Rolling bearing intelligent fault diagnosis method based on error self-correction

Abstract

The invention provides a rolling bearing intelligent fault diagnosis method based on error self-correction. The rolling bearing intelligent fault diagnosis method comprises the steps of obtaining angular domain equal-period vibration data of a rolling bearing; obtaining a total time-frequency matrix; generating a full frequency domain difference three-dimensional tensor, setting a diagnosis accuracy threshold, a diagnosis balance threshold and an initial tensor weight matrix, and adjusting the difference tensor; calculating a full-frequency-domain difference distinguishing capability, determining an optimal frequency band and performing sample generation; training and verifying a fault diagnosis model, and testing the model; carrying out prediction accuracy and balance calculation on the fault diagnosis model, and calculating a weight increment matrix; evaluating the accuracy and balance of the fault diagnosis model, and judging whether the prediction accuracy and balance of the fault diagnosis model meet conditions or not; if not, input and model optimization based on error self-correction are carried out, and if yes, the optimal input structure and fault diagnosis model are output. And rapid and reliable aero-engine rolling bearing fault diagnosis can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of aero-engine health management, and particularly relates to an intelligent fault diagnosis method for rolling bearings based on error self-correction. Background Art

[0002] In the prior art, a rolling bearing is one of the key components of an aero-engine, providing support and positioning for the rotor. Its stable and reliable operation is crucial for ensuring the normal operation of the engine. However, due to long-term operation under high temperature, high rotational speed, and complex load conditions, the rolling bearing is also one of the most easily damaged components of the engine. Minor bearing faults will increase the operation and maintenance costs of the engine, and in severe cases, it will cause damage to other components of the engine, directly affecting flight safety. Therefore, the identification and diagnosis of bearing faults are of particular significance.

[0003] In the aspect of rolling bearing fault diagnosis, traditional diagnosis methods relying on expert experience and theoretical models are difficult to meet the requirements in terms of efficiency and reliability. Data-driven intelligent fault diagnosis methods have been extensively studied due to their advantages such as low prior knowledge dependence, deeper feature extraction, and higher data processing efficiency. Since the frequency-domain characteristics of bearing faults show better discrimination than time-domain characteristics, they are widely used as the input for fault diagnosis.

[0004] Current research generally uses the full-band spectrum. However, the fault characteristics change with the evolution of damage. Using the full-band spectrum cannot locate the sensitive frequency band, resulting in the diagnostic model being easily affected by irrelevant frequency band components and reducing the generalization ability of the fault diagnosis model.

[0005] In addition, existing research mainly focuses on improving the overall accuracy of fault diagnosis, while paying less attention to the balance of the recognition effects of the diagnostic model for different categories, and unable to guarantee the diagnostic effects for small-sample categories.

[0006] In view of this, the inventor of the present application has designed an intelligent fault diagnosis method for rolling bearings based on error self-correction in order to overcome the above technical problems. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to overcome the defects in the prior art that using the full-band spectrum cannot locate the sensitive frequency band, resulting in the diagnostic model being easily affected by irrelevant frequency band components and reducing the generalization ability of the fault diagnosis model; and existing research mainly focuses on improving the overall accuracy of fault diagnosis, while paying less attention to the balance of the recognition effects of the diagnostic model for different categories and unable to guarantee the diagnostic effects for small-sample categories, and to provide an intelligent fault diagnosis method for rolling bearings based on error self-correction.

[0008] The present invention solves the above technical problems through the following technical solutions:

[0009] The present invention provides an intelligent fault diagnosis method for rolling bearings based on error self-correction, characterized in that the intelligent fault diagnosis method comprises the following steps: S 1 、Obtain the rolling bearing angular domain equi-periodic vibration data corresponding to each target category; S 2 、Based on the short-time Fourier transform method, obtain the total time-frequency matrix for each category; S 3 、Generate a three-dimensional tensor of full-frequency domain differences between categories, set a diagnostic accuracy threshold and a diagnostic balance threshold, set an initial tensor weight matrix, and adjust the difference tensor based on the tensor weight matrix;

[0010] S 4 、For all categories to be diagnosed, calculate the full-frequency domain difference discrimination ability, determine the optimal frequency band and generate samples; S 5 、Based on the generated sample sets of each category, train and verify the fault diagnosis model, and conduct model testing; S 6 、Calculate the prediction accuracy and balance of the fault diagnosis model, and calculate the weight increment matrix; S 7 、Evaluate the accuracy and balance of the fault diagnosis model, and determine whether the prediction accuracy and balance of the fault diagnosis model meet the conditions; if not, use the weight increment matrix for input and model optimization based on error self-correction, and if so, output the obtained optimal input structure and fault diagnosis model.

[0011] According to an embodiment of the present invention, the step S 1 comprises the following steps: S 11 、For the target diagnosis category, based on the historical operation data of the engine, obtain the rolling bearing vibration data segments corresponding to each target category; S 12 、Based on all data segments under each target category, determine the analysis parameters; S 13 、Convert the time-domain vibration data in each data segment into angular domain equi-periodic vibration data.

[0012] According to an embodiment of the present invention, the step S 11 comprises the following steps: S 111 、Determine the target diagnosis category, and denote the label corresponding to the i-th category as C i , i = 1, 2, …, C 0 , where C 0 is the total number of categories to be diagnosed, including 1 normal category and C 0 -1 fault categories; S 112 、Based on the historical operation data of the engine, obtain the rolling bearing vibration data segments corresponding to each target category, and each data segment includes: time-domain vibration data, time-domain rotational speed pulse data, and category label, and denote the j-th data segment under category C i as {Vij , P ij , C i}; Among them, V corresponds to the time-domain vibration data, P corresponds to the time-domain rotational speed pulse data, and C corresponds to the class label.

