Mechanical fault adaptive diagnosis method based on Granger causal inference and LSTM improvement

Abstract

The invention provides a mechanical fault adaptive diagnosis method based on Granger causal inference and LSTM improvement, and the method comprises the steps: firstly obtaining original vibration time sequence sample signals of a to-be-analyzed rotating machine in a normal state and N different fault states, and carrying out the noise reduction processing of the original vibration time sequence sample signals; the method comprises the following steps of: obtaining a de-noised rotating machinery vibration time sequence sample signal to be analyzed, taking the de-noised rotating machinery vibration time sequence sample signal as N + 1 time sequence variable sequence sets, performing Granger causal test on the N + 1 time sequence variable sequences in pairs, constructing a Granger causal matrix, and screening out time sequence variable sequence pairs with obvious causal; and training the LSTM model by using the time sequence variable sequence pair with significant causality and the state label to obtain a diagnosis model for performing fault adaptive diagnosis on the rotating machine, obtaining vibration time sequence data of the rotating machine to be analyzed, which needs fault diagnosis, performing noise reduction processing on the vibration time sequence data, inputting the processed vibration time sequence data into the trained diagnosis model, and performing fault diagnosis on the rotating machine to be analyzed. And obtaining a diagnosis result. According to the invention, the accuracy of rotary mechanical vibration time sequence signal analysis and mechanical fault diagnosis can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of mechanical fault diagnosis, in particular to a mechanical fault adaptive diagnosis method based on Granger causal inference and LSTM neural network. BACKGROUND

[0002] The causal reasoning method shows unique theoretical value and engineering advantage in rotating machinery fault diagnosis. This method can effectively identify the internal causal relationship between vibration anomalies and mechanical component damage by constructing an explicit causal graph model, and has stronger explainability and robustness than traditional correlation-based diagnosis methods. A "weight distribution fault bearing diagnosis method based on causal iterative correction" applied by Hefei University of Technology proposes a weight distribution fault bearing diagnosis method based on causal iterative correction. In the process of single training, the optimal weight distribution under each working condition is obtained, and the accuracy of fault diagnosis is improved through causal iterative correction. Another Chinese patent "bearing fault diagnosis method integrating feature extraction and ordering causal discovery" uses a bearing fault diagnosis method integrating feature extraction and ordering causal discovery. TSST is used for fault feature extraction to solve the problem of incomplete fault feature extraction. At the same time, the causal discovery theory is introduced to study the causal relationship between fault data, and the causal relationship is corrected through the ViT network model, which can more accurately diagnose bearing faults and improve the reliability of fault diagnosis. Jun Ho Park of Seoul University proposed a "framework for causal learning of neural networks", a causal learning framework for neural networks, which redefines the domain generalization problem by introducing a structural causal model, aiming to learn a causal representation independent of the domain, thereby improving the generalization ability of the model.

[0003] When the above-mentioned traditional fault diagnosis method is based on statistical correlation to construct the model, although it can capture the synchronous change trend between variables, it is difficult to distinguish the causal relationship and the false correlation (such as pseudo-correlation caused by noise). Especially under dynamic working conditions, the causal relationship between variables may change dynamically with parameters such as load and speed, and traditional static models cannot capture such time-varying characteristics, resulting in an increase in misdiagnosis rate. In addition, high-dimensional monitoring data often contain a large number of redundant or noise features, and traditional methods directly input all features, which will increase the computational burden and dilute the key signals. SUMMARY

[0004] In view of the problems existing in the prior art, the present application provides a mechanical fault adaptive diagnosis method based on Granger causal inference and improved LSTM, which combines the physical mechanism (causal relationship) of the mechanical system with the data-driven model (LSTM), and specifically includes the following steps:

[0005] Step 1: Obtain sample data of the rotating machinery to be analyzed, including the original vibration time series sample signals of the rotating machinery under normal conditions and N different fault conditions; perform noise reduction processing on the original vibration time series sample signals to obtain the noise-reduced vibration time series sample signals of the rotating machinery to be analyzed under N+1 conditions.

[0006] Step 2: Using the N+1 states of the denoised rotating machinery vibration time series sample signal obtained in Step 1 as a set of N+1 time series variable sequences, perform Granger causality tests on each pair of the N+1 time series variable sequences, construct the Granger causality matrix, and select the time series variable sequence pairs with significant causality.

[0007] Step 3: Using the causally significant time-series variable sequence pairs and state labels obtained through screening, train the LSTM model to obtain a diagnostic model for adaptive fault diagnosis of rotating machinery.

