructed based on the phase space matrix N.
6. The method for identifying heavy-duty port machinery faults based on homogeneous heterogeneous characteristic guided laminated vein network according to claim 1 is characterized in that: The step (3) comprises: According to the complementary isomeric traits determined in step (2), the phase space matrix N is combined with the time-frequency energy spectrum E to construct a fusion isomeric trait representation Θ = [E, N], A two-dimensional convolutional layer is used to perform fine-grained decomposition of the time-frequency energy spectrum E and phase space matrix N in the fusion homologous heterogeneous trait representation Θ. G E =BatchNorm(Conv2D(E,k E ))=BatchNorm(φ(W E *E+b E )) G N =BatchNorm(Conv2D(N,k N ))=BatchNorm(φ(W N *N+b N )) Among them, BatchNorm represents batch normalization, which is used to reduce internal covariate shift. Conv2D represents a two-dimensional convolution operation, k E and k N denote the convolution kernel size of the time-frequency energy spectrum E and the phase space matrix N, respectively, φ(·) is the nonlinear activation function of the element ReLU, and * denotes the convolution operator. W E and W N denote the convolution kernel of the time-frequency energy spectrum E and the phase space matrix N, respectively, and b E and b N They represent the bias terms of the time-frequency energy spectrum E and the phase space matrix N respectively; After fine-grained decomposition, G E and G N Reconstruction is performed and the time step information of the time-frequency energy spectrum E and the phase space matrix N is retained by position embedding.
7. The method for identifying heavy-duty port machinery faults based on homogeneous heterogeneous characteristic guided laminated vein network according to claim 6 is characterized in that: The time step information of retaining the time-frequency energy spectrum E and the phase space matrix N by position embedding is expressed as: Among them, PE(T,d f ) represents mapping the time step T to the feature dimension d f The position embedding function. and represents the initial characteristic information of the time-frequency energy spectrum E and phase space matrix N before position embedding, F E and F N It represents the reconstructed characteristic information of the time-frequency energy spectrum E and the phase space matrix N after position embedding, LayerNorm represents layer normalization, and STencoder represents the reconstruction operation.
8. The method for identifying heavy-duty port machinery faults based on homogeneous heterogeneous characteristic guided laminated vein network according to claim 1 is characterized in that: The step (4) comprises: According to step (3), the fusion homologous heterogeneous trait representation Θ = [E, N] is fine-grained decomposition and reconstruction to obtain the reconstructed trait information F after position embedding E and F N , establish a layered vein network model: Among them, DWConv(·) represents the operation of depthwise separable convolution, and Conv(·) represents the convolution operation. The trait information processed by convolution is linearly projected into a unified trait space, which promotes the establishment and optimization of the layered vein network model; Where L(·) represents the linear projection function, F add Represents the iterative output of the stacked vein network model.
9. The method for identifying heavy-duty port machinery faults based on homogeneous heterogeneous characteristic guided laminated vein network according to claim 8 is characterized in that: The stacked vein network model is optimized and the tanh activation function is used to optimize F add Perform nonlinear transformation and generate optimized output by softmax function. ψ s =softmax(tanh(F add )) Among them, ψ s Represents the final output of the optimized stacked vein network model.
10. The method for identifying heavy-duty port machinery faults based on homogeneous heterogeneous characteristic guided laminated vein network according to claim 1 is characterized in that: The step (5) comprises: According to the stacked network model established in step (4), the fault type of the heavy-duty port machinery structure is identified. The input characteristics are linearly transformed through the first fully connected layer of the stacked network model, and nonlinearly transformed using the ReLU activation function. The characteristics of the intermediate hidden layer are defined as: H fc =ReLU(W1·ψ s +b1) Among them, W1 and b1 are the weight matrix and bias vector of the first fully connected layer of the stacked vein network model respectively; The characteristics of the middle hidden layer are passed through the second fully connected layer of the stacked neural network model and a softmax activation function to obtain the final result for heavy-duty port machinery fault identification: Among them, W κ and b κ are the weight matrix and bias vector of the softmax layer respectively.