[0013] According to an embodiment of the present invention, in the step S 12 , the analysis parameters include the highest analysis order and the minimum order resolution. The step S 12 includes the following steps: S 121 Determine the highest analysis order N eo : Among them, Fs is the sampling frequency of the vibration data, and N max is the highest physical rotational speed corresponding to all class data segments, with the unit of rpm, and round(·) is the rounding operator; S 122 Determine the minimum order resolution Δf: Calculate the characteristic frequency orders F 0 corresponding to C i - 1 types of bearing fault classes respectively, where i = 1,..., C 0 - 1, and obtain the minimum characteristic frequency order F min : Among them, min(·) is the minimum value operator; the minimum order resolution Δf is determined by the following formula: Among them, ceil(·) is the ceiling operator.

[0014] According to an embodiment of the present invention, in the step S 13 under the class C i , the j-th data segment is {V ij , P ij , C i}. The step S 13 includes the following steps: S 131 Select the number of points N itp to be interpolated in each rotation period and meet the following requirements: N itp ≥ round(2.56 × N eo ); S 132 Divide the time range corresponding to the first rotation period into N itp equal parts, and obtain the corresponding moments T i , i = 1, 2,...,

[0015] N itp , and linearly interpolate the corresponding time-domain vibration signal at the above equal-point moments to obtain the angular-domain equi-periodic vibration data in this rotation period; S 133 Along the time axis, according to the step S 131 to the step S 132Obtain the angular domain equal-period vibration data within all rotation periods in sequence, and after merging, form the angular domain equal-period vibration data A corresponding to this data segment i,j .

[0016] According to an embodiment of the present invention, in the step S 2 : In category C i , the time-frequency matrix corresponding to the jth data segment is: TF ij = stft(A i,j ); where stft(·) is the short-time Fourier transform operator, and the window scale during the transformation process is The obtained time-frequency matrix TF i corresponding to the jth data segment in category C ij is a two-dimensional matrix, with the horizontal direction being the time dimension and the vertical direction being the frequency dimension.

[0017] According to an embodiment of the present invention, the steps of generating the three-dimensional tensor of the full-frequency domain difference between different categories in the step S 3 include: S 31 Generate the full-frequency domain difference vectors between any two categories, a total of ones, and the dimension of each vector is S 32 After obtaining all difference vectors, the three-dimensional tensor S of the full-frequency domain difference between different categories is expressed as follows: Among them, S is a symmetric structure, and its dimension is s ij represents the full-band difference vector between category C i and C j , and the dimension is

[0018] According to an embodiment of the present invention, in the step S 31 , under category C i and category C j , denote the full-band difference vector between C i and C j as s ij , and the solution process of s ij includes the following steps: S 311 Randomly shuffle each column of RTF (i) and RTF (j) along the time dimension to obtain RTF′ (i) and RTF′ (j) , and calculate the value of the kth element of s ij after this shuffling order: Among them, L = min(L i , L j ) is the number of effective comparison points, Li , L j are respectively RTF (i) and RTF (j) along the time dimension length, represents the element in the k-th row and the n-th column of RTF′ (i) ; represents the element in the k-th row and the n-th column of RTF′ (j) ; represents the value of the k-th element after this shuffling order; S ij ; 312 Repeat the above step S 311 for multiple times, and take the average value as the final value; S 313 Repeat the above step S 311 and the above step S 312 until all elements in s ij are solved.

[0019] According to an embodiment of the present invention, in the above step S 3 , setting a diagnostic accuracy rate threshold and a diagnostic balance threshold, setting an initial tensor weight matrix, and the step of adjusting the difference tensor based on the tensor weight matrix includes: S 33 Set a diagnostic accuracy rate threshold A thr and a diagnostic balance threshold B thr , and the initial tensor weight matrix W is expressed as follows: S 34 Adjust the difference tensor based on the tensor weight matrix W as follows: where S′ is the difference tensor adjusted based on the weight W, and S is the three-dimensional full-frequency domain difference tensor S between categories.

[0020] According to an embodiment of the present invention, the above step S 4 includes the following steps: S 41 For all categories to be diagnosed, calculate the full-frequency domain difference discrimination ability Q, and define the full-frequency domain discrimination ability Q i between category C i and all other categories as: Then, where σ(·) represents taking the standard deviation of all terms in the brackets, and the obtained Q orig has a dimension of Based on the above formula, normalize each element in Q orig to obtain Q, and its i-th element is expressed as: S 42 Determine the optimal frequency band and generate samples.

[0021] According to an embodiment of the present invention, the step S 42 includes the following steps: S 421 , determine a reasonable two-dimensional sample scale L t and L f and keep them unchanged, which respectively represent the lengths of a single sample in the time dimension and the frequency dimension; S 422 , sort Q in descending order and record the first L f frequency points; S 423 , based on the time-frequency matrices of various categories obtained in the step S 2 , extract and combine the rows where the recorded L f frequency points are located to obtain a new frequency doubling matrix. The frequency dimension of the new matrix is L f , and it is divided into samples of size L t along the time dimension; S 424 , obtain the sample sets of all categories.

