Time-varying working condition fault diagnosis method capable of explaining deep learning driving, equipment and medium

Abstract

The invention discloses a time-varying working condition fault diagnosis method capable of explaining deep learning driving, equipment and a medium. According to the method, a convolutional sparse intrinsic mode expansion network is provided, a convolutional sparse intrinsic mode dictionary is constructed based on an impact failure mechanism, a fast coding solution algorithm is designed and expanded, a network encoder is constructed, and intrinsic mode sparse coding is realized through multiple channels; and the decoder performs weighted stacking on the intrinsic mode impact response of each channel, and outputs time-varying fault impact characteristics with clear physical significance. According to the method, a fault mechanism is embedded in feature extraction, interpretable generalization of an unknown time-varying working condition can be realized without multi-source domain data, and the diagnosis accuracy and credibility of the mechanical fault under the time-varying working condition are remarkably enhanced.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent fault diagnosis and signal processing, and particularly relates to an interpretable deep learning driven time-varying working condition fault diagnosis method, device and medium. BACKGROUND

[0002] Rotary machines are key transmission components for power transmission and motion control, and are widely used in industrial robots, aerospace, precision machine tools and other high-end equipment. Rotary machines are long-term operated in complex working conditions such as time-varying speed and variable load, and are prone to faults such as thin-wall bearing inner / outer ring pitting and spalling. If not diagnosed in time and accurately, it will directly affect the operation accuracy and reliability of the whole machine, and even cause serious safety accidents and cause significant economic losses.

[0003] With the rapid development of artificial intelligence technology, deep learning based fault diagnosis methods have achieved remarkable results. Such methods adaptively extract fault features through network structure design and gradient backpropagation, and realize accurate classification of health status. However, existing methods usually assume that training data and test data are subject to the same distribution. In actual application, there are differences in speed under different working conditions, and there is a significant shift in data distribution, which limits the diagnosis generalization performance of traditional intelligent diagnosis methods under different working conditions.

[0004] To make up for the difference in data distribution caused by working conditions, the research introduces a transfer learning domain adaptation technology, the core of which is to learn domain-invariant features. In transfer learning, source domain and target domain data represent data under two different working conditions. The model is trained using source domain data and part of the target domain data, and through feature / model / sample instance / mixed transfer domain adaptation technology, accurate diagnosis under the working condition from the source domain to the target domain is realized. However, such methods still have obvious limitations in analyzing time-varying working conditions and unseen working conditions, such as in the scenario of continuous speed change, existing transfer methods are mostly based on the assumption of constant working conditions, making it difficult to decouple working condition interference and fault features; and when target domain data is not available, existing methods lack generalization ability outside the training working condition range.

[0005] In recent years, domain generalization methods try to learn more generalizable feature representations from multiple source data through adversarial learning, meta-learning or artificial disturbance. Some researches based on meta-learning strategy to mine common features among multiple source domains, or use causal learning to separate class-related factors and domain-related factors. However, existing researches still have two outstanding bottlenecks: 1) weak interpretability of feature extraction process: the network how to extract features and support decision-making lack transparency, which is difficult to understand from the perspective of human knowledge; 2) high data and computing cost: most methods rely on a large amount of multi-source domain data for training, which does not match the situation of difficult data acquisition and limited computing resources in actual industrial scenarios.

[0006] Interpretability is one of the most pressing problems in intelligent diagnosis. Black-box models are difficult to provide convincing decision-making basis, and lack of credibility in high-risk engineering applications. Existing interpretable analysis focuses on feature extraction, which can be divided into post-hoc explanation and pre-hoc explanation. Post-hoc explanation methods generate model decision-making basis explanation according to the features and parameters of the trained black-box network. Representative methods include saliency map and feature importance evaluation. Pre-interpretable methods embed physical or mathematical prior knowledge into network structure. Typical methods include algorithm unfolding technology to build algorithm-transparent network structure. Some research combines fast iterative shrinkage-thresholding algorithm (FISTA) and domain adaptation technology to build an interpretable and transferable diagnostic network. However, current interpretable research still faces the following challenges: 1) Doubt about the effectiveness of interpretation: due to the opaque feature transmission mechanism within the network, there is still controversy about whether post-hoc explanation is truly equivalent to model decision-making. 2) Abstract generalization mechanism: existing interpretable methods are difficult to cope with complex time-varying working conditions. Some newly proposed methods rely on abstract statistical distribution alignment for generalization mechanism, and lack of support for first-principle physical relationship.

[0007] The cross-condition fault diagnosis method of the interpretable three-feature extractor network (CN202510493430.X) converts the multi-layer sparse coding model into an interpretable feature extractor through algorithm unfolding, and constructs a three-feature extractor strategy and a prediction-feature adversarial loss function, which extracts shared features and private features of the source domain and the target domain, respectively, to realize interpretable cross-condition transfer fault diagnosis. However, the working condition generalization mechanism of the above scheme is still based on abstract statistical distribution alignment such as adversarial learning, and does not consider the physical relationship and fault mechanism caused by continuous change of speed. The interpretability of the diagnosis domain prior is weak, which limits the generalization ability of the model in unseen and speed-varying working conditions, and the output lacks basis explanation. SUMMARY

[0008] To address at least one of the problems existing in current technologies, this invention presents an interpretable deep learning-driven time-varying fault diagnosis method and apparatus. It proposes a Convolutional Sparse Modal Unrolling Network (CSMUNet), combining the fault mechanism of rotating machinery impact with prior knowledge of intrinsic modal impact responses. First, based on the physical mechanism that the fault impact response is composed of multiple modal components superimposed, the sparse coding algorithm is improved using the intrinsic modal impact response as a convolutional sparse dictionary, proposing a convolutional sparse modal coding algorithm to achieve interpretable feature extraction. Second, a network structure with complete physical meaning is constructed through algorithmic unrolling techniques, deeply integrating discriminative feature extraction and generalization capabilities with physical knowledge such as modal superposition. Furthermore, an equal-angle overlapping sample partitioning method is proposed. During the time-varying fault impact feature extraction process, pulse timing identification and pulse number control eliminate the influence of rotational speed changes, making the network robust to unseen operating conditions and reducing data dependence. Finally, interpretable intelligent fault diagnosis under unseen time-varying operating conditions is achieved.

