Machine tool fault diagnosis and analysis method and system based on big data

Abstract

The invention discloses a big data-based machine tool fault diagnosis and analysis method and system. The method comprises the following steps of: acquiring signal data of each part of a machine tool by utilizing a sensor; carrying out de-noising processing on the acquired signal data; the de-noised signals are decomposed by using successive variational mode decomposition, screening is carried out based on a Pearson's correlation coefficient, and the unscreened signals are fused and then are subjected to signal reconstruction with the screened signals; the method comprises the following steps: constructing a fault diagnosis model based on an Adaboost-Reformer integrated algorithm, and optimizing hyper-parameters of the Reformer model by using an improved water circulation optimization algorithm; the reconstructed signals are divided into a training set and a test set, the training set is input into the optimized model for training, machine tool signals collected in real time are input into the trained model after being subjected to denoising and reconstruction processing, and corresponding fault categories are obtained. The method can effectively extract fault features from the multi-mode sensor signals and perform accurate diagnosis, has strong robustness and high efficiency, and is suitable for intelligent fault diagnosis of industrial equipment.

Description

Technical Field

[0001] The present invention relates to a method and system for diagnosing and analyzing machine tool faults, and particularly to a method and system for diagnosing and analyzing machine tool faults based on big data. Background Art

[0002] With the rapid development of industrial automation and intelligence, machine tools, as core equipment in modern manufacturing, their operating status directly affects production efficiency and product quality. However, during long-term operation, machine tools are often affected by various factors, such as environmental changes, improper operation, etc., and are prone to failures, resulting in equipment downtime, reduced production efficiency, and even equipment damage. Traditional machine tool fault diagnosis methods mostly rely on manual experience or diagnostic algorithms based on single features, with low efficiency and the diagnostic results being easily affected by subjective factors. Especially in complex working conditions, traditional methods often have difficulty effectively identifying various types of faults.

[0003] With the development of sensor technology and data acquisition technology, modern machine tools are equipped with various sensors, such as vibration sensors, temperature sensors, pressure sensors, etc., which can collect the operating data of the equipment in real time. Fault diagnosis methods based on big data analysis and artificial intelligence have gradually become an effective way to solve this problem. These methods automatically identify and predict equipment faults by analyzing a large amount of historical data and real-time data, improving the accuracy and efficiency of fault diagnosis.

[0004] However, noise often exists in complex equipment signals, especially in industrial environments where the noise sources are complex and the signal quality is unstable. How to effectively remove noise and extract useful fault features has become an important challenge for fault diagnosis systems. Although existing signal processing technologies, such as wavelet denoising, Fourier transform, etc., can reduce the influence of noise to a certain extent, they still have certain limitations when facing multiple noise sources and complex signals. At the same time, successive variational mode decomposition is an adaptive method based on signal decomposition, which can decompose complex signals into multiple intrinsic mode functions and effectively extract the time-frequency features of the signals. Successive variational mode decomposition has good performance in the extraction of fault signals, especially when dealing with non-stationary and non-linear signals. However, when facing signals with strong noise, successive variational mode decomposition may lose some subtle fault features, affecting the subsequent fault diagnosis results. Therefore, how to make full use of various signal features on the basis of ensuring signal quality and improve the accuracy and robustness of fault diagnosis has become an urgent problem to be solved. Summary of the Invention

[0005] Object of the Invention: The object of the present invention is to provide a method and system for diagnosing and analyzing machine tool faults based on big data, which can effectively improve the accuracy and robustness of machine tool fault diagnosis.

[0006] Technical solution: A method for diagnosing and analyzing machine tool faults based on big data according to the present invention includes the following steps:

[0007] (1) Use sensors to collect vibration, temperature, pressure, and current signals of various parts of the machine tool, and classify the signals to ensure that each signal data has a label corresponding to a fault type;

[0008] (2) Perform wavelet denoising on the signal data collected in step (1) to remove noise interference;

[0009] (3) Use successive variational mode decomposition to decompose the denoised signal in step (2) into intrinsic mode functions, screen out the intrinsic mode functions with a relatively large correlation with the denoised signal based on the Pearson correlation coefficient, fuse the unselected signals, and perform signal reconstruction on the fused signal and the selected signal;