[0022] According to an embodiment of the present invention, the step S 5 includes the following steps: S 51 , randomly divide all samples into a training set, a validation set, and a test set, and the proportion of samples contained in the three is 6:2:2; S 52 , construct a two-dimensional convolutional neural network model and keep the structure unchanged, and perform model training based on the training set and the validation set to obtain an initial fault diagnosis model M; S 53 , use the sample of the test set divided in the step S 51 to perform model testing, obtain the predicted label and the probability vector of each category corresponding to each sample. The probability vector and the predicted label output by the i-th sample d i in the test set can be respectively expressed as:

[0023] where predict(a,b) represents using model a to predict sample b; p ij represents the probability value that the i-th sample is predicted to be category j; p i represents the probability vector output by the i-th sample d i in the test set, represents the predicted label output by the i-th sample d i in the test set.

[0024] According to an embodiment of the present invention, the step S 6 includes the following steps: S 61 , based on the result obtained in the step S 53 , obtain the confusion matrix F of the diagnosis result on the test set as follows: where Fij Denote the number of samples predicted as C i from C j ; calculate the prediction accuracy A and balance B of the calculation model M for all samples in the test set:

[0025] Wherein, S 62 , based on the prediction labels and probability vectors of all samples in the test set obtained in the step S 53 , calculate the weight increment matrix W'.

[0026] According to an embodiment of the present invention, the step S 62 includes the following steps: S 621 , initialize W' with a zero matrix, that is: S 622 , based on the probability vector and prediction label obtained in the step S 53 , complete the correction of each element in the weight increment matrix W'. In the first sample of the test set, assume that the prediction result of the model is p 1 and and the true label of the first sample is If c pred,1 ≠c true,1 , then update the element in the th row and th column of W' as follows: Wherein, represents the number of samples with the label that are mispredicted as the label among all samples, represents the element in the th row and th column of W'.

[0027] According to an embodiment of the present invention, the step S 7 includes the following steps: S 71 , repeatedly execute the step S 51 , the step S 52 , the step S 53 , the step S 61 and the step S 62 multiple times, and average the diagnostic accuracy A, balance B, and correction matrix W' obtained in each cycle to obtain the final diagnostic accuracy A, balance B, and weight increment matrix W'; S 72 , determine whether the following formula is satisfied simultaneously: If not, correct the tensor weight matrix W in the step S 34 : W = W + W'; and sequentially repeat the execution of the step S34 and the step S 41 and the step S 42 and the step S 51 and the step S 52 and the step S 53 and the step S 61 and the step S 62 and the step S 71 and the step S 72 ; if satisfied, stop the training and output the L 42 frequency points in step S f and the model M obtained from the last training in step S 51 .

[0028] The positive and progressive effects of the present invention are as follows:

[0029] The intelligent fault diagnosis method for rolling bearings based on error self-correction of the present invention has at least the following advantages:

[0030] 1. This method is a fast evaluation method for full-band discrimination of category differences, driven by data, which can achieve refined evaluation under the minimum frequency domain resolution, avoiding the limitations of traditional manual analysis while improving the evaluation efficiency.

[0031] 2. This method can be used for fast evaluation of full-band discrimination in single or multiple fault modes, and the upper frequency limit is the highest multiple frequency that can be obtained under the sampling frequency.

[0032] 3. This method establishes a band selection evaluation method for all categories with variable weights, which comprehensively considers the intra-category consistency and inter-category differences, and adjusts the weights automatically through the fault diagnosis results.

[0033] 4. This method establishes a bearing fault diagnosis model with an error self-regulation mechanism of input-model-input based on two-dimensional time-frequency input and convolutional neural network, realizing intelligent fault diagnosis of bearings. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The above-mentioned and other features, properties and advantages of the present invention will become more obvious through the following description in conjunction with the drawings and embodiments, in which the same reference numerals always represent the same features:

[0035] Figure 1 is the overall flow schematic diagram of the intelligent fault diagnosis method for rolling bearings based on error self-correction of the present invention.

[0036] Figure 2 is the schematic diagram of the evolution process of the full-frequency domain difference discrimination ability under the action of weights obtained by five iterations in an embodiment of the intelligent fault diagnosis method for rolling bearings based on error self-correction of the present invention.

[0037] Figure 3 It is a schematic diagram showing the change process of the fault diagnosis accuracy rate and the standard deviation of the accuracy rate between categories under five iterations in an embodiment of the intelligent fault diagnosis method for rolling bearings based on error self-correction of the present invention.

[0038] Figure 4 It is a confusion diagram of the diagnosis results on the test set during a certain model training in the fifth iteration process in an embodiment of the intelligent fault diagnosis method for rolling bearings based on error self-correction of the present invention. Detailed implementation manners

[0039] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific implementation manners of the present invention will be given in conjunction with the accompanying drawings.

[0040] Now, embodiments of the present invention will be described in detail with reference to the accompanying drawings. Now, preferred embodiments of the present invention will be described in detail, and examples thereof are shown in the drawings. In any possible case, the same reference numerals will be used throughout the drawings to represent the same or similar parts. In addition, although the terms used in the present invention are selected from well-known and commonly used terms, some of the terms mentioned in the specification of the present invention may be selected by the applicant according to his or her judgment, and their detailed meanings are described in the relevant parts of the description herein. In addition, it is required to understand the present invention not only through the actual terms used, but also through the meaning implied by each term.

[0041] As Figure 1 shown, the present invention discloses an intelligent fault diagnosis method for rolling bearings based on error self-correction, and the intelligent fault diagnosis method includes the following steps:

[0042] Step S 1 , Obtain the equal-period vibration data in the angular domain of the rolling bearing corresponding to each target category.

