A bearing fault characterization method based on working condition constraint decoupling convolutional autoencoder

Abstract

The present application relates to the technical field of mechanical equipment monitoring, in particular to a bearing fault characterization method based on working condition constraint and decoupling convolution autoencoder. It comprises the following steps: positioning and reading working condition information, coding, normalizing and strictly aligning with vibration slices; optional pre-processing of fast Fourier transform, logarithmic amplitude and unified standardization; dividing the health information latent space and the working condition information latent space in the latent space, jointly introducing the trend constraint, the independence constraint and the working condition correlation constraint, and logarithmically stabilizing and weighting the reconstruction error and the correlation target; in the diagnosis stage, the health channel outputs of the same timestamp are aggregated to form a health index. The present application can obtain a monotonic health trend conforming to the physical law under complex working conditions such as segmented variable speed, reduce the dependence on fault labels, and is suitable for health evaluation and early warning of rotating machinery.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment monitoring technology, and more specifically, to a bearing fault characterization method based on a decoupled convolutional self-encoder with operating condition constraints. Background Technology

[0002] With the increasing automation and sensing capabilities of industrial equipment, condition monitoring and fault diagnosis of rotating machinery have become crucial for ensuring stable operation and reducing maintenance costs. Traditional signal processing methods based on spectrum analysis, envelope demodulation, statistics, and empirical mode decomposition are effective under single or stable operating conditions. However, under time-varying conditions, the feature distribution is prone to drift, significantly reducing the robustness of diagnostic thresholds and models, leading to an increase in false alarms and false negatives.

[0003] In recent years, with the development of deep learning, end-to-end feature learning methods such as convolutional neural networks (CNNs) and autoencoders (AE / CAE / VAEs) have been widely used for fault identification and health assessment of time-series signals such as vibration. These methods reduce reliance on manual feature engineering by automatically learning high-dimensional features. However, in complex and time-varying operating conditions, these methods generally suffer from problems such as coupling of equipment health information with operating condition factors, lack of operating condition-oriented training constraints, and difficulty in synchronizing operating conditions and vibration segments. At the same time, the interpretability and visualization of diagnostic results in terms of operating condition dimension are insufficient, making it difficult to support operation and maintenance decisions.

[0004] Therefore, this invention provides a bearing fault characterization method based on decoupled convolutional autoencoder with operating condition constraints. Summary of the Invention

[0005] The purpose of this invention is to provide a bearing fault characterization method based on decoupled convolutional autoencoders with operating condition constraints, in order to solve the problems that end-to-end feature learning methods such as convolutional neural networks (CNN) and autoencoders (AE / CAE / VAE) proposed in the background art generally have problems such as coupling of equipment health information with operating condition factors, lack of operating condition-oriented training constraints, and difficulty in synchronizing operating conditions and vibration segments.

[0006] To achieve the above objectives, the present invention aims to provide a bearing fault characterization method based on decoupled convolutional self-encoder under operating condition constraints, comprising the following steps: S1. Real-time acquisition of bearing vibration signals and synchronous acquisition of corresponding working condition data. The vibration signals are segmented by a sliding window and strictly aligned with the working condition data to form a sample sequence. The sample sequence is then subjected to frequency domain transformation and standardization to obtain preprocessed vibration sample data. S2. Construct a decoupled convolutional autoencoder model containing an encoder and a decoder. Extract health latent variables and working condition latent variables from vibration sample data through the encoder, and output the reconstructed signal through the decoder. S3. Based on vibration sample data and reconstructed signals, calculate reconstruction loss, trend constraint loss, independence constraint loss and working condition correlation constraint loss, and use a logarithmic stabilization strategy to weight and fuse the various losses to form the total loss value. S4. To minimize the total loss, an optimization algorithm is used to train the decoupled convolutional autoencoder model. S5. Input the vibration data to be diagnosed into the trained decoupled convolutional autoencoder model, extract the health latent variables, and perform aggregate calculation on the health latent variables at the same time stamp to obtain the health index characterizing the health status of the bearing.

[0007] As a further improvement to this technical solution, in S1, the vibration signal is segmented by a sliding window and strictly aligned with the working condition data to form a sample sequence. The specific steps involved are as follows: Real-time acquisition of bearing vibration signals Simultaneously record the operating condition data corresponding to each time point. Including rotational speed radial load With temperature ;in, Indexed by time point; For the operating condition data corresponding to each time point, one of the following strategies—equal-weighted mean, maximum extraction, or weighted average—is used to compress it into a single scalar, resulting in the operating condition compressed scalar. ; Scalar compression under operating conditions Normalization is performed, and numerical protection is implemented during the normalization process to prevent the denominator from being too small, resulting in normalized operating condition values. ; Set a fixed window length According to the set overlap rate Calculate step size From vibration signals Segments of fixed length are extracted sequentially as samples. ;in, For window indexing; For each sample, record its corresponding original time point index and assign the normalized operating condition value corresponding to that time point to the sample, establishing the correspondence between samples, time indices, and operating condition values ​​to form a structured sample set. .