[0009] This invention embeds the intrinsic modal impulse response as a convolutional sparse dictionary into the network structure. By dividing samples with equal-angle overlapping and identifying the pulse moment, the influence of rotational speed changes is explicitly eliminated during feature extraction, enabling the network to have stable generalization ability and physical interpretability for unseen working conditions.

[0010] The present invention is achieved by at least one of the following technical solutions.

[0011] An interpretable deep learning-driven time-varying fault diagnosis method is proposed. The vibration acceleration signal to be measured is input into a convolutional sparse mode matching unfolding network to obtain time-varying fault impact features. The order envelope spectrum is obtained using the time-varying fault impact features, and the fault type is determined based on the order envelope spectrum. The convolutional sparse mode matching unrolling network includes an encoder and a decoder. The encoder includes branch 1 and branch 2. Branch 1 is a multi-convolutional channel parallel iterative structure, which is unrolled by a fast encoding solution algorithm. Each convolutional channel contains a number of basic unrolling modules connected in series. The basic unrolling modules are used to perform the operation process of the fast encoding solution algorithm to solve the sparse optimization objective of the convolutional sparse intrinsic mode to obtain the optimal sparse vector, thereby realizing the convolutional sparse mode encoding of the vibration acceleration signal. The operation process includes calculating the gradient based on the mask vector and the sparse vector output from the previous round to update the intermediate vector, performing a soft thresholding operation on the intermediate vector, and adaptively initializing the output to obtain the current iteration vector. Branch 2 is used to identify the impact moment to obtain the mask vector of each basic unrolling module that merges into branch 1. The decoder is used to reconstruct the intrinsic mode impact response in each channel based on the optimal sparse vector and output the time-varying fault impact characteristics through weighted superposition.

[0012] Furthermore, the convolutional sparse mode matching unfolding network is trained through the following steps: Vibration acceleration and synchronous speed signal data of rotating machinery under various variable speed conditions were collected. First, the real-time rotation angle was obtained by summing the accumulated speed signals. Second, multiple overlapping samples were synchronously extracted from the collected vibration acceleration signals according to the time corresponding to the fixed rotation angle value. Then, the vibration acceleration, speed and angle time series samples under one of the operating conditions were used as model training data, and the samples of other time-varying speed conditions were used as test datasets. The training samples are input into the convolutional sparse mode unfolding network. Combined with the sparsity constraint of the mode weights, the network parameters such as the intrinsic frequency and damping ratio of the impact kernel are updated through gradient descent backpropagation. This minimizes the reconstruction loss of the time-varying fault impact features, so as to adaptively match the intrinsic mode of the fault signal and reconstruct the healthy signal.

[0013] Furthermore, overlapping samples are synchronously extracted from the collected vibration acceleration signals at fixed rotation angle values ​​for corresponding time intervals. The extraction limits are determined by interpolating the vibration acceleration, rotation speed, and real-time rotation angle at corresponding moments to ensure that each basic sample still contains complete impact when the rotation speed changes. First, the extraction interval angle is set. , No. Basic sample length The time interval angle Φ can be used to measure the time. t Interpolation was performed to obtain the results. Subsequently, to ensure that each base sample still contains complete impact when the rotational speed changes, the overlapping area was proportionally extended at both ends of each base sample segment to obtain multiple overlapping sample segments. The extension length at both ends of the base sample is... and satisfy (1) In the formula, and The first and The basic sample length, This is an estimate of the order of fault characteristics. The size of the impact core is given by max(·,·), which takes the maximum value.

[0014] By using the equal-angle overlapping sample division method and preprocessing the time-varying speed signal at fixed rotation angle intervals, it can be ensured that each training and test sample contains complete time-varying fault impact characteristics.

[0015] Furthermore, the sample selection involves extracting vibration acceleration, rotational speed, and angle time-series signals according to the same start and end time. The above three signal segments within the same time period constitute a sample. The sample of one time-varying working condition is used to construct the model training set, while the samples of other time-varying working conditions are used to construct the model test set.

[0016] Furthermore, the impact time is identified using a time-varying impact location algorithm. The steps for identifying the impact time using the time-varying impact location algorithm include: 1) Discrete cross-correlation is performed on the vibration acceleration signal and dictionary atoms to filter out strong noise interference in the actual scene, and the filtered vibration acceleration signal is obtained, highlighting the time-varying impact waveform; 2) Filter the dynamic acceleration signal x flt The local maxima and the corresponding real-time rotation angles at the sampling time are sorted in descending order; 3) Filter the sorted maxima from largest to smallest, such as when it's the 1st... r There are local maxima (the corresponding real-time rotation angle is denoted as ). φ r When ), if the real-time rotation angle corresponding to its subsequent local maximum is φ and φ r The difference is less than 2π / O' If it is not found, it is filtered out; otherwise, it is left for the next iteration until all are found. x flt Elements are traversed or filtered out; 4) Finally, the sampling time corresponding to the local maxima that were not filtered out is taken as the estimated impact time, denoted as... .

[0017] Furthermore, the sparse intrinsic mode dictionary is convolved with the intrinsic mode impulse response function. As its atoms, they exhibit periodic oscillation waveforms with amplitudes decaying exponentially: (2) In the formula, the first c ( c = 1, 2, …) natural frequencies ω d,c Damping ratio ζ c These are called the intrinsic modal parameters of a second-order linear underdamped mechanical system. Subsequently, the... c 1-order convolution sparse intrinsic modality dictionary D conv,c Defined as (3) In the formula, dictionary atoms For the first i Vector form at each impact moment , I This represents the number of sampling points for atoms, which is also the total number of atoms in the dictionary.

[0018] Furthermore, the property of the convolutional sparse intrinsic modality dictionary in step 2 is: for any vectorα With dictionary D conv,c Multiplication is equivalent to performing discrete cross-correlation with each of its atoms.