[0043] Step S 2 , Based on the short-time Fourier transform method, obtain the total time-frequency matrix for each category.

[0044] Step S 3 , Generate a three-dimensional tensor of the full-frequency domain difference between categories, set a diagnosis accuracy rate threshold and a diagnosis balance threshold, set an initial tensor weight matrix, and adjust the difference tensor based on the tensor weight matrix.

[0045] Step S 4 , For all categories to be diagnosed, calculate the full-frequency domain difference discrimination ability, determine the optimal frequency band and perform sample generation.

[0046] Step S 5, based on the generated sample sets of each category, the fault diagnosis model is trained and verified, and the model is tested.

[0047] Step S 6 , calculate the prediction accuracy and balance of the fault diagnosis model, and calculate the weight increment matrix.

[0048] Step S 7 , evaluate the accuracy and balance of the fault diagnosis model to determine whether the prediction accuracy and balance of the fault diagnosis model meet the conditions; if not, use the weight increment matrix to perform input and model optimization based on error self-correction. If satisfied, output the optimal input structure and fault diagnosis model.

[0049] The intelligent fault diagnosis method for rolling bearings based on error self-correction of the present invention has the characteristics of input optimization and model error self-correction, and can be applied to intelligent fault diagnosis modeling of rolling bearings of aircraft engines. It improves the low efficiency caused by traditional diagnostic methods relying on manual experience and theoretical models, and the inability to balance diagnostic accuracy and balance, and can realize high-efficiency and intelligent fault diagnosis of rolling bearings.

[0050] The error self-correction is to automatically correct the input based on the prediction error during the fault diagnosis model training process, and re-train the model based on the corrected input.

[0051] The intelligent fault diagnosis uses artificial intelligence technology to automatically judge the status of input information instead of manual work.

[0052] like Figure 1 As shown, as a preferred implementation of the rolling bearing intelligent fault diagnosis method based on error self-correction of the present invention, the step S 1 The following steps may be included:

[0053] Step S 11 , for the target diagnosis category, based on the historical engine operation data, obtain the rolling bearing vibration data segments corresponding to each target category.

[0054] Step S 12 , determine the analysis parameters based on all data fragments under each target category.

[0055] Step S 13 , convert the time domain vibration data in each data segment into angular domain iso-periodic vibration data.

[0056] The preferred specific implementation steps of the rolling bearing intelligent fault diagnosis method based on error self-correction of the present invention are as follows:

[0057] The step S 11 The following steps may be included:

[0058] Step S 111 , determine the target diagnosis category, and denote the label corresponding to the i-th category as C i , where i = 1, 2, …, C 0 , and C 0 is the total number of categories to be diagnosed, including 1 normal category and C 0 - 1 fault categories.

[0059] Step S 112 , based on the historical operation data of the engine, obtain the rolling bearing vibration data segments corresponding to each target category. Each data segment consists of three parts: time-domain vibration data, time-domain rotational speed pulse data, and category label. Denote the j-th data segment under category C i as {V ij , P ij , C i}, where V, P, and C correspond to the above three parts respectively.

[0060] The analysis parameters in the said Step S 12 include the highest analysis order and the minimum order resolution. The said Step S 12 may include the following steps:

[0061] Step S 121 , determine the highest analysis order N eo :

[0062]

[0063] where Fs is the sampling frequency of the vibration data, and N max is the highest physical rotational speed corresponding to the data segments of all categories, with the unit of rpm, and round(·) is the rounding operator.

[0064] Step S 122 , determine the minimum order resolution Δf:

[0065] Calculate the characteristic frequency orders corresponding to C 0 - 1 bearing fault categories respectively, and denote them as F i , where i = 1, …, C 0 - 1, and obtain the minimum characteristic frequency order F min :

[0066]

[0067] where min(·) is the minimum value operator;

[0068] On this basis, the minimum order resolution Δf is determined by the following formula:

[0069]

[0070] Among them, ceil(·) is the ceiling operator.

[0071] In the said step S 13 , taking the j-th data segment under the category C i as an example, it includes {V ij , P ij , C i}, and the transformation process may include the following steps:

[0072] Step S 131 , select the number of points N to be interpolated in each rotation period itp and meet the following requirements:

[0073] N itp ≥ round(2.56×N eo )

[0074] Step S 132 , divide the time range corresponding to the first rotation period into N itp equal parts, obtain the moments T i corresponding to each equal division point, i = 1, 2,..., N itp , linearly interpolate the corresponding time-domain vibration signal at the above-mentioned equal division point moments to obtain the angular-domain equi-periodic vibration data within this rotation period.

[0075] Step S 133 , along the time axis, sequentially obtain the angular-domain equi-periodic vibration data within all rotation periods according to the said step S 131 to the said step S 132 , and after merging, form the angular-domain equi-periodic vibration data corresponding to this data segment, denoted as A i,j .

[0076] Step S 2 , based on the short-time Fourier transform (stft) method, obtain the total time-frequency matrix under each category. Taking the category C i as an example, the time-frequency matrix corresponding to its j-th data segment is:

[0077] TF ij = stft(A i,j )

[0078] Among them, stft(·) is the short-time Fourier transform operator, and the window scale of the transformation process is The obtained time-frequency matrix TF i corresponding to the j-th data segment in the category C ij is a two-dimensional matrix, with the horizontal direction being the time dimension and the vertical direction being the frequency dimension.