[0010] (4) Construct a fault diagnosis model based on the Adaboost-Reformer integration algorithm, use the improved water cycle optimization algorithm to optimize the hyperparameters of the Reformer model, further construct a comprehensive objective function considering the model classification accuracy, precision, and computational cost, and use the comprehensive objective function as the optimization goal to obtain the optimized Adaboost-Reformer model;

[0011] (5) Divide the reconstructed signal in step (3) into a training set and a test set, input the training set into the optimized Adaboost-Reformer model for training, evaluate the model performance through the test set, and input the real-time collected machine tool signal after being processed by steps (2) and (3) into the trained Adaboost-Reformer model to obtain the corresponding fault category.

[0012] Preferably, the sensors in step (1) include vibration sensors, current sensors, temperature sensors, and pressure sensors, and collect signal data to construct a machine tool fault data set R j =[R1, R2, …, R n as input features, where j is the number of fault sample numbers and n is the number of variables;

[0013] Add fault type labels L = [L1, L2,..., L m as output features, where m is the number of label types;

[0014] Construct a fault sample data set (R j , L m ).

[0015] Preferably, the wavelet denoising in step (2) is as follows: the signal X(t) is subjected to wavelet transform and decomposed into wavelet coefficients on different frequency bands or scales. The wavelet coefficients of each frequency band or scale are denoised by applying a threshold method, and the denoised coefficients are obtained. The threshold method includes a hard threshold and a soft threshold. Then, signal reconstruction is performed, and the denoised coefficients are subjected to inverse wavelet transform to obtain the denoised signal.

[0016] Preferably, in step (3), the Pearson correlation function of the decomposed intrinsic mode functions is calculated with the denoised signal and a correlation threshold is set for screening. The intrinsic mode functions with a correlation greater than the threshold are retained. The signals that are not screened are used to generate an auxiliary signal through weighted fusion, and the screened signals and the auxiliary signal are reconstructed.

[0017] Preferably, the Adaboost-Reformer model in step (4) combines the Adaboost algorithm with the optimized Reformer model. The Adaboost algorithm is used to weight and combine multiple Reformer weak classifiers to form a strong classifier. In each round of iteration, Adaboost adjusts its weight according to the classification error rate of each Reformer model, so that the samples with classification errors obtain higher weights in the next round of iteration.

[0018] Preferably, during the iteration process, the Adaboost algorithm calculates the classification error rate according to the output probability of the current model and updates the weight of the base classifier according to the error rate; the Reformer model updates the weight of each training sample, increasing the weight of the samples with classification errors and decreasing the weight of the correctly classified samples; through multiple iterations, the Adaboost algorithm gradually adjusts the weights of each Reformer model and combines the outputs of multiple Reformer models into a final diagnosis result through weighted voting; the Adaboost algorithm normalizes the comprehensive output through the Softmax layer to obtain the fault probability of each category.

[0019] Preferably, the improved water cycle optimization algorithm in step (4) associates the iteration step size with the number of iterations. The exponential iteration step size is:

[0020]

[0021] where C0 is the initial iteration step size, C max and C min are the maximum and minimum values of the iteration step size, and max_it is the maximum number of iterations.

[0022] A machine tool fault diagnosis and analysis system based on big data according to the present invention includes:

[0023] Data collection module: used to collect machine tool fault signal data and classify it to ensure that each signal data has a corresponding fault type label;

[0024] Preprocessing module: used to perform wavelet denoising on the collected signal data to remove noise interference;

[0025] Decomposition and reconstruction module: used to decompose the denoised signal into intrinsic mode functions using successive variational mode decomposition, screen out the intrinsic mode functions with a relatively large correlation with the denoised signal based on the Pearson correlation coefficient, fuse the unselected signals, and perform signal reconstruction on the fused signal and the selected signal;

[0026] Model construction module: used to construct a fault diagnosis model based on the Adaboost-Reformer integration algorithm, optimize the hyperparameters of the Reformer model using an improved water cycle optimization algorithm, further construct a comprehensive objective function considering the model classification accuracy, precision, and computational cost, and use the comprehensive objective function as the optimization goal to obtain an optimized Adaboost-Reformer model;

[0027] Fault diagnosis module: used to divide the reconstructed signal into a training set and a test set, input the training set into the optimized Adaboost-Reformer model for training, and input the real-time collected machine tool signal after denoising, decomposition, and reconstruction into the trained Adaboost-Reformer model to obtain the corresponding fault category.