[0019] Furthermore, based on the classical sparse representation optimization problem, a sparse optimization objective for convolutional sparse intrinsic modes is established: based on the convolutional sparse intrinsic mode dictionary. D conv,c and estimated impact time The following sparse reconstruction optimization objective is constructed: (4) In the formula, the first and second terms are the reconstruction accuracy target and the sparse vector, respectively. Sparse regularization; C The order of the intrinsic modes; This represents the element-wise multiplication of two vectors; reconstructing the time-varying fault impact feature vector. From the dictionary of intrinsic modes of each order D conv,c With sparse vectors The product is obtained by multiplication and superposition. For a mask vector, the elements that are set to 1 are located in... Sampling points, number of which M The rest are set to 0; and They represent l 1- and l 2-norm; λ To optimize the objective expression, the adjustment parameters between the two terms are determined. In the first term, the adjustment parameters between the two terms are determined. and The corresponding multiplication operation, based on the knowledge of the impact failure mechanism, ensures that the impact response occurs only at the estimated impact time. Reconstruction ensures the sparsity and physical meaning of the reconstructed features.

[0020] Furthermore, the basic expansion module includes three layers. The first layer uses the sparse vector and mask vector output by the previous basic expansion module to calculate the gradient to update the intermediate vector. The second layer performs a soft thresholding operation on the intermediate vector to obtain the initial sparse vector for this iteration, which is used to adaptively initialize the iteration vector in the third layer. The first layer also includes a convolutional layer for performing discrete cross-correlation and discrete convolution operations.

[0021] Furthermore, the fast encoding solution algorithm is based on the FISTA derivation, and in the... In each loop, the following steps are performed for each intrinsic mode: First, based on the sparse vector output from the previous round... Update intermediate vector , (5) In the formula, η For real parameters, For the gradient operator with respect to a sparse vector, the gradient can be obtained from the following equation.

[0022] (6) In the formula, for and Vectors obtained through discrete cross-correlation. Discrete convolution is defined as follows: ,in express The One element; Secondly, for the intermediate vector Applying a soft threshold operation ; Finally, adaptive initialization Output the sparse vector of this iteration. .

[0023] Furthermore, the "fastness" of the fast encoding solution algorithm is reflected in the time complexity of the gradient calculation process. ,in, K and C These represent the number of iteration rounds and the order of the intrinsic frequency, respectively; using the dictionary constructed above... D conv,c The property of multiplying with the input vector can be transformed into an inner product operation that can be computed using the Fast Fourier Transform, and the gradient computation complexity can be reduced by... Reduce to ,in N The value is usually less than 150, and the sample length is related to the number of sampling points of the atoms. I Equal, usually greater than 500.

[0024] Furthermore, the encoder section comprises two branches: Branch 1's multi-channel parallel iterative structure is constructed using a fast encoding solution algorithm. Each channel contains the same number of cascaded basic expansion modules. The vibration acceleration signal propagates forward in each channel, solving and outputting sparse vectors of the natural mode dictionary for each order. The one-dimensional convolutional layer and the transposed convolutional layer perform the discrete cross-correlation and convolution operations of the proposed solution algorithm, respectively. The basic unfolding module consists of three layers. The first layer utilizes the sparse vector and mask vector output from the previous module. Calculate the gradient to update the intermediate vector The second layer performs a soft thresholding operation on the intermediate vector to obtain... This is used to obtain the iteration vector for the current round in the third layer adaptive initialization. Branch 2 is constructed based on the proposed impact moment identification algorithm, and its output... Import the basic expansion modules of branch 1; thus, in the first... c There are 1 channel, and the encoder output is... .

[0025] Furthermore, the decoder consists of two layers: the first layer is a convolutional layer, where the impulse kernels of each channel share parameters with the impulse kernels of the corresponding channels in branch 1 of the encoder, and their inputs are respectively connected to the constructed dictionary. D conv,c Multiplication yields the reconstructed intrinsic modal impulse responses of each order; the second layer is a fully connected layer, where the intrinsic modal weight vectors contain training parameters, forcing the network to automatically select qualified reconstructed intrinsic modal impulse responses through training, which are then superimposed to reconstruct time-varying fault impulse features. Furthermore, the network training loss function includes the target reconstruction loss and the sparsity constraint of the network intrinsic modality weight vector; during training, only the basic sample fragments are used to calculate the target loss, and the overlapping extension of the model input and output will be discarded; based on the obtained loss, the parameters of the network's convolutional sparse intrinsic modality dictionary and its reconstruction weight vector are updated through gradient descent backpropagation.

[0026] Furthermore, test samples from various time-varying operating conditions are sequentially input into the constructed network. Based on the output time-varying fault impact characteristics, the order envelope spectrum is obtained by combining order tracking technology and envelope demodulation method. By analyzing the amplitude significance of different fault types at characteristic orders, fault state identification and diagnosis under cross-operating conditions are realized.

[0027] Furthermore, the order envelope spectrum is obtained by utilizing time-varying fault impact features, including: extracting basic sample fragments from the test set output of the same time-varying operating condition and splicing them into a complete time-varying fault impact feature in chronological order to obtain the envelope demodulation order spectrum; identifying the fault state under time-varying operating conditions by analyzing the amplitude significance of different fault types at the feature order; and further, combining the parameters obtained from network learning or solving, including intrinsic mode parameters of each order, sparse vectors, and intrinsic mode weights, with contribution transformation, time-frequency transformation, and other means, to interpret the diagnostic generalization performance of the model under time-varying unseen operating conditions and improve the credibility of the output results.

[0028] Compared with the prior art, the beneficial effects of the present invention are at least as follows: 1. This invention constructs a network structure with physical interpretability. By combining the impact failure mechanism with prior knowledge of the inherent modal impact response, and using algorithmic unfolding technology to construct a convolutional sparse modal unfolding network, the structure of each layer of the network and the feature extraction process have clear physical meaning. This effectively solves the problem of low decision credibility caused by the black box characteristics of traditional deep learning models, and provides a credible and interpretable means for the industrial practice of artificial intelligence, which has certain engineering application value.