[0008] As a further improvement to this technical solution, the specific steps involved in S1, which involve frequency domain transformation and standardization of the sample sequence to obtain preprocessed vibration sample data, are as follows: Based on structured sample sets For each sample Perform a real-number fast Fourier transform to convert the time-domain signal to the frequency domain, obtaining its frequency domain signal. ; Extracting frequency domain signals The amplitude is calculated and a logarithmic transformation is applied to obtain the logarithmic amplitude spectrum. ; Logarithmic amplitude spectrum Perform symmetrical splicing to restore it to the same length as the original sample; In the current training batch The logarithmic magnitude spectrum of each sample is calculated, with the mean calculated along the channel and length dimensions. and standard deviation And standardize each sample; in this process, if the calculated standard deviation If it is too small, the standard deviation will be... Set to 1; After frequency domain transformation and standardization, the preprocessed vibration sample data is obtained. .

[0009] As a further improvement to this technical solution, the specific steps involved in constructing the decoupled convolutional autoencoder model including an encoder and a decoder in step S2 are as follows: Construct a three-layer one-dimensional convolutional network as an encoder Based on vibration sample data, feature extraction is performed using three one-dimensional convolutional layers. After each convolutional layer, Tanh activation and batch normalization are executed sequentially. The feature map output from the third layer is flattened into a vector and input into two independent fully connected layers to obtain two latent vectors, including health latent variables. and working condition latent variables ; Construct a three-layer one-dimensional deconvolutional network as a decoder Potential health variables and working condition latent variables Each feature vector is projected onto the same dimension through a fully connected layer and then added element-wise to obtain the fused feature vector. ; Based on fused feature vectors The three one-dimensional deconvolutional layers are used sequentially for upsampling and reconstruction, and the output is a reconstructed signal with the same length as the input signal. By combining the encoder and decoder, a decoupled convolutional autoencoder model is constructed to obtain the normalized frequency domain sequence from the vibration sample data. The input is a decoupled convolutional autoencoder model. The encoder outputs health latent variables and operating condition latent variables, which are then processed by the decoder to output a reconstructed signal. .

[0010] As a further improvement to this technical solution, in step S3, the specific steps involved in calculating the reconstruction loss and the trend constraint loss are as follows: Based on reconstructed signal With the normalized frequency domain sequence extracted from vibration sample data Calculate the mean squared error loss between the two respectively. and absolute value loss The reconstruction loss is obtained by weighted fusion of these components. ; Based on health latent variables The original time index of its corresponding sample The health trend coefficient is obtained by calculating the Pearson correlation coefficient between the two, and the negative of the health trend coefficient is used as the trend constraint loss. .

[0011] As a further improvement to this technical solution, the specific steps involved in calculating the independence constraint loss and the operating condition correlation constraint loss in step S3 are as follows: Based on health latent variables Its corresponding normalized operating condition value The working condition fidelity coefficient is obtained by calculating the Pearson correlation coefficient between the two, and the absolute value of the working condition fidelity coefficient is taken as the independence constraint loss. ; Based on latent variables of working conditions Its corresponding normalized operating condition value The health condition coefficient is obtained by calculating the Pearson correlation coefficient between the two, and the negative of the health condition coefficient is used as the condition correlation constraint loss. .

[0012] As a further improvement to this technical solution, in step S3, the specific steps involved in weighted fusion of various losses using a logarithmic stabilization strategy to form the total loss value are as follows: Reconstructing loss Trend-based constraint loss Independence constraint loss and operating condition-related constraint loss , as input to the logarithmic stabilization strategy; Reconstruction loss With independence constraint loss The stable reconstruction loss was obtained by using the logarithmic transformation method. and stabilization independence constraint loss ; Trend-constrained loss Constraint loss related to operating conditions The stabilization trend constraint loss is obtained by using the signed logarithmic transformation method. and stable chemical condition constraint loss ; Stabilize and reconstruct loss Stabilization independence constraint loss Stabilization trend constraint loss , Stable chemical condition constraint loss The total loss is obtained by linearly weighting the values ​​according to preset weights. .

[0013] As a further improvement to this technical solution, in step S4, the specific steps involved in training the decoupled convolutional autoencoder model using an optimization algorithm with the objective of minimizing the total loss are as follows: S4.1, Based on vibration sample data The normalized frequency domain sequence Corresponding time index Normalized chemical condition value Construct a training set and initialize the encoder parameters of the decoupled convolutional autoencoder model. With decoder parameters The encoder parameters include the weights, biases, and corresponding batch normalization parameters of the three-layer one-dimensional convolution, while the decoder parameters include the weights, biases, and corresponding fully connected layer parameters of the three-layer one-dimensional transposed convolution. S4.2. Randomly select a batch of samples from the training set and input them into the encoder. To obtain potential health variables With working condition latent variables Then through the decoder Output reconstructed signal; S4.3, Based on the normalized frequency domain sequence Potential health variables With working condition latent variables Time Index Normalized chemical condition value Calculate the reconstruction loss sequentially. Trend constraint loss Independent constraint loss loss due to operating conditions The total loss value was obtained by weighting the data using a logarithmic stabilization strategy. ; S4.4, Based on total loss value The Adam optimizer is used to perform backpropagation and update the encoder parameters. With decoder parameters ; S4.5 Repeat steps S4.2 to S4.4, iterate the training to the preset number of rounds, save the model weights, and complete the model training.