[0028] A computer device, including one or more processors, a memory, and one or more programs, where the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the program is executed by the processor, it implements the steps of the method for analyzing machine tool fault diagnosis based on big data.

[0029] A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the method for analyzing machine tool fault diagnosis based on big data.

[0030] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: (1) The wavelet denoising method can retain the low-frequency components of the signal while removing high-frequency noise, thereby reducing the interference of noise on subsequent fault diagnosis; the SVMD method decomposes the signal into multiple intrinsic mode functions through adaptive decomposition, and can accurately extract the time-frequency features in the signal, providing high-quality data input for subsequent fault diagnosis; through the optimization of the data preprocessing method, the availability and stability of the signal are significantly improved, and the robustness of the fault diagnosis system is enhanced; (2) The Reformer model is combined with the Adaboost integration algorithm for machine tool fault diagnosis. As an efficient self-attention model, the Reformer model can capture long-distance dependencies in the signal and process large-scale data inputs. After being combined with Adaboost, it can improve the generalization ability and accuracy of the model by weighted integration of multiple weak classifiers. This combined model can give full play to their respective advantages, improve the accuracy and robustness of fault diagnosis. Through this way of ensemble learning, the diagnosis system has a high accuracy rate and strong adaptability when dealing with complex fault types; (3) The improved water cycle algorithm is used to optimize the hyperparameters of the Reformer model, further improving the performance of the model, being able to effectively find the optimal solution in the parameter search space, and avoiding the local optimum problem in traditional optimization algorithms; through the optimization of the hyperparameters, the performance of the Reformer model has been significantly improved, and the accuracy and efficiency of fault diagnosis have also been enhanced; in addition, as a natural heuristic algorithm, the improved water cycle algorithm can achieve a high optimization effect in a short time, reduce the computational overhead, and improve the real-time performance and operability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0032] The technical solution of the present invention will be further described below with reference to the drawings.

[0033] A machine tool fault diagnosis and analysis method based on big data, as Figure 1 shown, the specific steps are as follows:

[0034] (1) Use sensors to collect vibration, temperature, pressure, and current signals of various parts of the machine tool, and classify the signals to ensure that each signal data has a label corresponding to the fault type.

[0035] The sensors include vibration sensors, current sensors, temperature sensors, and pressure sensors, and collect vibration, temperature, pressure, and current data signals at different parts of the machine tool through the sensors.

[0036] First, number the sensors, including vibration sensor 1, temperature sensor 2, pressure sensor 3, and current sensor 4, to collect multi-dimensional data under the operating state of the machine tool, including x vi (t), x ci (t), x ti (t), x pi (t), which respectively represent the signals collected by vibration, current, temperature, and pressure sensor i at time t.

[0037] Construct a fault data set R j =[R1, R2, …, R n as the input feature, where j is the number of fault sample numbers and n is the number of variables.

[0038] The types of faults include but are not limited to motor overload, motor blocked rotation, hydraulic abnormality, and bearing failure.

[0039] Add fault type labels L = [L1, L2, ..., L m as the output feature, where m is the number of label types.

[0040] Construct a fault sample data set (R j , L m ).

[0041] (2) Perform wavelet denoising on the signal data collected in step (1) to remove noise interference.

[0042] The implementation steps of the wavelet denoising are as follows:

[0043] (21) Perform wavelet transform on the collected signal data X(t) to decompose it into wavelet coefficients W k (X(t)) on different frequency bands or scales, where k represents different scales or frequency bands.

[0044] (22) Apply the threshold method to each scale of wavelet coefficients to remove noise and obtain the denoised coefficients Common threshold methods include hard thresholding and soft thresholding. In hard thresholding, the part of the wavelet coefficient whose absolute value is less than the threshold is set to zero, while in soft thresholding, the part of the wavelet coefficient whose absolute value is less than the threshold is set to zero, and the part greater than the threshold is subtracted by the threshold. The threshold expressions are as follows:

[0045] Hard threshold:

[0046]

[0047] Soft threshold:

[0048]

[0049] (23) Perform signal reconstruction. Perform inverse wavelet transform on the denoised coefficients to obtain the denoised signal. The formula is as follows:

[0050]

[0051] Among them, IDWT represents inverse wavelet transform, is the coefficient after threshold processing.