[0029] 2. This invention proposes an equal-angle overlapping sample partitioning method. By partitioning and overlapping time-varying working condition samples at equal angles, it overcomes the problems of splitting time-varying fault impact features and edge distortion caused by equal-duration sample partitioning, and significantly improves the robustness and generalization ability of the network in feature extraction under unseen time-varying working conditions.

[0030] 3. This invention realizes the sparse reconstruction and extraction of time-varying fault impact features: Based on the convolutional sparse mode coding algorithm, the network can adaptively optimize the natural frequency, damping ratio and other parameters of each channel impact kernel, and screen effective modes through sparse constraints, thereby accurately reconstructing physically interpretable time-varying fault impact features under strong noise background.

[0031] 4. This invention integrates order tracking and envelope demodulation to support fault diagnosis under unseen time-varying operating conditions: By combining the time-varying fault impact characteristics output by the network with order analysis technology, envelope demodulation is performed in the order domain to effectively separate speed fluctuation interference, highlight the fault characteristic order, and achieve reliable diagnosis under time-varying operating conditions.

[0032] 5. This invention reduces the dependence on multi-source training data and computing resources: the network structure is based on mechanism-driven design, the parameters have clear physical meanings, and the training process does not need to rely on a large amount of multi-source working condition data, thus improving the engineering applicability and deployment efficiency of the method. Attached Figure Description

[0033] Figure 1 This is an overall flowchart of the interpretable deep learning fault diagnosis according to an embodiment of the present invention; Figure 2 This is a diagram of the convolutional sparse intrinsic mode matching unfolded network structure according to an embodiment of the present invention; Figure 3 This is a basic unfolded module structure diagram of an embodiment of the present invention; Figure 4 This is a diagram showing the experimental setup for fault implantation in a harmonic reducer according to an embodiment of the present invention. Figure 5 This is a comparison diagram of inner ring fault features extracted under two test conditions in an embodiment of the present invention; Figure 6 This is a normal signal diagram reconstructed by the convolutional sparse intrinsic mode matching unfolding network according to an embodiment of the present invention. Figure 7 This is a diagram showing the reconstruction contribution of the effective and filtering channels to the inner ring fault characteristics under different operating conditions in this embodiment of the invention. Figure 8 This is a channel weight distribution diagram after training on the inner circle fault state according to an embodiment of the present invention; Figure 9 This is a wavelet transform time-frequency diagram of the inner ring fault signal according to an embodiment of the present invention; Figure 10 This is a channel weight distribution diagram after normal training according to an embodiment of the present invention; Figure 11 This is a diagram showing the reconstruction contribution of each channel to the normal signal under different operating conditions in an embodiment of the present invention. Detailed Implementation

[0034] To make the technical solution and objectives of the present invention clearer, the present invention will be described in detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific implementation steps described herein are only used to better illustrate the application of the present invention, but the technical features involved in the embodiments of the present invention are not limited thereto.

[0035] Please see Figure 1 The present invention provides an interpretable deep learning-driven time-varying fault diagnosis method, comprising the following steps: Step 1: Collect vibration acceleration signals and synchronous speed signals of rotating machinery under various variable speed conditions. First, obtain the real-time rotation angle by summing the accumulated speed signals. Second, synchronously extract multiple overlapping samples from the collected vibration acceleration signals according to the time corresponding to the fixed rotation angle value. Then, use the vibration acceleration, speed and angle time series samples under one of the operating conditions as model training data, and use the samples from other time-varying speed conditions as test datasets.

[0036] In some embodiments of the present invention, step 1 involves synchronously extracting multiple overlapping samples from the acquired vibration acceleration signal at fixed rotation angle values ​​over a corresponding time interval. The extraction limits are determined by interpolating the vibration acceleration, rotation speed, and real-time rotation angle at corresponding moments. First, the extraction interval angle is set. , No. Basic sample length The time interval angle Φ can be used to measure the time. t The interpolation solution yields the following result: (1) In the formula, and The start and end times of the basic sample. For real-time rotational speed, For integration variables, The time sampling resolution is set; subsequently, to ensure that each base sample still contains the complete impact when the rotational speed changes, the overlapping area is proportionally extended at both ends of each base sample segment to obtain multiple overlapping sample segments. The extension length at both ends of the base sample is... and satisfy: (2) In the formula, and The first and The basic sample length, This is an estimate of the order of fault characteristics. The impact kernel size is defined by max(·,·), which takes the maximum value to ensure the integrity of the impact waveform at the edge of the base sample.

[0037] In some embodiments of the present invention, the sample is extracted by extracting the vibration acceleration, rotational speed and angle time-series signals according to the same start and end time. The above three signal segments in the same time period constitute a sample. The sample of one time-varying working condition is used to construct the model training set, and the samples of other time-varying working conditions are used to construct the model test set.

[0038] Step 2: Construct a convolutional sparse intrinsic mode dictionary, design a time-varying impact localization algorithm, and then, based on the classical sparse representation optimization problem, simultaneously solve the convolutional sparse intrinsic mode sparse optimization objective. Optimize the solution using a fast encoding algorithm to output the optimal sparse vector of the intrinsic mode dictionary. Multiply the optimal sparse vector with the convolutional sparse intrinsic mode dictionary matrix to obtain the time-varying impact fault features, thereby realizing the extraction of time-varying impact fault features from the collected vibration acceleration signal.

[0039] In some embodiments of the present invention, step 2 involves convolving a sparse intrinsic mode dictionary and selecting intrinsic mode impulse response functions. As dictionary atoms, they exhibit periodic oscillation waveforms with amplitudes decaying exponentially: (3) In the formula, the first ( = 1, 2, …) natural frequencies Damping ratio These are called the intrinsic modal parameters of a second-order linear underdamped mechanical system. It is a step function.

[0040] No. 1-order convolution sparse intrinsic modality dictionary Defined as (4) In the formula, dictionary atoms For the first Vector form at each impact moment Defined as: (5) In the formula, For dictionary atoms In the The amplitude at each sampling time point, This represents the number of sampling points for atoms, which is also the total number of atoms in the dictionary. This is denoted as transpose.