[0008] Step 4: Obtain the vibration time series data of the rotating machinery to be analyzed for fault diagnosis. After noise reduction processing of the vibration time series data, input it into the trained diagnostic model to obtain the diagnostic results.

[0009] In a further preferred embodiment, in step 1, wavelet transform denoising method is used to denoise the original vibration time series sample signal.

[0010] In a further preferred embodiment, step 1 selects five fault states: misalignment of the shaft angle, misalignment of the shafts, rotor mass imbalance, rotor-stator rubbing, and bearing failure.

[0011] In a further optimized approach, step 2, the process of constructing the Granger causality matrix is ​​as follows:

[0012] For a given pair of time series variables Determine the causal variable through the following process. Is it a result variable? Granger reason:

[0013] Step 2.1: Set the lag step size Constructing the hysteresis matrix and ;

[0014] Step 2.2: Based on the hysteresis matrix and Constructing constrained and unconstrained models:

[0015] Constrained Model: Causal Variables Not a result variable Granger reason:

[0016]

[0017] in The intercept of the regression line is represented in the model as the value when all lags are 0. The baseline value; for Its own lag term on the current The weighting coefficients of the values, where each element Represented as right The degree of impact; For the residuals of the constrained model;

[0018] Unrestricted model: Causal variable It is the outcome variable Granger reason:

[0019]

[0020] in for The lag term for the current The weighting coefficients of the values, where each element express right The degree of impact; The residuals are for the unconstrained model;

[0021] Step 2.3: Calculate the sum of squared residuals for the constrained model and the unconstrained model respectively:

[0022] Sum of squared residuals in the constrained model:

[0023]

[0024] in , , ;

[0025] Unconstrained model residual sum of squares:

[0026]

[0027] in , , ;

[0028] Step 2.4: According to the formula:

[0029]

[0030] calculate Statistics; among which To limit the number of model parameters, The number of parameters in the unrestricted model. The effective sample size;

[0031] Step 2.5: Significance test:

[0032] According to the formula

[0033]

[0034] Calculate decision parameters ,in Denotes the cumulative distribution function of the F-distribution; if Then it represents the causal variable. It is the outcome variable Granger's reason;

[0035] For each pair of time series variables, the judgment is made through steps 2.1 to 2.5, thereby constructing the Granger causality test matrix. Matrix elements Defined as:

[0036] .

[0037] Furthermore, the present invention also proposes an electronic device and a computer-readable storage medium for implementing the above method:

[0038] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the diagnostic method described above.

[0039] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described diagnostic method.

[0040] Beneficial effects

[0041] The proposed mechanical fault adaptive diagnosis method based on Granger causal inference and LSTM improvement adopts a Granger-LSTM fusion architecture, which combines the physical mechanism of the mechanical system with a data-driven model. By filtering features and dynamically allocating weights through a causal matrix, it can improve the accuracy of rotating machinery vibration time-series signal analysis and mechanical fault diagnosis.

[0042] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0043] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0044] Figure 1 Training process of LSTM neural network based on Granger causal inference;

[0045] Figure 2 Diagnostic results for six types of fault labels. Detailed Implementation

[0046] The embodiments of the present invention are described in detail below. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0047] This embodiment addresses the problem that traditional rotating machinery fault diagnosis methods cannot capture such time-varying characteristics, leading to a higher misdiagnosis rate, due to the dynamic changes in causal relationships between variables under rotating machinery operating conditions, such as load and rotational speed. It proposes an adaptive fault diagnosis method for machinery based on Granger causal inference and an improved LSTM, comprising the following steps:

[0048] Step 1: Obtain sample data of the rotating machinery to be analyzed, including the original vibration time series sample signals of the rotating machinery under normal conditions and N different fault conditions; perform noise reduction processing on the original vibration time series sample signals to obtain the noise-reduced vibration time series sample signals of the rotating machinery to be analyzed under N+1 conditions.

[0049] This embodiment targets a rotor-stator rubbing test specimen for rotating mechanical components. The selected fault states include five types: misalignment of the shaft angle, misalignment of the shafts, rotor mass imbalance, rotor-stator rubbing, and bearing failure.