[0052] (3) Use successive variational mode decomposition to decompose the denoised signal in step (2) into intrinsic mode functions. Screen the intrinsic mode functions with relatively high correlation with the denoised signal based on the Pearson correlation coefficient, fuse the unselected signals, and perform signal reconstruction on the fused signal and the selected signal.

[0053] The core of the SVMD is to decompose the signal into a set of intrinsic mode functions IMF k (t) and a residual signal R(t) to reveal the characteristics of the signal at different frequency components. The decomposition objective formula is:

[0054]

[0055] The decomposition of SVMD is achieved by solving an optimization problem, and the goal is to minimize the frequency bandwidth of each IMF. The frequency bandwidth can be defined by time-frequency representation as:

[0056]

[0057] Among them, IMF k (t) is the k-th intrinsic mode function, and w k is the corresponding central frequency.

[0058] By introducing the Lagrange multiplier λ(t) and the constraint term it is transformed into the Lagrangian objective function:

[0059]

[0060] By alternately updating IMF k (t) and w k , gradually approaching the optimal decomposition result. After SVMD decomposition, each obtained IMF k (t) represents the component of the signal in a certain frequency range, and R(t) represents the low-frequency trend or residual part of the signal.

[0061] Calculate the Pearson correlation coefficient of each decomposed IMF k (t) with the denoised signal one by one. The formula is:

[0062]

[0063] Wherein, is the covariance of IMF k (t) and ; is the standard deviation of IMF k (t) and .

[0064] Set a correlation threshold for screening, and retain the IMF whose correlation with t is greater than the threshold ρ k (t) to form a new set The un-screened signal is IMF‘ k (t). The screening conditions are as follows:

[0065]

[0066] The screened signal mainly retains the part with a large correlation with the denoised signal, but the un-screened signal may still contain low-correlation or auxiliary information valuable for diagnosis. Directly discarding it may lead to information loss. To avoid missing key information, the un-screened signals are fused and then signal reconstruction is performed with the screened signals. The un-screened signals are weighted and fused to generate an auxiliary signal IMF aux (t). The formula is as follows:

[0067]

[0068] Wherein, S represents the set of screened signals, and w k is the weight of the un-screened signal.

[0069] The formula for reconstructing the screened signal and the auxiliary signal is as follows:

[0070]

[0071] Wherein, γ is the fusion coefficient, which is used to balance the weights of the screened signal and the auxiliary signal.

[0072] (4) Construct a fault diagnosis model based on the Adaboost-Reformer integrated algorithm, optimize the hyperparameters of the Reformer model using the improved water cycle optimization algorithm, further construct a comprehensive objective function considering the model classification accuracy, precision, and computational cost, and use the comprehensive objective function as the optimization goal to obtain the optimized Adaboost-Reformer model.

[0073] The Reformer model is an efficient deep learning architecture designed specifically for long sequence modeling. By optimizing techniques such as the attention mechanism, introducing Locality-Sensitive Hashing (LSH) attention, and residual connections, it significantly improves computational efficiency and enhances the model's generalization ability. Through multi-layer feature extraction of signal data, the model can effectively process complex time-series data in machine tool fault diagnosis. The specific structure and implementation methods are as follows:

[0074] First, the input layer receives the reconstructed X f (t) signal and maps it to a high-dimensional embedding space. To enhance the model's ability to process time series, Reformer adds a positional encoding P after the embedding layer. This encoding is generated through sine and cosine functions to ensure that the model can understand the time order of the input data:

[0075] H 0 = E + P

[0076] Among them, P is usually generated using sine and cosine functions, and the specific form is:

[0077]

[0078] Where i represents the sequence position and k represents the dimension index.