[0041] In some embodiments of the present invention, the property of the convolutional sparse intrinsic mode dictionary in step 2 is: for any vector With the c 1-order convolution sparse intrinsic modality dictionary Multiplication is equivalent to performing discrete cross-correlation with each of its atoms. ,Right now: (6) In the formula, the first Cross-correlation numbers at each impact moment , For vectors In the The element values ​​at each sampling time.

[0042] The time-varying impact localization algorithm in this embodiment of the invention includes the following steps: 1) processing the vibration acceleration signal. x and dictionary atoms Discrete cross-correlation is performed to filter out strong noise interference in the actual scene, resulting in the filtered vibration acceleration signal. x flt 1) Highlight the time-varying impact waveform; 2) Filter the vibration acceleration signal x flt 3) Sort the local maxima and the corresponding real-time rotation angles at the sampling time in descending order; 4) Used to filter the sorted local maxima from largest to smallest, such as when it is the 1st... There are local maxima (the corresponding real-time rotation angle is denoted as ). When ), if the real-time rotation angle corresponding to its subsequent local maximum is If the condition does not meet formula (7), the signal is filtered out; otherwise, it is left for the next iteration until all vibration acceleration signals are obtained. x flt Elements are traversed or filtered out; 4) Finally, the sampling time corresponding to the local maxima that were not filtered out is taken as the estimated impact time, denoted as... .

[0043] (7) In this embodiment of the invention, based on the classical sparse representation optimization problem, a sparse optimization objective of convolutional sparse intrinsic modes is proposed, which means optimizing the convolutional sparse intrinsic mode dictionary. D conv,c and estimated impact time The following sparse reconstruction optimization objective is constructed: (8) In the formula, the first and second terms in the optimization objective are the reconstruction accuracy objective and the sparse vector objective, respectively. Sparse regularization; C The order of the intrinsic modes; This represents the element-wise multiplication of two vectors; reconstructing the time-varying fault impact feature vector. sparse intrinsic modal dictionary of each order of convolution With sparse vectors The product is obtained by multiplication and superposition. This is a mask vector, where the 1st element is located at the moment of impact. Sampling points ( The number of sampling points), and the rest are set to 0; and They represent l 1- and l 2-norm; λ This is the adjustment parameter between the two terms in formula (8). In the first term, and The corresponding multiplication operation is based on the knowledge of the impact failure mechanism, so that the impact response occurs only at the moment of impact. Reconstruction ensures the sparsity and physical meaning of the reconstructed features.

[0044] In one embodiment of the present invention, formula (8) is solved using a fast encoding solution algorithm to obtain the optimal sparse vector. The fast encoding solution algorithm is derived based on the Fast Iterative Shrinking Thresholding Algorithm (FISTA), and in the first... In each loop, the following steps are performed for each intrinsic mode: First, based on the sparse vector output from the previous round... Update intermediate vector : (9) In the formula, For real parameters, For the gradient operator with respect to a sparse vector, the gradient can be obtained from the following equation:

[0045] (10) In the formula, for With sparse vectors Vectors obtained through discrete cross-correlation. Discrete convolution is defined as follows: ,in express The Each element.

[0046] Secondly, for the intermediate vector Apply soft threshold operation: (11) In the formula, Let be the initial sparse vector for this iteration, and the sign function be defined as follows: .

[0047] Finally, update the momentum coefficient from the previous cycle. Obtain the momentum coefficient for this round. Used for adaptive initialization Output the sparse vector of this iteration. : (12) (13) Iterate the sparse vector using the above steps. K Next, we can obtain the first... c Sparse vectors of the sparse intrinsic modality dictionary of convolutional order This is considered as a solution to the optimization problem.

[0048] The "fastness" of the fast encoding solution algorithm proposed in this embodiment is reflected in the time complexity of the gradient calculation process (Equation (10)). , K and C These represent the number of iteration rounds and the order of intrinsic frequencies, respectively; the constructed convolutional sparse intrinsic mode dictionary is used. With the input vector (the optimal sparse vector) The properties of multiplication can be transformed into inner product operations that can be computed using the Fast Fourier Transform, and the gradient computation complexity can be reduced by... Reduce to In one embodiment, N The value is usually less than 150, and the sample length is related to the number of sampling points of the atoms. Equal, usually greater than 500.

[0049] Step 3: Design a convolutional sparse mode matching unfolding network. Its convolutional kernel is composed of atoms from the constructed convolutional sparse intrinsic mode dictionaries of each order. The convolutional sparse mode matching unfolding network includes an encoder and a decoder. The encoder is obtained by unfolding using a fast encoding solution algorithm. It is a multi-convolutional channel parallel iterative structure to realize the convolutional sparse mode encoding of vibration acceleration signals. Using the intrinsic mode sparse vectors output by the encoder, the decoder reconstructs the intrinsic mode impact response in each channel and outputs the time-varying fault impact characteristics through weighted superposition.

[0050] In some embodiments of the present invention, please refer to the structure of the convolutional sparse intrinsic mode matching unfolding network designed in step 3. Figure 2 The encoder part of the convolutional sparse mode matching unfolding network contains two branches: the multi-channel parallel iterative structure of branch 1 is composed of the unfolding of the fast encoding solution algorithm (the unfolding of the algorithm into a model structure is an existing method, such as Qin Yi, Yang Rui, Zhao Lijuan, et al. Unsupervised noise adaptive matching tracking algorithm for bearing fault diagnosis unfolding denoising network[J]. Journal of Mechanical Engineering, 2025, 61(12):26-38, and the specific unfolding steps of the fast encoding solution algorithm will not be elaborated here). Each convolutional channel contains basic unfolding modules connected in series with the same number of modules. The vibration acceleration signal propagates forward in each channel, and the optimal sparse vector is output by solving. Please refer to the basic expanded module structure diagram. Figure 3 It contains three layers, corresponding to the operation processes of formulas (9)-(10), (11), and (12)-(13), respectively. The first layer utilizes the sparse vector and mask vector output by the previous basic expansion module. Calculate the gradient to update the intermediate vector The second layer performs a soft thresholding operation on the intermediate vector to obtain the initial sparse vector for this iteration. This is used to obtain the iteration vector for the current round in the third layer adaptive initialization. Furthermore, the first layer includes one one-dimensional convolutional layer and one transposed convolutional layer, which respectively perform the discrete cross-correlation and discrete convolution operations of the fast encoding algorithm. Branch 2 is constructed based on the impact moment identification algorithm, and its output mask vector... By merging the basic expansion modules of branch 1, thus, in the first... c There are 1 channel, and the encoder output is... .