[0014] As a further improvement to this technical solution, in step S5, the specific steps involved in inputting the vibration data to be diagnosed into the trained decoupled convolutional autoencoder model to extract potential health variables are as follows: After model training is complete, the inference phase begins. The vibration data to be diagnosed is processed according to the same procedure as in the training phase, including sliding window segmentation, frequency domain transformation, and standardization, to form the frequency domain sequence to be diagnosed. ; The frequency domain sequence to be diagnosed The trained decoupled convolutional autoencoder model is input, and after forward propagation by the encoder, the corresponding health latent variables and working condition latent variables are output. The health latent variables are then used for subsequent health indicator calculations.

[0015] As a further improvement to this technical solution, in step S5, the specific steps involved in aggregating and calculating the health potential variables at the same time point to obtain the health index characterizing the bearing's health status are as follows: For the same original time point There are multiple sliding window samples, each corresponding to a potential health variable. The health index value at that time point is obtained by taking the arithmetic mean of all latent health variables at the same time point. ; Health indicator values Perform a monotonic linear mapping to obtain the mapped health indicators. .

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This bearing fault characterization method based on decoupled convolutional autoencoders under operating condition constraints explicitly decouples latent variables of health information and latent variables of operating conditions at the model level. It also introduces three types of constraints: trend, independence, and operating condition correlation. This ensures a physically consistent separation of health features and operating condition factors under time-varying operating conditions. Simultaneously, logarithmic stabilization coordinates multiple objective dimensions and suppresses gradient explosion / vanishing. Numerical protection is provided for boundary conditions such as excessively small standard deviation and insufficient samples, resulting in a smoother, more controllable, and more convergent training process. Relying on a target design that prioritizes self-supervised reconstruction and supplements it with correlation constraints, the method significantly reduces its reliance on extensive manual fault labeling, further enhancing its feature representation capabilities.

[0017] 2. In this bearing fault characterization method based on decoupled convolutional autoencoder with operating condition constraints, at the model level, dual latent variables are used in conjunction with triple constraints of trend, independence, and operating condition correlation to achieve explicit decoupling between health and operating condition, thus suppressing information leakage. At the data engineering level, a closed-loop process of adaptive parsing of operating conditions, encoding / normalization, and slice-level synchronous amplification is proposed to ensure strict alignment of sample, original time series, and operating condition, and to eliminate format heterogeneity and index mismatch. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is a structural diagram of the working condition constraint decoupling model proposed in this invention; Figure 3 This is a graph showing the RMS variation trend of the bearing throughout its entire lifespan. Figure 4 This is a graph showing the changes in operating conditions of the bearing throughout its entire lifespan. Figure 5 The image of the health indicator obtained by decoupling in this invention; Figure 6 This is an image of the bearing operating condition obtained by decoupling according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example: Please see Figure 1-6 As shown, this embodiment provides a bearing fault characterization method based on decoupled convolutional self-encoder under operating condition constraints, including the following steps: S1. The vibration signal of the bearing is collected in real time by the vibration acceleration sensor and the corresponding working condition data is acquired simultaneously. The vibration signal is divided into sliding windows and strictly aligned with the working condition data to form a sample sequence. The sample sequence is frequency domain transformed and standardized to obtain the preprocessed vibration sample data. An accelerometer was installed in the bearing housing to acquire vibration signals at a sampling frequency of 20480Hz and a radial load of 4KN. Simultaneously, a displacement sensor was placed at the coupling to collect key phase signals, used to accurately measure the rotational speed and serve as the phase reference for condition alignment and order analysis. The experiment was performed using a segmented speed change program of "600r / min → 3000r / min → 600r / min," with the speed gradually increasing and decreasing in 600r / min increments, maintaining a constant speed for 30 minutes at each speed plateau; the complete cycle was repeated over a timescale of 240 minutes. To facilitate subsequent modeling and alignment, a 3-second sampling duration was used as the basic analysis window for each vibration segment.