[0079] The Reformer is characterized by using the Locality-Sensitive Hashing (LSH) attention mechanism, which effectively reduces the computational complexity of traditional self-attention from O(T 2 ) to O(T log T), which is particularly important for processing long time-series data. The basic formula for attention calculation is:

[0080]

[0081] Where: Q, K, and V are the query, key, and value matrices respectively, calculated as:

[0082] Q = H l W Q , K = H l W K , V = H l W V

[0083] These matrices are calculated with the help of hash mapping, further improving the computational efficiency of the model.

[0084] To avoid the problems of gradient vanishing and model degradation during model training, Reformer uses residual connections between each sub-layer and regularizes the output of each sub-layer. The specific operation is:

[0085] H l+1= LayerNorm(H l + Attention(Q, K, V))

[0086] Among them, LayerNorm represents layer normalization.

[0087] Through the residual connection, the model can train deep networks more easily, and at the same time, the regularization technique helps to prevent overfitting. After the attention layer, Reformer introduces a feed-forward connection layer (FeedForwardNetwork, FFN), which performs a non-linear transformation on the output of the attention layer to enhance the model's expressive power. The calculation formula of the feed-forward connection layer is:

[0088] H l+1 = LayerNorm(H l + FFN(H l ))

[0089] Among them, the formula of the feed-forward neural network FFN(H) is:

[0090] FFN(H) = ReLU(HW1 + b1)W2 + b2

[0091] Through these transformations, the model can learn more complex feature representations.

[0092] The final output layer of Reformer consists of a linear transformation and a Softmax layer, which is used to generate the fault probability distribution of each category. According to the probability distribution output by the Softmax layer, the category with the highest probability is selected as the final diagnosis result:

[0093] P = Softmax(HW0 + b0)

[0094] Among them, W0 and b0 are the weights and biases of the output layer, and P is the category probability distribution.

[0095] In order to improve the accuracy and robustness of the fault diagnosis system, the present invention proposes a method based on the combination of the Adaboost ensemble learning algorithm and multiple optimized Reformer models. This method effectively improves the system's ability to identify complex fault patterns by integrating multiple independent Reformer models. The specific steps are as follows:

[0096] Use the Adaboost algorithm to weight and combine multiple Reformer weak classifiers to form a strong classifier. In each round of iteration, Adaboost adjusts its weight according to the classification error rate of each Reformer model, so that the samples with classification errors obtain higher weights in the next round of iteration. In each iteration step, the Adaboost algorithm first calculates according to the output probability P of the current modeli Calculate the classification error rate ∈ i , and then update the weight α of this base classifier according to the error rate i :

[0097]

[0098] Secondly, the model updates the weights of each training sample, increases the weight for samples with classification errors, and decreases the weight for correctly classified samples. The adjusted sample weight is:

[0099] ω t+1 (x) = ω t (x) · exp(-α i y i P i (x))

[0100] Among them, ω t (x) is the weight of sample x in the t-th iteration, y i is the true label of the sample, and P i (x) is the failure probability of the i-th Reformer model.

[0101] Through multiple iterations, Adaboost will gradually adjust the weights of each Reformer model and combine the outputs of multiple Reformer models into a final diagnosis result by means of weighted voting.

[0102] The final output P f of Adaboost is the weighted sum of the outputs of all base classifiers (Reformer models):

[0103]

[0104] Among them, α i is the weight of the i-th base classifier (Reformer model), and P i is the output probability of the i-th model.

[0105] Finally, Adaboost will normalize the comprehensive output through the Softmax layer to obtain the failure probability of each category:

[0106]

[0107] This output is the diagnosis result, representing the probability of each failure type. Finally, the category with the highest probability is selected as the diagnosis result.

[0108] The Reformer model is a deep learning model for fault diagnosis, which can perform fault diagnosis based on input multi-dimensional data. However, the performance of the Reformer model highly depends on the selection of hyperparameters. Traditional hyperparameter tuning methods often require a large amount of computing resources and time, and are prone to falling into local optimal solutions.

[0109] The present invention proposes a method for optimizing the hyperparameters of the Reformer model based on an improved water cycle algorithm (WCA), optimizing the number of attention mechanism heads, the number of neurons in the hidden layer, and the learning rate, thereby improving the accuracy and efficiency of fault diagnosis.