[0051] The decoder part of the convolutional sparse mode matching unrolling network constructed in this embodiment of the invention comprises two layers: the first layer is a convolutional layer, where the impulse kernels of each channel share parameters with the impulse kernels of the corresponding channels in branch 1 of the encoder, and their inputs are respectively connected to the constructed convolutional sparse intrinsic mode dictionary. Multiply to obtain the impulse responses of each reconstructed intrinsic mode. ,Right now: (14) The second layer is a fully connected layer, containing intrinsic mode weight vectors. The parameters in the training parameters force the convolutional sparse modality matching network to automatically select qualified reconstructed intrinsic modal impulse responses through training (the network adjusts itself). w The relative sizes of the elements in the vector are used to filter out qualified vectors. ), used to superimpose and reconstruct the time-varying fault impact characteristics, namely: (15) in, As the activation function, the weights of each intrinsic mode are... The domain is mapped over the interval [0,1].

[0052] Step 4: Input the training samples of the same time-varying working conditions into the convolutional sparse mode unfolding network. Combined with the sparsity constraint of the mode weights, update the network parameters such as the natural frequency and damping ratio of the impact kernel through gradient descent backpropagation, minimize the reconstruction loss of the time-varying fault impact features, and adaptively match the natural mode of the fault signal and the reconstructed healthy signal to achieve robust extraction of fault features.

[0053] In some embodiments of the present invention, during training, the network training loss function includes the target reconstruction loss and the network intrinsic mode weight vector. w The sparsity constraint is given by the first and second terms in the following equations: (16) In the formula, To train hyperparameters, the same function is used as in formula (4). λ , Q The number of samples; To constrain the sparsity of modal weights; during training, only basic sample fragments are used to calculate the target loss, and the overlapping extensions of the model input and output will be discarded; based on the obtained loss, the parameters of the network's convolutional sparse intrinsic modal dictionary and its reconstructed weight vectors are updated through backpropagation of gradient descent.

[0054] Step 5: Input the test samples of various time-varying operating conditions into the constructed convolutional sparse mode matching unfolding network in sequence. Based on the output time-varying fault impact characteristics, combine the order tracking technology and the envelope demodulation method to obtain the order envelope spectrum (the method of obtaining the order envelope spectrum is an existing method and will not be described in detail here). By analyzing the amplitude significance of different fault types at the characteristic order, fault state identification and diagnosis under cross-operating conditions can be realized.

[0055] In some embodiments of the present invention, step 5 extracts basic sample fragments from the test set output of the same time-varying operating condition and splices them into complete time-varying fault impact features in chronological order, and calculates the envelope demodulation order spectrum; by analyzing the amplitude significance of different fault types at the characteristic order, the fault state under the time-varying operating condition is identified; further, combined with the parameters obtained by network learning or solving, including intrinsic mode parameters of each order, sparse vectors and intrinsic mode weights, etc., combined with contribution transformation, time-frequency transformation and other means, the diagnostic generalization performance of the model under the time-varying unseen operating condition is explained, and the credibility of the output results is improved.

[0056] In one embodiment of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the methods described in the foregoing embodiments.

[0057] In one embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the foregoing embodiments.

[0058] In one embodiment of the present invention, the effectiveness of the method provided by the embodiment of the present invention is verified in conjunction with the accompanying drawings and experimental examples.

[0059] Conduct a fault implantation experiment on the harmonic reducer. Please refer to the experimental setup diagram. Figure 4 In the experimental setup, an accelerometer was mounted on the housing of a harmonic reducer to collect vibration acceleration signals. A servo motor controlled and synchronously acquired the input rotational speed. A thin-walled bearing inside the reducer was located between the wave generator and the flexspline, transmitting force and providing support through deformation. The experiment tested three healthy states of the thin-walled bearing: normal, inner ring fault, and outer ring fault. The characteristic orders of the bearings with inner and outer ring faults were 11.643 and 9.357, respectively. To verify the generalization ability of the constructed convolutional sparse mode matching unfolding network, a 12-second signal acquired at a constant speed of 500 rpm was used to train the network. The trained network was then tested using vibration data from two variable speed conditions (condition I and condition II). The input speed was accelerated from 0 rpm to 1000 rpm or 2000 rpm over 5 seconds, maintained at a constant speed for 2 seconds, and then decreased to 0 rpm for 5 seconds. The signal sampling frequency was set to 12.8 kHz. O' This is the average of the two fault orders, which is 10.5.

[0060] Table 1 shows the training parameter settings for the convolutional sparse modality matching unfolded network.

[0061] The training parameters for the convolutional sparse modality matching unfolded network are shown in Table 1. In formula (15)... Use the Sigmoid function to constrain the weights to the [0,1] interval.

[0062] To demonstrate the advantages of the constructed convolutional sparse modality matching unfolding network, three state-of-the-art methods suitable for bearing fault diagnosis under time-varying conditions were selected for comparison: Deep Expert Network (DEN), Single-Domain Generalized Physically Interpretable Network (SDGPIN), and Regularized Domain Adaptive Weight Optimization (RDAWO). For fair comparison, the recognition parts of these methods were all replaced with fully connected layers of the same input size, and mean squared error was used as the reconstruction loss function. Simultaneously, the input sample size was set to 1024 for all methods, the training epochs were 200, and the learning rate was set to the same value.