[0021] In this embodiment, the bearing vibration signal is acquired in real time from the vibration acceleration sensor. Simultaneously record the operating condition data corresponding to each time point. Including rotational speed radial load With temperature ;in, Indexed by time point; in, The collected vibration signal is of length [missing information]. One-dimensional time-series data; sampling frequency of 20480Hz, meaning 20480 data points are collected per second, and stored in HDF5 / Mat (v7.3+) format; synchronously collected operating condition data are saved in triplet form: in, For operating condition data, Indicates rotational speed. Indicates radial load, Represents temperature, corresponding to a point in time. Operating data is stored in the same format as vibration signals to ensure complete timing synchronization. For the operating condition data corresponding to each time point, one of the following strategies—equal-weighted mean, maximum extraction, or weighted average—is used to compress it into a single scalar, resulting in the operating condition compressed scalar. ; Compress the operating data corresponding to each timestamp into a scalar. Three equivalent implementations: The strategies, from top to bottom, are: equal-weighted average strategy, suitable for overall performance representation; maximum extraction strategy, suitable for extreme conditions; and weighted average strategy, which adjusts the corresponding weights according to the actual scenario. When no strategy is specified or an invalid strategy is specified, the first equal-weighted average method is used by default. For the first The compression scalar of the operating condition corresponding to each time point; For the first Each operating parameter at a given time point The value, ; For the first The weighting coefficients of each operating condition parameter in the weighted average are summed to 1. Scalar compression under operating conditions Normalization is performed (Min-Max normalization or Z-score normalization can be selected), and numerical protection is implemented during the normalization process to prevent the denominator from being too small, resulting in the normalized operating condition value. ; in, These are the normalized operating conditions. This represents the minimum value of the scalar sequence of compression under operating conditions. This represents the maximum value of the scalar sequence of compression under operating conditions. The mean of the scalar sequence of compression under operating conditions. The standard deviation of the compression scalar sequence under operating conditions; the compression scalar sequence under operating conditions is the compression scalar of all operating conditions. The sequence is arranged in chronological order; during the normalization process, if the denominator is too small (e.g. Numerical protection is implemented in cases where division by zero occurs to avoid anomalies. Vibration signals A fixed-length sliding window is used for data augmentation: a fixed window length is set. According to the set overlap rate Calculate step size From vibration signals Segments of fixed length are extracted sequentially as samples. ;in, For window indexing; For the Vibration fragments corresponding to each timestamp Segmentation: in, For the first The original vibration segment corresponding to each time point; This represents the window length, corresponding to 0.1 seconds of data. The overlap rate, Step size; For the first At the 1st time point The starting index of the window; From The total number of samples that can be extracted from the sample; For the first At the 1st time point Samples from one window; Strict alignment: For each sample, record its corresponding original time point index and assign the normalized operating condition value corresponding to that time point to the sample, establishing a correspondence between samples, time indices, and operating condition values ​​to form a structured sample set. ; in, For the sample The corresponding original timestamp index; For the sample The corresponding normalized working condition value; The final sample sequence is a set: in, This represents the total number of segments in the original signal. The sample set is structured, and each sample contains a vibration slice. Original time index and normalized chemical condition values This is used as input for subsequent preprocessing and model training.

[0022] Furthermore, based on structured sample sets For each sample Perform a real-number fast Fourier transform to convert the time-domain signal to the frequency domain, obtaining its frequency domain signal. : in, The transformed frequency domain signal; It is the real number Fast Fourier Transform function; Extracting frequency domain signals The amplitude is calculated and a logarithmic transformation is applied to obtain the logarithmic amplitude spectrum. This enhances the weaker fault characteristics in the frequency domain while compressing the dynamic range and suppressing noise effects. in, This is a numerically stable term to prevent taking the logarithm of zero. For the first At the 1st time point The logarithmic magnitude spectrum of a window, the length of which depends on the window length. Parity; Logarithmic amplitude spectrum Perform symmetrical splicing to restore it to the same length as the original sample; Specifically, to maintain consistency with the input dimension of subsequent convolutional networks, Based on window length Symmetrical splicing based on parity restores the length. : If window length If it is even, then Length is Completed by mirror symmetry ; If window length If the number is odd, the symmetry rules will be adjusted accordingly. In the current training batch The logarithmic magnitude spectrum of each sample is calculated, with the mean calculated along the channel and length dimensions. and standard deviation And standardize each sample; in this process, if the calculated standard deviation If it is too small, the standard deviation will be... Set to 1; Standardize each sample: in, This represents the number of samples in the training batch. This represents the average of the current training batches. This represents the standard deviation of the current training batch. The frequency domain sequence is the normalized version. To address numerical instability caused by excessively small standard deviations within a batch, a protection mechanism is implemented: if the standard deviation... Then Set to 1, that is: After frequency domain transformation and standardization, the preprocessed vibration sample data is obtained. : in, For preprocessed vibration sample data; each The normalized frequency domain sequence and its corresponding original time index and normalized chemical condition value These samples are used together as standardized samples for subsequent decoupled convolutional autoencoder models.