[0110] The water cycle algorithm is an optimization algorithm that simulates the natural water cycle process. Its basic principle is to perform global search to find the optimal solution by simulating processes such as water evaporation, precipitation, and flow. The advantages of this algorithm include good global search ability and strong ability to jump out of local optima, and it is suitable for high-dimensional complex optimization problems. The process is as follows:

[0111] The initial population of the water cycle consists of three types of individuals: the sea (Sea), the river (River), and the stream (Stream). Among them, the Sea is the optimal individual with the smallest fitness function in the current population, the River is the sub-optimal individual with a smaller fitness function in the current population, and the remaining crossover individuals with larger fitness functions are Streams. The initial population can be expressed as:

[0112]

[0113] where, N pop is the population size, and N sr is the sum of the number of seas and rivers.

[0114] The fitness function of an individual can be expressed as:

[0115]

[0116] In the present invention, the core of the water cycle optimization algorithm (WCA) is to optimize the performance of the Reformer model in the fault diagnosis task. To comprehensively evaluate the diagnostic ability of the model, the present invention constructs a comprehensive objective function, combines multiple key performance indicators, enables the optimization process to balance the accuracy, precision, and computational cost of the model, and ensures the accuracy and computational efficiency of fault diagnosis.

[0117] The present invention adopts the following comprehensive objective function:

[0118] J = ω A A + ω Y Y + ω C CO′

[0119] Among them: J is the comprehensive objective function value, which measures the overall optimization effect of the Reformer model. The larger the value, the better the model performance. ω A 、ω Y and ω C are the weight coefficients of each index, which can avoid the problem of single objective dominance caused by artificially fixing weights in the optimization process.

[0120] The accuracy rate (A) of the balanced model is as follows:

[0121] It measures the classification accuracy rate of the model for all test samples:

[0122]

[0123] Among them: TP (True Positive) is the number of samples that are actually in the fault category and are correctly classified as faults. TN (True Negative) is the number of samples that are actually in the normal category and are correctly classified as normal. FP (False Positive) is the number of samples that are actually in the normal category but are misclassified as faults, that is, false alarms. FN (False Negative) is the number of samples that are actually in the fault category but are misclassified as normal, that is, missed alarms.

[0124] The precision rate (Y) of the balanced model is as follows:

[0125] It measures how many of the fault categories predicted by the model are real faults:

[0126]

[0127] That is: among the samples predicted as faults, the proportion of real fault samples.

[0128] The computational cost (CO) of the balanced model is as follows:

[0129] In order to avoid too high computational complexity of the model during the optimization process, a computational cost constraint is particularly introduced. And since the optimization direction of the computational cost CO is opposite to other indicators, the present invention introduces an exponential decay transformation to convert it into the same optimization direction as the accuracy rate and precision rate, which is defined as follows:

[0130]

[0131] CO’ = e -αCO

[0132] Among them: CO is the computational cost, which measures the computational resource consumption of the model. T is the model training time (seconds), N is the total number of model parameters, that is, the complexity of the model, ρ and is the weight coefficient of the calculation cost, and α is a hyperparameter that controls the influence of the calculation cost, making its optimization direction consistent with other objectives through exponential decay.

[0133] During the optimization process of the water cycle algorithm, the comprehensive objective function value J is used to evaluate the quality of each hyperparameter combination. The algorithm updates the positions of individuals iteratively to find the hyperparameter combination that maximizes the comprehensive objective function value.

[0134] In the WCA iteration process, the positions of different individuals are updated according to the principle in nature that Stream flows into River and Sea, and River flows into Sea. The update formula is as follows:

[0135]

[0136] Among them, and are the positions of Stream, River, and Sea respectively; the superscript t is the current iteration number; U is the position update coefficient set to a fixed value within the range of [0, 1].

[0137] To avoid the algorithm falling into a local optimal solution and increase the population diversity, a rainfall process is carried out to generate new individuals. The conditions for triggering the rainfall process are:

[0138]

[0139] where d max is a number close to 0.

[0140] When the conditions for rainfall are met, the rainfall process generates new individuals, thereby increasing the population diversity. There are mainly the following two ways:

[0141] Randomly generate new individuals in the problem space to increase the population diversity:

[0142]

[0143] Rainfall near the ocean and search for the optimal value near the optimal value:

[0144]

[0145] Among them, UB and LB are the upper and lower bounds of the space respectively, rand is a random number uniformly distributed between 0 and 1, and μ is the search range of the sea area. The smaller the value of μ, the smaller the search range.