[0050] Wavelet transform denoising is used to reduce noise in the original vibration time series signal. The specific process is as follows:

[0051] Step 1.1: Signal Wavelet basis function decomposition into coefficients at different scales. Mathematical expression:

[0052]

[0053] in: It is a wavelet function, representing high-frequency details; It is a scaling function, representing a low-frequency approximation; and These are the detail coefficients and approximation coefficients, respectively. The detail coefficients are extracted through wavelet decomposition. and approximation coefficients This forms the initial feature set, which is used for subsequent causal analysis.

[0054] Step 1.2: Threshold the detail coefficients to remove noise-dominated high-frequency components. The threshold setting uses a soft thresholding standard as follows:

[0055]

[0056] threshold Use a common threshold:

[0057]

[0058] in It is the noise standard deviation. This is the signal length. Noise standard deviation σ:

[0059]

[0060] Step 1.3: Signal Reconstruction. Reconstruct the denoised signal using the thresholded coefficients:

[0061]

[0062] Step 2: Using the denoised time-series sample signals of the rotating machinery to be analyzed, obtained from Step 1 under N+1 states, as a set of N+1 time-series variable sequences, perform Granger causality tests on each pair of the N+1 time-series variable sequences to construct a Granger causality matrix, and select time-series variable sequence pairs with significant causality. Specifically, this includes the following process:

[0063] Using the signals from the normal state and the five fault states after noise reduction and reconstruction in step 1 Establish a set of time series variables , This is the reconstructed and denoised signal under normal conditions. The reconstructed denoised signal is given under five fault conditions.

[0064] Extract pairwise time series sequences from the time series sequence set Perform Granger causality test, where Represents the cause variable, taking , , Represents the result variable, taking , ,and .

[0065] For a given pair of time series variables Determine the causal variable through the following process. Is it a result variable? Granger reason:

[0066] Step 2.1: Set the lag step size Constructing the hysteresis matrix and :

[0067]

[0068]

[0069] set up ,Pick ,but

[0070]

[0071] in This indicates the total number of time points in the time series data.

[0072] Step 2.2: Based on the hysteresis matrix and Constructing constrained and unconstrained models:

[0073] Constrained Model: Causal Variables Not a result variable Granger reason:

[0074]

[0075] in The intercept of the regression line is represented in the model as the value when all lags are 0. The baseline value is estimated using the least squares method; for Its own lag term on the current The weighting coefficients of the values, where each element Represented as right The degree of influence of (t∈T) is estimated by the least squares method; The residuals are for the constrained model.

[0076] Unrestricted model: Causal variable It is the outcome variable Granger reason:

[0077]

[0078] in for The lag term for the current The weighting coefficients of the values, where each element express right The degree of influence was estimated using the least squares method; The residuals are those of the unconstrained model.

[0079] Step 2.3: Calculate the residual sum of squares (RSS) for the constrained model and the unconstrained model respectively:

[0080] Sum of squared residuals in the constrained model:

[0081]

[0082] in , , ;

[0083] Unconstrained model residual sum of squares:

[0084]

[0085] in , , .

[0086] Step 2.4: According to the formula:

[0087]

[0088] calculate Statistics; among which To limit the number of model parameters, The number of parameters in the unrestricted model. This is the effective sample size.

[0089] Step 2.5: Significance test:

[0090] According to the formula

[0091]

[0092] Calculate decision parameters ,in Denotes the cumulative distribution function of the F-distribution; if Then it represents the causal variable. It is the outcome variable The Granger reason.

[0093] For each pair of time series variables, the judgment is made through steps 2.1 to 2.5, thereby constructing the Granger causality test matrix. Matrix elements Defined as:

[0094]

[0095] Step 3: Using the causally significant time-series variable sequence pairs and state labels obtained through screening, train the LSTM model to obtain a diagnostic model for adaptive fault diagnosis of rotating machinery.

[0096] Step 2 yields the Granger causality matrix, where non-zero elements correspond to time series variable pairs with significant causal relationships. The LSTM model is trained using the time series variable pairs with significant causal relationships and their corresponding state labels as the sample set.

[0097] The LSTM model captures temporal dependencies, learns the time lag effect between variables, that is, the influence of the historical value of X1 on the current value of X2, and indirectly learns causal patterns through state label classification tasks (normal, fault 1, fault 2, fault 3, fault 4, fault 5). For example, the label of fault 1 corresponds to the causal chain of X1→X2. The causal LSTM model recognizes this pattern and classifies it correctly.

[0098] In this embodiment, the sample set is divided into a training set and a test set using the HoldOut method, that is, the data is randomly divided according to a ratio (20%) to ensure that the data distribution of the training set and the test set is consistent.