[0023] S2. Construct a decoupled convolutional autoencoder model containing an encoder and a decoder. Extract health latent variables and working condition latent variables from vibration sample data through the encoder, and output the reconstructed signal through the decoder. In this embodiment, a three-layer one-dimensional convolutional network is constructed as the encoder. Based on vibration sample data, feature extraction is performed using three one-dimensional convolutional layers. After each convolutional layer, Tanh activation and batch normalization are executed sequentially. The feature map output from the third layer is flattened into a vector and input into two independent fully connected layers to obtain two latent vectors, including health latent variables. and working condition latent variables ; Specifically, such as Figure 2 As shown, encoder A three-layer one-dimensional convolutional structure is adopted, with Tanh activation function and batch normalization applied after each convolutional layer to enhance nonlinear expressive power and stabilize the training process. This structure is used to extract latent features from the preprocessed vibration signal. The specific parameter configuration is as follows: First convolutional layer: kernel size 64, stride 16, padding 24; The second convolutional layer has a kernel size of 32, a stride of 8, and padding of 12. The third convolutional layer has a kernel size of 16, a stride of 4, and padding of 6. After three convolutional layers, the feature map is flattened and fed into two independent fully connected branches, which output the health latent variables respectively. With working condition latent variables This achieves explicit decoupling of the feature space; in a typical setting, For the dimensions of latent health variables, latent health variables Scalar representation of the corresponding health degradation trend; For the dimensions of the latent variables of the working condition, the latent variables of the working condition High-dimensional embedding of the corresponding working condition effects; Construct a three-layer one-dimensional deconvolutional network as a decoder Potential health variables and working condition latent variables Each feature vector is projected onto the same dimension through a fully connected layer and then added element-wise to obtain the fused feature vector. ; in, To fuse feature vectors, The dimensions are consistent with the flattened output dimensions of the encoder. Based on fusion feature vectors The three layers of one-dimensional deconvolution (transposed convolution) are used sequentially for upsampling reconstruction, and the output is a reconstructed signal with the same length as the input signal (i.e., vibration sample data). like Figure 2 As shown, the decoder Specific parameters and encoder symmetry: First deconvolutional layer: kernel size 16, stride 4, padding 6; Second deconvolutional layer: kernel size 32, stride 8, padding 12; The third deconvolutional layer has a kernel size of 64, a stride of 16, and padding of 24. Final output reconstructed signal Its dimension is the same as the input standardized frequency domain sequence. (Obtained from vibration sample data) Same; By combining the encoder and decoder, a decoupled convolutional autoencoder model is constructed to obtain the normalized frequency domain sequence from the vibration sample data. The input is a decoupled convolutional autoencoder model. The encoder outputs health latent variables and operating condition latent variables, which are then processed by the decoder to output a reconstructed signal. ; The forward propagation process of the decoupled convolutional autoencoder model is described as follows: in, Indicates the first The timestamp, the first A frequency domain sequence standardized by a sliding window; These are potential health variables used to construct health indicators later. These are potential variables of the operating condition, used to characterize operating condition factors such as speed and load; To reconstruct the signal for self-supervised training.

[0024] S3. Based on vibration sample data and reconstructed signals, calculate reconstruction loss, trend constraint loss, independence constraint loss and working condition correlation constraint loss, and use a logarithmic stabilization strategy to weight and fuse the various losses to form the total loss value. In this embodiment, the reconstruction loss is calculated based on the reconstruction signal. With the normalized frequency domain sequence extracted from vibration sample data Calculate the mean squared error loss between the two respectively. and absolute value loss The reconstruction loss is obtained by weighted fusion of these components. ; The specific calculation method is as follows: in, This is the window length (typically 2048, corresponding to 0.1 seconds of data). For the first The timestamp, the first The sample at the th Standardized frequency domain amplitude at each frequency point To reconstruct the value at the corresponding position of the signal; For mean square error loss, For absolute value loss, For reconstruction loss; weighting coefficients of 0.7 and 0.3 are used to balance the contributions of the two types of loss, taking into account both smoothness and robustness; Calculating trend-constrained loss: based on latent health variables The original time index of its corresponding sample The health trend coefficient is obtained by calculating the Pearson correlation coefficient between the two, and the negative of the health trend coefficient is used as the trend constraint loss. ; Pearson correlation coefficient: in, This represents the latent health variables of all samples in the current batch. This represents the time index sequence of all samples in the current batch. for The mean, for The mean, for standard deviation for Standard deviation; For health trend coefficient; in, The loss is a trend-constrained loss, which is achieved through inverse optimization. Approaching 1, thus driving the health characteristics to increase monotonically over time, which conforms to the actual physical laws of the degradation process such as bearing wear and fatigue.

[0025] In this embodiment, the independence constraint loss is calculated based on the health latent variable. Its corresponding normalized operating condition value The working condition fidelity coefficient is obtained by calculating the Pearson correlation coefficient between the two, and the absolute value of the working condition fidelity coefficient is taken as the independence constraint loss. ; in, This is the normalized sequence of operating condition values. For the sequence of potential health variables, both correspond to the timestamp of the same sample; for standard deviation for Standard deviation; For operating condition fidelity factor, Minimize the independence constraint loss. This can make health characteristics less sensitive to changes in working conditions, thereby improving the stability and interpretability of health indicators; Calculate the loss due to the correlation between operating conditions and the operating conditions: based on the latent variables of the operating conditions Its corresponding normalized operating condition value The health condition coefficient is obtained by calculating the Pearson correlation coefficient between the two, and the negative of the health condition coefficient is used as the condition correlation constraint loss. ; in, This is the normalized sequence of operating condition values. These are sequences of latent variables related to the operating conditions, and both correspond to the timestamps of the same sample. for standard deviation for Standard deviation; For health condition factor, The loss is due to the constraint of operating condition correlation; by minimizing This ensures that the decoupled operating conditions are consistent with the actual operating conditions, prevents operating condition information from leaking into the health channel, and guarantees the purity of health indicators.