[0146] Although the water cycle algorithm performs well in solving optimization problems, in practical applications, there are still deficiencies: the iteration step size C is set to a fixed value of [0, 1], resulting in an oscillating iteration phenomenon when approaching the optimal solution in the later stage of iteration.

[0147] To solve this problem, the iteration step size is associated with the number of iterations, and the exponential iteration step size is proposed as follows:

[0148]

[0149] where C0 is the initial iteration step size, C max and C min are the maximum and minimum values of the iteration step size, and max_it is the maximum number of iterations.

[0150] The method of associating the iteration step size with the number of iterations has a larger step size in the initial stage, thus having a faster search speed. As the iteration progresses, the iteration step size gradually decreases to obtain higher search accuracy.

[0151] (5) Divide the reconstructed signal in step (3) into a training set and a test set. Input the training set into the optimized Adaboost-Reformer model for training, evaluate the model performance through the test set, and input the real-time collected machine tool signal after being processed by steps (2) and (3) into the trained Adaboost-Reformer model to obtain the corresponding fault category.

[0152] The reconstructed signal is divided into the training set D train and the test set D test in a ratio of 8:2. By comparing the fault category diagnosed by the model and the real category of the test set, after denoising, decomposing, and reconstructing the real-time input signal data, input it into the trained Adaboost-Reformer model, and the model outputs the diagnosed fault type according to the characteristics of the input signal.

[0153] The present invention provides a machine tool fault diagnosis and analysis system based on big data, including:

[0154] Data collection module: used to collect machine tool fault signal data and classify it to ensure that each signal data has a label corresponding to the fault type;

[0155] Preprocessing module: used to perform wavelet denoising on the collected signal data to remove noise interference;

[0156] Decomposition and reconstruction module: used to decompose the denoised signal into intrinsic mode functions using successive variational mode decomposition, screen out the intrinsic mode functions with a large correlation with the denoised signal based on the Pearson correlation coefficient, fuse the unselected signals, and perform signal reconstruction on the fused signal and the selected signal;

[0157] Model construction module: It is used to construct a fault diagnosis model based on the Adaboost-Reformer integration algorithm, and optimize the hyperparameters of the Reformer model using an improved water cycle optimization algorithm to obtain an optimized Adaboost-Reformer model;

[0158] Fault diagnosis module: It is used to divide the reconstructed signal into a training set and a test set, input the training set into the optimized Adaboost-Reformer model for training, and input the real-time collected machine tool signals into the trained Adaboost-Reformer model after denoising, decomposition and reconstruction processing to obtain the corresponding fault categories.

Claims

1. A machine tool fault diagnosis and analysis method based on big data, characterized in that: The following steps are involved: (1) Use sensors to collect vibration, temperature, pressure, and current signals from various parts of the machine tool, and classify the signals to ensure that each signal data has a label corresponding to the fault type; (2) performing wavelet denoising on the signal data collected in step (1) to remove noise interference; (3) using successive variational mode decomposition to decompose the denoised signal in step (2) into intrinsic mode functions, selecting the intrinsic mode functions with greater correlation with the denoised signal based on the Pearson correlation coefficient, fusing the unfiltered signals, and reconstructing the fused signals and the filtered signals; (4) Construct a fault diagnosis model based on the Adaboost-Reformer integrated algorithm, use the improved water cycle optimization algorithm to optimize the hyperparameters of the Reformer model, further construct a comprehensive objective function that considers the model classification accuracy, precision and computational cost, and use the comprehensive objective function as the optimization target to obtain the optimized Adaboost-Reformer model; (5) The reconstructed signal in step (3) is divided into a training set and a test set. The training set is input into the optimized Adaboost-Reformer model for training. The model performance is evaluated by the test set. The real-time machine tool signal is processed in steps (2) and (3) and then input into the trained Adaboost-Reformer model to obtain the corresponding fault category.