[0063] Fault characteristic shocks must possess both significance and sparse distribution characteristics, which differs from normal state signals. To quantify the extraction quality, four evaluation metrics are introduced: based on geometric mean... pq- mean (GM) 2to1 ), kurtosis, negative entropy, and Gini exponent. GM 2to1 Kurtosis is used to characterize the significance of the impact component; signals containing time-varying fault impact characteristics have higher values ​​than signals without impact. Negative entropy and the Gini exponent measure sparsity by quantifying the difference in distribution between the signal and white noise; larger values ​​indicate less noise interference. The original signal and the index results for each network are listed in Tables 2 and 3, respectively.

[0064] Table 2 Impact Indicators of Original Vibration Signals

[0065] Table 3 shows the impact indexes of the proposed method and the comparative method.

[0066] For both inner and outer ring faults, the fault features extracted by the constructed convolutional sparse mode matching (CMM) unrolling network showed improvements over the original signal in all four metrics, with the improvement being superior to the comparison methods. As shown in Table 3, under two test conditions, the constructed CMM unrolling network increased the kurtosis of the vibration acceleration signal by approximately 1.3 times and 3.2 times, respectively, reaching 1221.7 and 485.0, while the DEN method reduced it to 12.6% and 13.7% of the original value, respectively. The Gini index increased from 0.14 to over 0.35 after propagation through the constructed network, more than double that of the original signal. This significant increase indicates that the network is superior in maintaining the significance of fault impact and suppressing noise interference.

[0067] Analysis of signals under normal health conditions indicates that, due to the lack of time-varying fault impact characteristics, their indices should approach their minimum values. SDGPI and RDAWO outperform the constructed network under normal health conditions, but their index values ​​are closer to those under fault conditions. (The last sentence appears to be incomplete and possibly refers to a different network.) 2to1 Taking the Gini index as an example, the constructed convolutional sparse modality matching unfolded network achieves minimum differences exceeding 84% and 35% between normal and fault states, respectively, while RDAWO achieves maximum differences of less than 9% and 14% between the three health states. The results indicate that although the constructed network reconstructs the impact component, its ability to distinguish between different health states is significantly superior to the comparative model.

[0068] To visually demonstrate the diagnostic generalization capability of the constructed convolutional sparse mode matching unfolded network, please refer to the comparison of the extracted inner ring fault features under two different test conditions. Figure 5 Under operating condition I, the time-varying fault impact features extracted from noise by the constructed convolutional sparse mode matching unfolded network are clearer than those of other methods. Figure 5 (a)], the time-domain results are consistent with the data in Table 2. In the plotted order demodulation spectrum [ Figure 5 (b)], Only the constructed network reconstructs the fault order. O t and its modulation frequency conversion order O r Accurately indicates inner ring faults. When the model is generalized to operating condition II with a wider speed range, only the convolutional sparse mode matching unfolded network of this invention maintains significant time-varying fault impact characteristics and fault components in both the time domain and the order domain. Figure 5 (c) and (d)]. As shown in Table 2, the normal signal diagram reconstructed by the constructed convolutional sparse mode matching unfolded network can be found in [reference]. Figure 6 Although the output contains similar impact components, its periodicity is more ambiguous than that of the inner ring fault characteristics; the order envelope demodulation spectrum shows the fault order. O t The result is not significant, indicating that the output of the constructed network can effectively distinguish between normal and fault states.

[0069] The generalization ability of the constructed convolutional sparse mode matching unfolding network has clear physical interpretability. This is explained in the following two aspects: The explanation of how the constructed convolutional sparse mode matching unrolling network extracts fault features is provided. Please refer to the graphs showing the contribution of effective and filtered channels to the reconstruction of inner circle fault features under different operating conditions. Figure 7 The number of impact pulses identified by the samples remained stable. For example, the number of pulses in the acceleration phase of Condition I was 14, which is close to the number of pulses (12) in the 50th sample of the constant speed phase of Condition II. This matching number of pulses ensures that the network can accurately identify the impact intervals. Please refer to the channel weight distribution diagram after training for the inner loop fault state. Figure 8The constructed convolutional sparse modality matching unfolded network selected only 4 features from 32 candidate intrinsic modal impulse responses for feature reconstruction. For example... Figure 8 The first four weights contributed over 99% at almost all impact moments, and this sparse distribution is highly consistent with the fault mechanism. To verify whether the network selects the true modal order, the resonant frequency band of the fault signal was visualized using wavelet transform. Please refer to the wavelet transform time-frequency plot of the inner fault signal. Figure 9 It can be seen that the resonance components are concentrated in the [1700, 2600] Hz frequency band. On the other hand, as shown in Table 4, the natural frequencies obtained by training all fall within this frequency band, and the impulse kernel waveform is similar to the input features. Through convolution and cross-correlation operations, noise interference is effectively suppressed, thereby realizing the reconstruction of fault features.

[0070] Table 4. Modal parameter identification results of qualified channels in bearing inner ring faults.

[0071] The reasons for the difference in output between normal and fault states. Please refer to the channel weight distribution diagram after training in the normal state. Figure 10 Under normal input conditions, the weights trained by the network exhibit a uniform distribution, which is significantly different from that under fault conditions. Each weight value is tiny and can be considered zero. Furthermore, please refer to the diagram showing the contribution of each channel to the reconstruction of the normal signal under different operating conditions. Figure 11 The gradually changing colors in each contribution bar indicate that the contributions of the 32 intrinsic modes of impulse response are similar, which contrasts sharply with the phenomenon that a few intrinsic modes dominate under fault conditions. These results demonstrate that even when the impulse moment is identified, the network still determines that no impulse component exists, and therefore does not reconstruct fault features. This ensures the network's ability to extract discriminative fault features under time-varying conditions, thus accurately distinguishing between fault and normal states.

[0072] It should be noted that although the implementation of the present invention has been described in detail with reference to examples, those skilled in the art will readily understand that any modifications, substitutions, and improvements made without departing from the spirit and principles of the invention as set forth in the appended claims should be included within the scope of protection of the present invention.