[0079] On this basis, retain TFij The matrix takes the first data points along the frequency dimension to obtain class C i The time-frequency matrix RTF corresponding to the jth data segment ij ; Using the same method, obtain the time-frequency matrices corresponding to all data segments under class C i and merge them along the time dimension to form the total time-frequency matrix under class C i , denoted as RTF (i) .

[0080] In the step S 3 , generating the three-dimensional tensor of the full-frequency domain difference between classes may include the following steps:

[0081] Step S 31 、Generate the full-frequency domain difference vectors between any two classes, a total of ones, and the dimension of each vector is . The method is as follows:

[0082] Taking class C i and class C j as an example, denote the full-band difference vector between C i and C j as s ij , and its solution process is as follows:

[0083] Step S 311 、Randomly shuffle the columns of RTF (i) and RTF (j) along the time dimension, that is, randomly arrange the columns to obtain RTF′ (i) and RTF′ (j) , and calculate the value of the kth element of s ij after this shuffle:

[0084]

[0085] where L = min(L i , L j ) is the number of effective comparison points, and L i , L j are the lengths of RTF (i) and RTF (j) along the time dimension respectively, represents the nth element of the kth row of RTF′ (i) , Similarly, represents the value of the kth element of s ij after this shuffle.

[0086] Step S 312 、Repeat the above step S 311 multiple times to obtain multiple Its average value is used as the final value;

[0087] Step S 313 is repeated, and the above-mentioned step S 311 and the above-mentioned step S 312 are repeated until all elements in s ij are solved.

[0088] After step S 32 obtains all difference vectors, the full-frequency domain difference three-dimensional tensor S can be expressed as follows:

[0089]

[0090] where S is a symmetric structure, and its dimension is s ij represents the full-frequency band difference vector between category C i and C j , and its dimension is

[0091] In the above-mentioned step S 3 , setting the diagnostic accuracy rate threshold and the diagnostic balance threshold, and setting the initial tensor weight matrix, the steps of adjusting the difference tensor based on the tensor weight matrix may include:

[0092] Step S 33 sets the diagnostic accuracy rate threshold A thr and the diagnostic balance threshold B thr , and the initial tensor weight matrix W is expressed as follows:

[0093]

[0094] Step S 34 adjusts the difference tensor based on the tensor weight matrix W as follows:

[0095]

[0096] where S′ is the difference tensor adjusted based on the weight W.

[0097] The above-mentioned step S 4 may include the following steps:

[0098] Step S 41 calculates the full-frequency domain difference discrimination ability Q for all categories to be diagnosed, and defines the full-frequency domain discrimination ability Q i between category C i and all other categories as:

[0099]

[0100] Then,

[0101]

[0102] where σ(·) represents taking the standard deviation of all terms within the parentheses, and the resulting Q orig has a dimension of

[0103] Based on the above formula, normalize each element in Q orig to obtain Q, and its i-th element is expressed as:

[0104]

[0105] Step S 42 , determine the optimal frequency band and generate samples. The said Step S 42 may include the following steps:

[0106] Step S 421 , according to the computer processing capacity, determine a reasonable two-dimensional sample scale L t and L f and keep them unchanged, where the two respectively represent the lengths of a single sample in the time dimension and the frequency dimension.

[0107] Step S 422 , sort Q in descending order and record the first L f frequency points.

[0108] Step S 423 , based on the time-frequency matrices RTF of each category obtained in the said Step S 2 , extract and combine the rows where the recorded L f frequency points are located to obtain a new octave matrix. The frequency dimension of the new matrix is L f , and it is divided into samples of size L t along the time dimension.

[0109] Step S 424 , use the above method to obtain the sample sets of all categories.

[0110] The said Step S 5 may include the following steps:

[0111] Step S 51 , randomly divide all samples into a training set, a validation set, and a test set, and the sample ratios of the three are 6:2:2.

[0112] Step S 52 , construct a two-dimensional convolutional neural network model and keep the structure unchanged, and perform model training based on the training set and the validation set to obtain an initial fault diagnosis model, denoted as M.

[0113] Step S 53 、Use the test set samples divided in the said Step S 51 to perform model testing, and obtain the predicted labels and category probability vectors corresponding to each sample. Taking the i-th sample d i in the test set as an example, its output probability vector p i and predicted label c pred can be respectively expressed as:

[0114]

[0115]

[0116] Among them, predict(a,b) represents using model a to predict sample b; p ij represents the probability value that the i-th sample is predicted as category j; represents the i-th sample d i in the test set

[0117] The said Step S 6 may include the following steps:

[0118] Step S 61 、Based on the result obtained in the said Step S 53 obtain the confusion matrix F of the diagnostic results on the test set as follows:

[0119]

[0120] Among them, F ij represents the number of samples that are predicted as C i and predicted as C j ;

[0121] On this basis, calculate the prediction accuracy A and balance B of model M for all samples in the test set:

[0122]

[0123]

[0124] Among them,

[0125]

[0126] Step S 62 、Based on the predicted labels and probability vectors of all samples in the test set obtained in the said Step S 53 calculate the weight increment matrix W′. The said Step S 62 may include the following steps:

[0127] Step S 621, initialize W′ with a zero matrix, i.e.:

[0128]

[0129] Step S 622 , based on the predicted probability vectors and labels of the test set samples obtained in the above-mentioned Step S 53 , complete the correction of each element in the weight increment matrix W′. Taking the first sample in the test set as an example, assume the prediction result of the model is p 1 and and the true label of the first sample is If then update the element in the th row and th column of W′ as follows:

[0130]

[0131] where represents the number of samples with label that are mispredicted as label among all samples, and represents the element in the th row and th column of W′.