2. The method for machine tool fault diagnosis and analysis based on big data according to claim 1, characterized in that: The sensors in step (1) include vibration sensors, current sensors, temperature sensors and pressure sensors, which collect signal data to construct a machine tool fault data set R j =[R1,R2,…,R n ] is the input feature, where j is the number of fault samples and n is the number of variables; Add fault type label L = [L1, L2, ..., L m ] as the output feature, where m is the number of label types; Construct a fault sample data set (R j , L m ).

3. The method for machine tool fault diagnosis and analysis based on big data according to claim 1, characterized in that: The wavelet denoising in step (2) is as follows: the signal X(t) is subjected to wavelet transform to decompose into wavelet coefficients at different frequency bands or scales, and the wavelet coefficients of each frequency band or scale are subjected to denoising by a threshold method to obtain denoised coefficients, wherein the threshold method includes a hard threshold and a soft threshold, and then the signal is reconstructed, and the denoised coefficients are subjected to an inverse wavelet transform to obtain a denoised signal.

4. The method for machine tool fault diagnosis and analysis based on big data according to claim 1, characterized in that: The step (3) calculates the decomposed intrinsic mode function and the denoised signal Pearson correlation function, set the correlation threshold for screening, and retain For the intrinsic mode functions with correlation greater than the threshold, the unfiltered signals are weighted fused to generate auxiliary signals, and the filtered signals and the auxiliary signals are reconstructed.

5. The method for machine tool fault diagnosis and analysis based on big data according to claim 1, characterized in that: The Adaboost-Reformer model in step (4) is a combination of the Adaboost algorithm and the optimized Reformer model. The Adaboost algorithm is used to weightedly combine multiple Reformer weak classifiers to form a strong classifier. In each iteration, Adaboost adjusts the weight of each Reformer model according to the classification error rate, so that the misclassified samples obtain a higher weight in the next iteration.

6. A machine tool fault diagnosis and analysis method based on big data according to claim 5, characterized in that: In the iterative process, the Adaboost algorithm calculates the classification error rate according to the output probability of the current model, and updates the weight of the base classifier according to the error rate; the Reformer model updates the weight of each training sample, increases the weight of the sample with classification error, and reduces the weight of the sample with correct classification; through multiple iterations, the Adaboost algorithm gradually adjusts the weight of each Reformer model, and combines the outputs of multiple Reformer models into a final diagnosis result by weighted voting; the Adaboost algorithm normalizes the comprehensive output through the Softmax layer to obtain the fault probability of each category.

7. The method for machine tool fault diagnosis and analysis based on big data according to claim 1, characterized in that: The improved water cycle optimization algorithm described in step (4) associates the iteration step length with the number of iterations, and the exponential iteration step length is: Among them, C0 is the initial iteration step size, C max and C min are the maximum and minimum values ​​of the iteration step, and max_it is the maximum number of iterations.

8. A machine tool fault diagnosis and analysis system based on big data, characterized in that: include: Data collection module: used to collect machine tool fault signal data and classify it to ensure that each signal data has a label corresponding to the fault type; Preprocessing module: used to perform wavelet denoising on the collected signal data to remove noise interference; Decomposition and reconstruction module: used to decompose the denoised signal into intrinsic mode functions using successive variational mode decomposition, screen the intrinsic mode functions with greater correlation with the denoised signal based on the Pearson correlation coefficient, fuse the unfiltered signals, and reconstruct the fused signals and the filtered signals; Model building module: used to build a fault diagnosis model based on the Adaboost-Reformer integrated algorithm, optimize the hyperparameters of the Reformer model using the improved water cycle optimization algorithm, and further build a comprehensive objective function that considers the model classification accuracy, precision and computational cost. The comprehensive objective function is used as the optimization target to obtain the optimized Adaboost-Reformer model. Fault diagnosis module: It is used to divide the reconstructed signal into a training set and a test set, input the training set into the optimized Adaboost-Reformer model for training, and input the real-time collected machine tool signal into the trained Adaboost-Reformer model after denoising, decomposition and reconstruction to obtain the corresponding fault category.

9. A computer device, characterized in that: The method comprises one or more processors, a memory and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of a machine tool fault diagnosis and analysis method based on big data are implemented as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, the steps of a machine tool fault diagnosis and analysis method based on big data as described in any one of claims 1 to 7 are implemented.

Images