[0099] The LSTM model captures temporal causality using the standard LSTM structure and dynamically adjusts the weights by monitoring the validation set. The dynamic weight adjustment mechanism is as follows:

[0100] (1) Adam optimizer update rules:

[0101]

[0102] in: and These are the first and second moments estimated after bias correction.

[0103] (2) LSTM gating mechanism:

[0104]

[0105] in: (Forget Gate) controls the degree to which historical memory is retained; (Input gate) controls the fusion weights of the current input, assigning higher weights to causal salient features; This represents the current state of the unit (memory carrier). The state of memory from the previous moment; The output of the sigmoid function determines the proportion of historical memory retained (0-1). The output of the sigmoid function controls the intensity of new information fusion; For the weight matrix (initialized using the Xavier method), perform a linear transformation on the concatenated vector. ; The bias term is calculated by backpropagation by minimizing the cross-entropy loss L; The hyperbolic tangent function compresses candidate memories to the range [-1, 1]. The Hadamard product is used to achieve element-wise modulation of the gated vector and memory.

[0106] (3) Loss function (cross-entropy):

[0107]

[0108] in: For the sample The true label; This represents the predicted probability of the output.

[0109] Basic multi-condition classification is achieved through HoldOut data partitioning and standard LSTM training process. The model weights are indirectly guided by causal simulation in the data and Granger matrix selection, and the weights are dynamically updated through the above three mechanisms.

[0110] Step 4: After the diagnostic model training is complete, acquire the vibration time-series data of the rotating machinery to be analyzed for fault diagnosis. After noise reduction processing, input the vibration time-series data into the trained diagnostic model to obtain the diagnostic results. Based on the diagnostic results and actual maintenance conditions, collect feedback information. Add new fault data and maintenance results as new training samples to the training set.

[0111] Figure 1This paper presents the dynamic changes of the Granger-LSTM fusion model during training, with two sub-graphs illustrating the evolution of the model's performance metrics and loss function. The upper sub-graph shows the performance metric curve exhibiting significant fluctuations in the early training phase (1-200 iterations), with the metric value remaining at a high level and exhibiting large amplitudes, reflecting the model's initial inability to establish effective feature representation capabilities. As training progresses to the 200-400 iteration range, the curve shows a clear monotonically decreasing trend, indicating that the model begins to capture effective patterns in the data, and the gradient descent algorithm is driving the parameters towards the optimal solution. Notably, the curve exhibits periodic oscillations between the 400-800 iterations, with the amplitude gradually decreasing. This may be due to parameter space exploration behavior caused by the learning rate setting or batch sampling randomness. Finally, after more than 1000 iterations, the curve converges to a stable plateau, verifying that the model has reached a local optimum. The lower sub-graph shows a more typical optimization trajectory, with the loss value decreasing sharply in the initial stage (first 100 iterations), consistent with the common pattern of rapid fitting in the early stages of neural networks. The brief rebound observed around iterations 300-400 may be related to noise in the training data or the optimizer overshooting local minima. Of particular note is that both curves maintained relatively stable low-fluctuation states after 500 iterations, and the black validation curve and the orange training curve eventually converged to similar levels, indicating that the model achieved good generalization performance on both the training and validation sets, without significant overfitting. The entire training process took approximately 2000 iterations, conforming to the standard convergence characteristics of supervised learning models: rapid initial improvement, oscillating optimization in the middle stage, and stable convergence in the later stage. The subtle fluctuations in the curves reflect the essential characteristics of the stochastic gradient descent algorithm, while the overall monotonic downward trend verifies the effectiveness of the backpropagation algorithm. This training dynamic highly matches theoretical expectations, proving that the model architecture and hyperparameter settings are reasonable and can effectively capture the potential causal relationships between different types of faults in the data.

[0112] Figure 2 The confusion matrix results are given. When calculating the confusion matrix, the input parameters are the true label and the diagnostic label. The output confusion matrix contains the following information: rows (true class), columns (diagnostic class), diagonal (number of correctly classified samples), and off-diagonal (number of misclassified samples).