[0132] The above-mentioned Step S 7 may include the following steps:

[0133] Step S 71 , repeatedly execute the above-mentioned Step S 51 , the above-mentioned Step S 52 , the above-mentioned Step S 53 , the above-mentioned Step S 61 and the above-mentioned Step S 62 , and average the diagnostic accuracy rate A, balance B, and correction matrix W′ obtained in each cycle to obtain the final diagnostic accuracy rate A, balance B, and weight increment matrix W′.

[0134] Step S 72 , determine whether the following formula is satisfied simultaneously:

[0135]

[0136] If not, correct the tensor weight matrix W in the above-mentioned Step S 34 :

[0137] W = W + W′

[0138] and then repeatedly execute the above-mentioned Step S 34 , the above-mentioned Step S 41, the step S 42 , the step S 51 , the step S 52 , the step S 53 , the step S 61 , the step S 62 , the step S 71 and the step S 72 ;

[0139] If satisfied, stop the training and output the L 42 in the step S f frequency points and the model M obtained in the last training in the step S 51 .

[0140] The intelligent fault diagnosis method for rolling bearings based on error self - correction of the present invention has the characteristics of input optimization and model error self - correction, can be applied to the intelligent fault diagnosis modeling of rolling bearings in aero - engines, and improves the problems such as low efficiency caused by the dependence on manual experience and theoretical models in traditional diagnosis methods, and the inability to balance diagnostic accuracy and balance. It can achieve high - efficiency and intelligent fault diagnosis of rolling bearings.

[0141] The method of the present invention is an intelligent fault diagnosis modeling method that can automatically focus on effective inputs and balance diagnostic accuracy and balance, and can achieve fast and reliable fault diagnosis of rolling bearings in aero - engines.

[0142] A specific embodiment of fault diagnosis using the above - mentioned intelligent fault diagnosis method for rolling bearings based on error self - correction is as follows:

[0143] Taking the fault diagnosis of a certain rolling bearing as an example, the target categories for diagnosis are divided into four categories: normal - 0, inner - ring fault - 1, rolling - element fault - 2, and outer - ring fault - 3. The numbers after "-" are the corresponding category labels respectively.

[0144] Set the set diagnosis correct - rate threshold A thr and the diagnosis balance threshold B thr to be 94% and 3% respectively. The number of samples in the training set, validation set, and test set are 672, 224, and 224 respectively. The fault diagnosis model training process meets the above requirements after optimizing the input 5 times. The average correct - rate and the standard deviation of the correct - rate of the finally obtained fault diagnosis model on the test set are 94.64% and 1.49% respectively.

[0145] As Figure 2 shown, it intuitively shows the evolution process of the full - frequency - domain difference discrimination ability under the weights obtained in five iterations. The key difference frequency bands identified in each iteration are within the dashed boxes.

[0146] As Figure 3As shown, it intuitively shows the change process of the fault diagnosis accuracy rate and the standard deviation of the accuracy rate between categories for each iteration.

[0147] As Figure 4 shown, it is the confusion diagram of the diagnosis results on the test set after a certain training during the 5th iteration process (A = 93.30%, B = 1.96%).

[0148] Through Figures 2 to 3 the following conclusions can be obtained:

[0149] (1) Different frequency bands in the full frequency domain have different abilities to distinguish the differences between the above four categories. Obviously, the higher the discrimination degree of the frequency band, the more effective fault diagnosis information it carries.

[0150] (2) There is no definite relationship between the fault diagnosis accuracy rate and the diagnosis balance (i.e., the difference in the diagnosis effects for different categories). The increase of the former does not necessarily lead to the decrease of the latter. For example, Figure 3 in the 2nd and 3rd iterations in , and both are crucial for the evaluation of the fault diagnosis effect.

[0151] (3) Observing Figure 3 the change trend in , it can be seen that by adopting the fault diagnosis method proposed in this patent, the accuracy rate and balance of fault diagnosis can be taken into account, and the fault diagnosis effect is improved.

[0152] The above results prove the effectiveness of the rolling bearing fault diagnosis method proposed in this patent.

[0153] In summary, the intelligent fault diagnosis method of rolling bearing based on error self-correction of the present invention has the following many advantages:

[0154] First, this method is a fast evaluation method for the full-band discrimination degree of category differences, driven by data, which can realize the refined evaluation under the minimum frequency domain resolution, avoiding the limitations of traditional manual analysis while improving the evaluation efficiency.

[0155] Second, this method can be used for the fast evaluation of the full-band discrimination degree in single or multiple fault modes, and the upper limit of the frequency band is the highest multiple frequency that can be obtained under the sampling frequency.

[0156] Third, this method establishes a band selection evaluation method for all categories with variable weights, which comprehensively considers the within-category consistency and between-category differences, and adjusts the weights automatically through the fault diagnosis results.

[0157] Fourth, this method establishes a bearing fault diagnosis model with an error self-regulation mechanism of input-model-input based on two-dimensional time-frequency input and convolutional neural network to realize the intelligent fault diagnosis of bearings.

[0158] Although the specific embodiments of the present invention have been described above, those skilled in the art should understand that these are only illustrative examples, and the protection scope of the present invention is defined by the appended claims. Without departing from the principle and essence of the present invention, those skilled in the art can make various changes or modifications to these embodiments, but these changes and modifications all fall within the protection scope of the present invention.