[0026] In this embodiment, the reconstruction loss is... Trend-based constraint loss Independence constraint loss and operating condition-related constraint loss , as input to the logarithmic stabilization strategy; These losses may have different numerical distributions. For example, reconstruction loss is usually a small positive value, while correlation coefficient losses (such as trend term, independent term, and condition term) may be close to zero or contain positive or negative signs. Direct weighting can easily lead to gradient instability. The logarithmic stabilization strategy obtains scale-stable and training-controllable loss values ​​based on the original loss data and through logarithmic transformation. For loss terms of different natures, the following stabilization transformations are applied respectively: Reconstruction loss With independence constraint loss The stable reconstruction loss was obtained by using the logarithmic transformation method. and stabilization independence constraint loss ; Non-negative loss terms include reconstruction loss. and independence constraint loss Both are non-negative values, and a logarithmic transformation method is used for compression to reduce the growth rate in the large value region while preserving monotonicity and preventing gradient explosion. in, To calculate the natural logarithm, add one term to ensure the input is positive and avoid taking the logarithm of zero. To stabilize the reconstruction loss, To stabilize the independence constraint loss; Trend-constrained loss Constraint loss related to operating conditions The stabilization trend constraint loss is obtained by using the signed logarithmic transformation method. and stable chemical condition constraint loss ; The signed loss term includes trend constraint loss. loss related to operating conditions The value can be positive or negative. A signed logarithmic transformation is used to compress the absolute value while preserving the sign information. in, This is the health trend coefficient, reflecting the monotonicity of health trends over time. This is a sign function, outputting ±1 or 0; To stabilize the trend-constrained loss, To stabilize chemical conditions and limit losses; Stabilize and reconstruct loss Stabilization independence constraint loss Stabilization trend constraint loss Stable chemical condition constraint loss The total loss is obtained by linearly weighting the values ​​according to preset weights. : in, For the weights of the reconstructed items, The weight of the trend term, For the weights of independent terms, The weights of the operating condition constraints are all hyperparameters; This represents the total loss. In typical implementation scenarios, to balance the contributions of various constraints and ensure the dominance of health indicators, it is recommended to set it to... This configuration has been verified in accelerated fatigue life tests and is suitable for monitoring bearing degradation under multiple operating conditions such as segmented speed change and variable load.

[0027] S4. To minimize the total loss, an optimization algorithm is used to train the decoupled convolutional autoencoder model. In this embodiment, S4.1, based on vibration sample data Normalized frequency domain sequence Corresponding time index Normalized chemical condition value Construct a training set and initialize the encoder parameters of the decoupled convolutional autoencoder model. With decoder parameters The encoder parameters include the weights, biases, and corresponding batch normalization parameters of the three-layer one-dimensional convolution, while the decoder parameters include the weights, biases, and corresponding fully connected layer parameters of the three-layer one-dimensional transposed convolution. Initially, set the batch size. Learning rate 100 rounds of training; S4.2 Randomly select a batch of samples from the training set. Input encoder To obtain potential health variables With working condition latent variables Then through the decoder Output reconstructed signal; Latent variables via decoder Reconstructed into a signal: S4.3, Based on the normalized frequency domain sequence Potential health variables With working condition latent variables Time Index Normalized chemical condition value Calculate the reconstruction loss sequentially. Trend constraint loss Independent constraint loss loss due to operating conditions The total loss value was obtained by weighting the data using a logarithmic stabilization strategy. ; S4.4, Based on total loss value The Adam optimizer is used to perform backpropagation and update the encoder parameters. With decoder parameters ; in, For learning rate, ; Total loss value For encoder parameters The gradient is the total loss value. decoder parameters gradient, This indicates an assignment operation, which updates the original parameter by subtracting the adjustment amount from the current parameter value; S4.5 Repeat steps S4.2 to S4.4, iterate the training to the preset number of rounds (e.g., 100 rounds), save the model weights, and complete the model training.

[0028] S5. In the inference stage, the vibration data to be diagnosed is input into the trained decoupled convolutional autoencoder model, health latent variables are extracted, and the health latent variables at the same time point are aggregated and calculated to obtain health indicators characterizing the health status of the bearing. In this embodiment, after model training is completed, the inference phase begins. The vibration data to be diagnosed is processed according to the same procedure as in the training phase, including sliding window segmentation, frequency domain transformation, and standardization, to form the frequency domain sequence to be diagnosed. ; The frequency domain sequence to be diagnosed The trained decoupled convolutional autoencoder model is input, and after forward propagation by the encoder, the corresponding health latent variables and working condition latent variables are output. Gradient calculation is not performed in this process, and the health latent variables are used for subsequent health indicator calculations.