[0113] Figure 2The confusion matrix in the figure visually presents the performance of the multivariate time series classification model based on the Granger-LSTM network in six types of mechanical fault diagnosis tasks, reflecting the effectiveness of data generation, feature engineering, and model architecture design in the code implementation. As can be seen from the diagonal elements of the matrix, the model achieves 100% accuracy in identifying faults of types 0-3 (all 360 samples are correctly classified), which verifies that the various fault features designed in the data generation function (such as the strong causal relationship X1→X2 for type 1 and the bidirectional causal relationship X5↔X6 for type 3) have sufficient discriminative power. Particularly noteworthy is the phenomenon that two samples in type 5 (nonlinear causal fault) were misclassified as type 4, with an accuracy of 98.8%. This misclassification in a specific direction may stem from the partial similarity between the two types of faults in Granger causal analysis—the cascade causality of type 4 and the nonlinear causality of type 5 in the code may produce similar time-series patterns under certain parameter combinations. The recall rates (true class accuracy) displayed in the right column of the matrix are highly consistent with the prediction precision (predicted class accuracy) in the bottom row, with only the prediction precision for class 4 slightly decreasing to 98.8%, indicating that the model maintains good prediction stability while maintaining high sensitivity. No cross-class misclassifications were found in the confusion matrix. Overall, this confusion matrix confirms the synergistic effectiveness of the temporal feature extraction and classification decision modules, and its overall accuracy of 99.69% has reached the practical standard of industrial-grade fault diagnosis systems.

[0114] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. An adaptive mechanical fault diagnosis method based on Granger causal inference and LSTM improvement, characterized in that: Includes the following steps: Step 1: Obtain sample data of the rotating machinery to be analyzed, including the original vibration time series sample signals of the rotating machinery under normal conditions and N different fault conditions; perform noise reduction processing on the original vibration time series sample signals to obtain the noise-reduced vibration time series sample signals of the rotating machinery to be analyzed under N+1 conditions. Step 2: Using the N+1 states of the denoised rotating machinery vibration time series sample signal obtained in Step 1 as a set of N+1 time series variable sequences, perform Granger causality tests on each pair of the N+1 time series variable sequences, construct the Granger causality matrix, and select the time series variable sequence pairs with significant causality. Step 3: Using the causally significant time-series variable sequence pairs and state labels obtained through screening, train the LSTM model to obtain a diagnostic model for adaptive fault diagnosis of rotating machinery. Step 4: Obtain the vibration time series data of the rotating machinery to be analyzed for fault diagnosis. After noise reduction processing of the vibration time series data, input it into the trained diagnostic model to obtain the diagnostic results.

2. The adaptive mechanical fault diagnosis method based on Granger causal inference and LSTM improvement according to claim 1, characterized in that: In step 1, wavelet transform denoising method is used to denoise the original vibration time series sample signal.

3. The adaptive mechanical fault diagnosis method based on Granger causal inference and LSTM improvement according to claim 1, characterized in that: In step 1, the selected fault states include five types: misalignment of the shaft angle, misalignment of the shafts, rotor mass imbalance, rotor-stator rubbing, and bearing failure.

4. The adaptive mechanical fault diagnosis method based on Granger causal inference and LSTM improvement according to claim 1, characterized in that: In step 2, the process of constructing the Granger causality matrix is ​​as follows: For a given pair of time series variables Determine the causal variable through the following process. Is it a result variable? Granger reason: Step 2.1: Set the lag step size Constructing the hysteresis matrix and ; Step 2.2: Based on the hysteresis matrix and Constructing constrained and unconstrained models: Constrained Model: Causal Variables Not a result variable Granger reason: in The intercept of the regression line is represented in the model as the value when all lags are 0. The baseline value; for Its own lag term on the current The weighting coefficients of the values, where each element Represented as right The extent of the impact; For the residuals of the constrained model; Unrestricted model: Causal variable It is the outcome variable Granger reason: in for The lag term for the current The weighting coefficients of the values, where each element express right The extent of the impact; The residuals are for the unconstrained model; Step 2.3: Calculate the sum of squared residuals for the constrained model and the unconstrained model respectively: Sum of squared residuals in the constrained model: in , , ; Unconstrained model residual sum of squares: in , , ; Step 2.4: According to the formula: calculate Statistics; among which To limit the number of model parameters, The number of parameters in the unrestricted model. The effective sample size; Step 2.5: Significance test: According to the formula Calculate decision parameters ,in Denotes the cumulative distribution function of the F-distribution; if Then it represents the causal variable. It is the outcome variable Granger's reason; For each pair of time series variables, the judgment is made through steps 2.1 to 2.5, thereby constructing the Granger causality test matrix. Matrix elements Defined as: 。 5. An electronic 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 program, it implements the diagnostic method as described in any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the diagnostic method as described in any one of claims 1 to 4.

Images