Claims

1. A method for diagnosing time-varying faults under interpretable deep learning-driven conditions, characterized in that, The vibration acceleration signal to be measured is input into a convolutional sparse mode matching unfolding network to obtain time-varying fault impact features. The order envelope spectrum is obtained using the time-varying fault impact features, and the fault type is determined based on the order envelope spectrum. The convolutional sparse mode matching unrolling network includes an encoder and a decoder. The encoder includes branch 1 and branch 2. Branch 1 is a multi-convolutional channel parallel iterative structure, which is unrolled by a fast encoding solution algorithm. Each convolutional channel contains a number of basic unrolling modules connected in series. The basic unrolling modules are used to perform the operation process of the fast encoding solution algorithm to solve the sparse optimization objective of the convolutional sparse intrinsic mode to obtain the optimal sparse vector, thereby realizing the convolutional sparse mode encoding of the vibration acceleration signal. The operation process includes calculating the gradient based on the mask vector and the sparse vector output from the previous round to update the intermediate vector, performing a soft thresholding operation on the intermediate vector, and adaptively initializing the output to obtain the current iteration vector. Branch 2 is used to identify the impact moment to obtain the mask vector of each basic unrolling module that merges into branch 1. The decoder is used to reconstruct the intrinsic mode impact response in each channel based on the optimal sparse vector and output the time-varying fault impact characteristics through weighted superposition.

2. The time-varying fault diagnosis method driven by interpretable deep learning according to claim 1, characterized in that, The convolutional sparse modality matching unfolding network is trained through the following steps: Vibration acceleration signals and synchronous speed signals of motors under various variable speed operating conditions are collected. The real-time rotation angle is obtained by summing the accumulated speed signals. Multiple overlapping samples are synchronously extracted from the collected vibration acceleration signals according to the time corresponding to the fixed rotation angle value. The vibration acceleration, speed and angle time series samples under one operating condition are used as network training data, and the samples under other time-varying speed operating conditions are used as test datasets. The training samples are input into the convolutional sparse mode unfolding network. Combined with the sparsity constraint of the mode weights, the network parameters such as the intrinsic frequency and damping ratio of the impact kernel are updated through gradient descent backpropagation. This minimizes the reconstruction loss of the time-varying fault impact features, so as to adaptively match the intrinsic mode of the fault signal and reconstruct the healthy signal.

3. The time-varying fault diagnosis method driven by interpretable deep learning according to claim 2, characterized in that, Multiple overlapping samples are synchronously extracted from the acquired vibration acceleration signal at fixed rotation angle values ​​for corresponding time intervals. The extraction limits are determined by interpolating the vibration acceleration, rotation speed, and real-time rotation angle at corresponding moments. Set the interception interval angle , No. Basic sample length The time interval angle Φ can be used to measure the time. t The interpolation solution yields the following result: In the formula, and The start and end times of the basic sample. For real-time rotational speed, For integration variables, For time sampling resolution; The overlapping regions are expanded proportionally at both ends of each base sample segment to obtain multiple overlapping sample segments. The expansion length at both ends of the base sample is... and satisfy: In the formula, and The first and The basic sample length, This is an estimate of the order of fault characteristics. The size of the impact core is given by max(·,·), which takes the maximum value.

4. The time-varying fault diagnosis method driven by interpretable deep learning according to claim 1, characterized in that, The impact time is identified using a time-varying impact location algorithm. The steps involved in identifying the impact time using this algorithm include: Discrete cross-correlation is performed on the vibration acceleration signal and dictionary atoms to filter out strong noise interference in the actual scene and obtain the filtered vibration acceleration signal. The local maxima of the filtered vibration acceleration signal and the real-time rotation angle at the corresponding sampling time are sorted. The sorted local maxima are filtered out. If the real-time rotation angle corresponding to the local maximum does not meet the preset condition, it is filtered out. Otherwise, it is left to be traversed in the next step until all vibration acceleration signal elements are traversed or filtered out. The sampling time corresponding to the local maxima that were not filtered out is taken as the estimated impact time.

5. The time-varying fault diagnosis method driven by interpretable deep learning according to claim 1, characterized in that, The sparse reconstruction optimization objective is constructed based on the convolutional sparse intrinsic mode dictionary and the estimated impact time, and is expressed as: In the formula, C The order of the intrinsic modes; This indicates that the elements of two vectors are multiplied together. This represents the reconstructed time-varying fault impact feature vector; This represents a dictionary of sparse intrinsic modes of convolution, using intrinsic mode impulse response functions as dictionary atoms; Represents a sparse vector; It is a mask vector; Indicates the estimated moment of impact; λ To adjust the parameters; It is a vibration acceleration signal; and They represent l 1- and l 2-norm.

6. The time-varying fault diagnosis method driven by interpretable deep learning according to claim 1, characterized in that, The basic expansion module consists of three layers. The first layer uses the sparse vector and mask vector output by the previous basic expansion module to calculate the gradient to update the intermediate vector. The second layer performs a soft thresholding operation on the intermediate vector to obtain the initial sparse vector for this iteration, which is used to adaptively initialize the iteration vector in the third layer. The first layer also includes a convolutional layer for performing discrete cross-correlation and discrete convolution operations.

7. The time-varying fault diagnosis method driven by interpretable deep learning according to claim 1, characterized in that, The decoder includes: In the convolutional layer, the impulse kernel of each channel shares parameters with the impulse kernel of the corresponding channel of branch 1 of the encoder. Its input is multiplied by the convolutional sparse intrinsic mode dictionary to obtain the impulse response of each order of reconstructed intrinsic mode. The fully connected layer, with its intrinsic modality weight vector, forces the convolutional sparse modality matching network to automatically select qualified reconstructed intrinsic modality impact responses through training, so as to superimpose reconstructed time-varying fault impact features.

8. The interpretable deep learning-driven time-varying operating condition fault diagnosis method according to any one of claims 1-7, characterized in that, When training a convolutional sparse mode unfolding network, the network training loss function includes the target reconstruction loss and the network intrinsic mode weight vector. w Sparsity constraints, network training loss function The expression is: During training, the target loss is calculated using only basic sample fragments, and the overlapping extensions of the network input and output are discarded. Based on the obtained loss, the parameters of the network's convolutional sparse intrinsic modal dictionary and its reconstructed weight vectors are updated through backpropagation via gradient descent.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-8.

Images