[0029] In this embodiment, for the same original time point There are multiple sliding window samples (corresponding to different window start positions), and each sample corresponds to a health latent variable. To eliminate random fluctuations caused by window sampling and enhance the robustness of health indicators, the arithmetic mean of all latent health variables at the same time point is taken to obtain the health indicator value at that time point. ; in, These are latent health variables, usually scalars, used to characterize the health status of the sample. For time points The corresponding total number of samples; The aggregated health indicators reflect the overall health status of the bearing at that point in time. Health indicator values Perform a monotonic linear mapping to obtain the mapped health indicators. ; Specifically, if the encoder output layer uses the tanh activation function, the health latent variable The range of is restricted to [-1, 1]. To facilitate interpretation and threshold setting in practical applications, it can be further mapped to the [0, 1] interval through a linear transformation. The transformation formula is as follows: in, This represents the mapped health indicators; the mapping is a monotonically linear transformation and does not change the health indicator values. The trend over time and the relative order between samples are adjusted only in terms of numerical range; and the original time index is followed. The sorting is visualized to show its evolution over time, allowing for an intuitive examination of the decoupling relationship between operating condition characteristics and health indicators.

[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A bearing fault characterization method based on decoupled convolutional autoencoder under operating condition constraints, characterized in that: Includes the following steps: S1. Real-time acquisition of bearing vibration signals and synchronous acquisition of corresponding working condition data. The vibration signals are segmented by a sliding window and strictly aligned with the working condition data to form a sample sequence. The sample sequence is then subjected to frequency domain transformation and standardization to obtain preprocessed vibration sample data. S2. Construct a decoupled convolutional autoencoder model containing an encoder and a decoder. Extract health latent variables and working condition latent variables from vibration sample data through the encoder, and output the reconstructed signal through the decoder. S3. Based on vibration sample data and reconstructed signals, calculate reconstruction loss, trend constraint loss, independence constraint loss and working condition correlation constraint loss, and use a logarithmic stabilization strategy to weight and fuse the various losses to form the total loss value. S4. To minimize the total loss, an optimization algorithm is used to train the decoupled convolutional autoencoder model. S5. Input the vibration data to be diagnosed into the trained decoupled convolutional autoencoder model, extract the health latent variables, and aggregate the health latent variables at the same time point to obtain the health index characterizing the health status of the bearing. In step S3, the specific steps involved in calculating the reconstruction loss and the trend constraint loss are as follows: Based on reconstructed signal With the normalized frequency domain sequence extracted from vibration sample data Calculate the mean squared error loss between the two respectively. and absolute value loss The reconstruction loss is obtained by weighted fusion of these components. ; Based on health latent variables The original time index of its corresponding sample The health trend coefficient is obtained by calculating the Pearson correlation coefficient between the two, and the negative of the health trend coefficient is used as the trend constraint loss. ; In step S3, the specific steps involved in calculating the independence constraint loss and the operating condition correlation constraint loss are as follows: Based on health latent variables Its corresponding normalized operating condition value The working condition fidelity coefficient is obtained by calculating the Pearson correlation coefficient between the two, and the absolute value of the working condition fidelity coefficient is taken as the independence constraint loss. ; Based on latent variables of working conditions Its corresponding normalized operating condition value The health condition coefficient is obtained by calculating the Pearson correlation coefficient between the two, and the negative of the health condition coefficient is used as the condition correlation constraint loss. .

2. The bearing fault characterization method based on decoupled convolutional self-encoder with operating condition constraints according to claim 1, characterized in that: In step S1, the specific steps involved in performing sliding window segmentation of the vibration signal and strictly aligning it with the operating condition data to form a sample sequence are as follows: Real-time acquisition of bearing vibration signals Simultaneously record the operating condition data corresponding to each time point. Including rotational speed radial load With temperature ;in, Indexed by time point; For the operating condition data corresponding to each time point, one of the following strategies—equal-weighted mean, maximum extraction, or weighted average—is used to compress it into a single scalar, resulting in the operating condition compressed scalar. ; Scalar compression under operating conditions Normalization is performed, and numerical protection is implemented during the normalization process to prevent the denominator from being too small, resulting in normalized operating condition values. ; Set a fixed window length According to the set overlap rate Calculate step size From vibration signals Segments of fixed length are extracted sequentially as samples. ;in, For window indexing; For each sample, record its corresponding original time point index and assign the normalized operating condition value corresponding to that time point to the sample, establishing the correspondence between samples, time indices, and operating condition values ​​to form a structured sample set. .

3. The bearing fault characterization method based on decoupled convolutional self-encoder with operating condition constraints according to claim 2, characterized in that: In step S1, the specific steps involved in performing frequency domain transformation and standardization on the sample sequence to obtain preprocessed vibration sample data are as follows: Based on structured sample sets For each sample Perform a real-number fast Fourier transform to convert the time-domain signal to the frequency domain, obtaining its frequency domain signal. ; Extracting frequency domain signals The amplitude is calculated and a logarithmic transformation is applied to obtain the logarithmic amplitude spectrum. ; Logarithmic amplitude spectrum Perform symmetrical splicing to restore it to the same length as the original sample; In the current training batch The logarithmic magnitude spectrum of each sample is calculated, with the mean calculated along the channel and length dimensions. and standard deviation And standardize each sample; In this process, if the calculated standard deviation If it is too small, the standard deviation will be... Set to 1; After frequency domain transformation and standardization, the preprocessed vibration sample data is obtained. .