Claims

1. An intelligent fault diagnosis method for rolling bearings based on error self-correction, Characterized in that, The intelligent fault diagnosis method includes the following steps: S 1 Obtain the angular domain equal-cycle vibration data of the rolling bearings corresponding to each target category; S 2 , obtain the total time-frequency matrix for each category based on the short-time Fourier transform method; S 3 , generate a three-dimensional tensor of full-frequency domain differences between various categories, set a diagnostic accuracy threshold and a diagnostic balance threshold, set an initial tensor weight matrix, and adjust the difference tensor based on the tensor weight matrix; S 4 For all the categories to be diagnosed, calculate the full-frequency domain discrimination ability, determine the optimal frequency band and generate samples; S 5 Train and validate the fault diagnosis model based on the generated sample sets of various categories, and conduct model testing; S 6 、Calculate the prediction accuracy and balance of the fault diagnosis model, and calculate the weight increment matrix; S 7 Perform the evaluation of the accuracy and balance of the fault diagnosis model to determine whether the prediction accuracy and balance of the fault diagnosis model meet the conditions; If not satisfied, use the weight increment matrix for input and model optimization based on error self-correction. If satisfied, output the obtained optimal input structure and fault diagnosis model.

2. The intelligent fault diagnosis method for rolling bearings based on error self-correction according to claim 1, Characterized in that, The said step S 1 comprises the following steps: S 11 、For the target diagnosis category, based on the historical operation data of the engine, obtain the vibration data segments of the rolling bearings corresponding to each target category; S 12 Determine analysis parameters based on all data segments under each target category; S 13 Convert the time-domain vibration data within each data segment into angular-domain equi-periodic vibration data.

3. The intelligent fault diagnosis method for rolling bearings based on error self-correction according to claim 2, Characterized in that, The said step S 11 comprises the following steps: S 111 Determine the target diagnosis category, and denote the label corresponding to the i-th category as C i , where i = 1, 2, …, C 0 , and C 0 is the total number of categories to be diagnosed, including 1 normal category and C 0 - 1 fault categories; S 112 、Based on the historical operation data of the engine, obtain the vibration data segments of the rolling bearings corresponding to each target category. Each data segment includes: time-domain vibration data, time-domain rotational speed pulse data, and category labels. Denote the j-th data segment under category C i as {V ij , P ij , C i}; Among them, V corresponds to time-domain vibration data, P corresponds to time-domain rotational speed pulse data, and C corresponds to category labels.

4. The intelligent fault diagnosis method for rolling bearings based on error self-correction according to claim 3, Characterized in that, The said step S 12 in which the analysis parameters include the highest analysis order and the minimum order resolution, the said step S 12 comprises the following steps: S 121 Determine the highest analysis order N eo : where Fs is the vibration data sampling frequency, and N max is the highest physical rotational speed corresponding to all category data segments, with the unit of rpm, and round(·) is the rounding operator; S 122 , determine the minimum order resolution Δf: Calculate C separately 0 - The characteristic frequency order F corresponding to one bearing fault category i , where i = 1, …, C 0 -1, and obtain the minimum characteristic frequency order F min : Among them, min(·) is the minimum value operator; The minimum order resolution Δf is determined by the following formula: Among them, ceil(·) is the ceiling operator.

5. The intelligent fault diagnosis method for rolling bearings based on error self-correction according to claim 4, Characterized in that, The said step S 13 in category C i the j-th data segment is {V ij , P ij , C i}, and the said step S 13 includes the following steps: S 131 Select the number of points N to be interpolated in each rotation period itp and meet the following requirements: N itp ≥ round(2.56 × N eo ) S 132 Divide the time range corresponding to the first rotation period into N itp equal parts to obtain the moments T corresponding to each equal division point i , where i = 1, 2, …, N itp , linearly interpolate the corresponding time-domain vibration signals at the above-mentioned moments of each equal division point to obtain the angular-domain equi-periodic vibration data within this rotation period; S 133 Along the time axis, according to the said step S 131 to the said step S 132 successively obtain the angular domain equi-periodic vibration data within all rotation periods, and after merging, form the angular domain equi-periodic vibration data A corresponding to this data segment i,j .

6. The intelligent fault diagnosis method for rolling bearings based on error self-correction according to claim 5, Characterized in that, The said step S 2 Among them: In category C i The time-frequency matrix corresponding to the j-th data segment is as follows: TF ij = stft(A i，j ) where stft(·) is the short-time Fourier transform operator, and the window scale during the transformation process is The obtained time-frequency matrix TF corresponding to the j-th data segment in class C i is a two-dimensional matrix, with the horizontal direction being the time dimension and the vertical direction being the frequency dimension. ij ​ 7. The intelligent fault diagnosis method for rolling bearings based on error self-correction according to claim 6, Characterized in that, The step S 3 The step of generating a three-dimensional tensor of full-frequency domain differences between various categories in S 31 Generate full-frequency domain difference vectors between any two categories, with a total of ones, and the dimension of each vector is S 32 , after obtaining all differential vectors, the three-dimensional tensor S of the full-frequency domain difference between various categories is expressed as follows: Among them, S is a symmetric structure, and its dimension is s ij represents the full-band difference vector between class C i and C j , and its dimension is 8. The intelligent fault diagnosis method for rolling bearings based on error self-correction according to claim 7, Characterized in that, The said step S 31 In category C i and category C j under, denote the full-band difference vector between C i and C j as s ij , the solution process of s ij includes the following steps: S 311 Randomly shuffle the RTF along the time dimension (i) and the RTF (j) for each column to obtain RTF' (i) and RTF' (j) , and calculate the value of the k-th element of s ij after this shuffle: where, L = min(L i , L j ) is the number of valid comparison points, L i , L j are the lengths of RTF (i) and RTF (j) along the time dimension respectively, represents the element at the n-th column and the k-th row of RTF′ (i) , represents the element at the n-th column and the k-th row of RTF′ (j) , represents the value of the k-th element of s after this shuffling ij ; S 312 Repeat the step S multiple times 311 to obtain multiple and use their average value as the final value; S 313 Repeat the said step S 311 and the said step S 312 until all elements in s ij are solved.