4. The bearing fault characterization method based on decoupled convolutional self-encoder with operating condition constraints according to claim 1, characterized in that: In S2, the specific steps involved in constructing a decoupled convolutional autoencoder model containing an encoder and a decoder are as follows: Construct a three-layer one-dimensional convolutional network as an encoder Based on vibration sample data, feature extraction is performed using three one-dimensional convolutional layers. After each convolutional layer, Tanh activation and batch normalization are executed sequentially. The feature map output from the third layer is flattened into a vector and input into two independent fully connected layers to obtain two latent vectors, including health latent variables. and working condition latent variables ; Construct a three-layer one-dimensional deconvolutional network as a decoder Potential health variables and working condition latent variables Each feature vector is projected onto the same dimension through a fully connected layer and then added element-wise to obtain the fused feature vector. ; Based on fusion feature vectors The three one-dimensional deconvolutional layers are used sequentially for upsampling and reconstruction, and the output is a reconstructed signal with the same length as the input signal. By combining the encoder and decoder, a decoupled convolutional autoencoder model is constructed to obtain the normalized frequency domain sequence from the vibration sample data. The input is a decoupled convolutional autoencoder model. The encoder outputs health latent variables and operating condition latent variables, which are then processed by the decoder to output a reconstructed signal. .

5. The bearing fault characterization method based on decoupled convolutional self-encoder with operating condition constraints according to claim 1, characterized in that: In step S3, the specific steps involved in weighted fusion of various losses using a logarithmic stabilization strategy to form the total loss value are as follows: Reconstructing loss Trend-based constraint loss Independence constraint loss and operating condition-related constraint loss , as input to the logarithmic stabilization strategy; Reconstruction loss With independence constraint loss The stable reconstruction loss was obtained by using the logarithmic transformation method. and stabilization independence constraint loss ; Trend-constrained loss Constraint loss related to operating conditions The stabilization trend constraint loss is obtained by using the signed logarithmic transformation method. and stable chemical condition constraint loss ; Stabilize and reconstruct loss Stabilization independence constraint loss Stabilization trend constraint loss Stable chemical condition constraint loss The total loss is obtained by linearly weighting the values ​​according to preset weights. .

6. The bearing fault characterization method based on decoupled convolutional self-encoder with operating condition constraints according to claim 1, characterized in that: In step S4, the specific steps involved in training the decoupled convolutional autoencoder model using an optimization algorithm with the objective of minimizing the total loss are as follows: S4.1, Based on vibration sample data The normalized frequency domain sequence Corresponding time index Normalized chemical condition value Construct a training set and initialize the encoder parameters of the decoupled convolutional autoencoder model. With decoder parameters The encoder parameters include the weights, biases, and corresponding batch normalization parameters of the three-layer one-dimensional convolution, while the decoder parameters include the weights, biases, and corresponding fully connected layer parameters of the three-layer one-dimensional transposed convolution. S4.

2. Randomly select a batch of samples from the training set and input them into the encoder. To obtain potential health variables With working condition latent variables Then through the decoder Output reconstructed signal; S4.3, Based on the normalized frequency domain sequence Potential health variables With working condition latent variables Time Index Normalized chemical condition value Calculate the reconstruction loss sequentially. Trend constraint loss Independent constraint loss loss due to operating conditions The total loss value was obtained by weighting the data using a logarithmic stabilization strategy. ; S4.4, Based on total loss value The Adam optimizer is used to perform backpropagation and update the encoder parameters. With decoder parameters ; S4.5 Repeat steps S4.2 to S4.4, iterate the training to the preset number of rounds, save the model weights, and complete the model training.

7. The bearing fault characterization method based on decoupled convolutional autoencoder with operating condition constraints according to claim 1, characterized in that: In step S5, the specific steps involved in inputting the vibration data to be diagnosed into the trained decoupled convolutional autoencoder model to extract latent health variables are as follows: After model training is complete, the inference phase begins. The vibration data to be diagnosed is processed according to the same procedure as in the training phase, including sliding window segmentation, frequency domain transformation, and standardization, to form the frequency domain sequence to be diagnosed. ; The frequency domain sequence to be diagnosed The trained decoupled convolutional autoencoder model is input, and after forward propagation by the encoder, the corresponding health latent variables and working condition latent variables are output. The health latent variables are then used for subsequent health indicator calculations.

8. The bearing fault characterization method based on decoupled convolutional self-encoder with operating condition constraints according to claim 1, characterized in that: In step S5, the specific steps involved in aggregating and calculating health potential variables at the same time point to obtain health indicators characterizing the bearing's health status are as follows: For the same original time point There are multiple sliding window samples, each corresponding to a potential health variable. The health index value at that time point is obtained by taking the arithmetic mean of all latent health variables at the same time point. ; Health indicator values Perform a monotonic linear mapping to obtain the mapped health indicators. .

Images