9. The intelligent fault diagnosis method for rolling bearings based on error self-correction according to claim 8, Characterized in that, The step S 3 includes setting a diagnostic accuracy threshold and a diagnostic balance threshold, setting an initial tensor weight matrix, and adjusting the difference tensor based on the tensor weight matrix, as follows: S 33 Set the diagnostic accuracy rate threshold A thr and the diagnostic balance threshold B thr , and set the initial tensor weight matrix W as follows: S 34 , adjust the difference tensor based on the tensor weight matrix W as follows: Among them, S′ is the difference tensor adjusted based on the weight W, and S is the three-dimensional full-frequency domain difference tensor S between categories.

10. The intelligent fault diagnosis method for rolling bearings based on error self-correction according to claim 9, Characterized in that, The said step S 4 comprises the following steps: S 41 For all categories to be diagnosed, calculate the full-frequency domain discrimination ability Q, and define the full-frequency domain discrimination ability Q i between category C i and all other categories as: Then there is, where σ(·) represents taking the standard deviation of all terms within the parentheses, and the resulting Q orig The dimension is Based on the above formula, perform normalization on each element in Q orig to obtain Q, and its i-th element is expressed as: S 42 Determine the optimal frequency band and generate samples.

11. The intelligent fault diagnosis method for rolling bearings based on error self-correction according to claim 10, Characterized in that, The said step S 42 comprises the following steps: S 421 Determine a reasonable two-dimensional sample scale L according to the computer processing capacity t and L f and keep them unchanged, where they respectively represent the lengths of a single sample in the time dimension and the frequency dimension; S 422 Sort Q in descending order and record the first L f frequency points; S 423 Based on the time-frequency matrices under various categories obtained in the above step S 2 , extract and combine the rows where the recorded L f frequency points are located to obtain a new frequency-doubling matrix. The frequency dimension of the new matrix is L f , and it is segmented along the time dimension into samples of size L t . S 424 Obtain the sample sets under all categories.

12. The intelligent fault diagnosis method for rolling bearings based on error self-correction according to claim 11, Characterized in that, The said step S 5 comprises the following steps: S 51 , all samples are randomly divided into a training set, a validation set, and a test set, and the proportion of samples contained in the three is 6:2:2; S 52 Construct a two-dimensional convolutional neural network model and keep its structure unchanged. Train the model based on the training set and the validation set to obtain the initial fault diagnosis model M; S 53 、Use the test set samples divided in the above step S 51 to perform model testing, obtain the predicted labels and category probability vectors corresponding to each sample. For the i-th sample d i in the test set, the output probability vector and predicted label can be respectively expressed as: Among them, predict(a,b) represents using model a to predict sample b; p ij represents the probability value that the i-th sample is predicted as class j; p i represents the i-th sample d in the test set i output probability vector, represents the i-th sample d in the test set i output predicted label.

13. The intelligent fault diagnosis method for rolling bearings based on error self-correction according to claim 12, Characterized in that, The said step S 6 includes the following steps: S 61 Based on the result obtained from the above step S 53 the following confusion matrix F of the diagnostic result on the test set is obtained: Among them, F ij represents the number of samples that i predict C as j C; Calculate the prediction accuracy A and balance B of the model M for all samples in the test set: Among them, S 62 Based on the predicted labels and probability vectors of all samples in the test set obtained in the above step S 53 , calculate the weight increment matrix W'.

14. The intelligent fault diagnosis method for rolling bearings based on error self-correction according to claim 13, Characterized in that, The said step S 62 comprises the following steps: S 621 , initialize W′ with a zero matrix, i.e.: S 622 Based on the probability vector and the predicted label obtained in the above step S 53 , complete the correction of each element in the weight increment matrix W'. In the first sample of the test set, assume that the prediction result of the model is p 1 and and the true label of the first sample is If c pred,1 ≠c true,1 , then the update of the -th row and -th element in W' is as follows: Among them, represents the number of samples in all samples that are mispredicted as the label to the label . represents the -th row and the -th element in W'.

15. The intelligent fault diagnosis method for rolling bearings based on error self-correction according to claim 14, Characterized in that, The said step S 7 comprises the following steps: S 71 and repeatedly execute the said step S 51 the said step S 52 the said step S 53 the said step S 61 and the said step S 62 , and average the diagnostic accuracy rate A, balance B, and correction matrix W′ obtained in each cycle to obtain the final diagnostic accuracy rate A, balance B, and weight increment matrix W′; S 72 , determine whether the following formula is satisfied simultaneously: If not satisfied, correct the tensor weight matrix W in the step S 34 as follows: W = W + W′ and sequentially repeat the execution of the said step S 34 、the said step S 41 、the said step S 42 、the said step S 51 、the said step S 52 、the said step S 53 、the said step S 61 、the said step S 62 、the said step S 71 and the said step S 72 ; If the condition is satisfied, stop the training and output L in step S 42 and the model M obtained from the last training in step S f at L frequency points 51 ​

Images