Vibration-based intelligent fault diagnosis of rotating machinery large model modeling method

Abstract

A vibration-based intelligent fault diagnosis large model modeling method for rotating machinery: in terms of data, a monitoring big data time-frequency structure alignment method is proposed, which normalizes different scene vibration data and aligns time-frequency information, solving the problems of data amplitude difference and time-frequency information misalignment caused by cross-working condition, cross-equipment and cross-sensor; in terms of model, a multi-scale state collaborative intelligent diagnosis network is proposed to solve the problem of multi-entity data and multi-health state; in terms of adaptation, a target entity model self-adaptive fine-tuning adaptation method is proposed to solve the misclassification problem caused by field difference, so that the model can better adapt to the target scene; the present application solves the problem that the existing model is difficult to be used in multiple entity equipment, and realizes the purpose of rapid adaptation and generalization of the target equipment by constructing an intelligent fault diagnosis large model.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and health management technology for rotating machinery and equipment, specifically to a vibration-based intelligent fault diagnosis large-scale modeling method for rotating machinery. Background Technology

[0002] Traditional intelligent fault diagnosis methods perform well under single-device, single-operating-condition diagnostic models. However, when dealing with cross-device and variable-condition problems, these models struggle to generalize effectively to the target device. Some researchers have employed intelligent transfer diagnosis methods based on transfer learning, such as domain adaptation, to generalize to target domain devices. However, these methods have limitations regarding the data distribution in the source and target domains.

[0003] The field of intelligent fault diagnosis aims to leverage massive amounts of data to build large-scale intelligent fault diagnosis models, and to use these models to quickly and efficiently address the shortcomings of existing intelligent diagnosis methods in terms of poor generalization ability to the target domain (LI Yan-Fu, WANGHuan, SUN Muxia. ChatGPT-like large-scale foundation models for prognostics and health management: A survey and roadmaps[J]. Reliability Engineering & System Safety, 2024, 243:109850.). However, the establishment of large-scale intelligent fault diagnosis models still faces many challenges: how to preprocess multi-entity data before building the large-scale fault diagnosis model; how to set the model structure during the large-scale model building process; and how to adapt the large-scale model to the target domain after its establishment. These challenges can be further summarized into three levels: data, model, and adaptation.

[0004] (1) Data Level: The training of large models requires the support of big data, especially in the construction of intelligent diagnostic large models. This process requires massive amounts of mechanical equipment condition monitoring data, which cover various working conditions and equipment types. However, when faced with such diverse mechanical equipment data, existing methods such as Max-Min normalization (YIN Peizhe, NIEJie, LIANG Xinyue, et al. A multi-scale graph convolutional neural network framework for fault diagnosis of rolling bearing[J]. IEEE Transactions on Instrumentation and Measurement, 2023.) can transform the data to the same amplitude range, but they have the disadvantage of completely discarding the amplitude difference information of the original signal. Therefore, how to ensure that the above data can be used for subsequent model training is an urgent problem to be solved.

[0005] (2) Model Level: Due to the differences in health status types among multi-entity datasets, there is overlap in the health statuses of different datasets. Different strategies exist when dealing with similar health status types across different datasets: one strategy treats similar health statuses from different entities as independent and classifies them as different types; another strategy considers the similar health statuses as having the same health status information and classifying them as the same type. Strategy one may lead to a decrease in model generalization ability, while strategy two has the disadvantages of information loss and reduced diagnostic accuracy. Therefore, how to label similar health statuses across entity data is particularly important during model construction.

[0006] (3) Adaptation Level: When adapting to the target entity, traditional intelligent diagnostic methods do not consider the differences in data distribution between the source domain and the target domain. In practical applications, models trained on the source domain data are directly used, resulting in poor generalization performance when there are distribution differences between the data. Existing technical solutions include adversarial domain adaptation methods (CHENZhuyun, HE Guolin, LI Jipu, et al. Domain adversarial transfer network for cross-domain fault diagnosis of rotary machinery[J]. IEEE Transactions on Instrumentation and Measurement, 2020, 69(11): 8702-8712.), but they have the disadvantage of potentially causing negative transfer when there are large domain differences. Therefore, how to generalize to the target domain is also a problem that needs to be considered. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, the present invention aims to propose a vibration-based intelligent fault diagnosis large-scale modeling method for rotating machinery. At the data level, it solves problems such as differences in data amplitude and information misalignment caused by cross-operating conditions, cross-equipment, and cross-sensor; at the model level, it solves the problem of multiple entity data and multiple health states; and at the adaptation level, it solves the problem of misclassification caused by domain differences, so that the model can better adapt to the target scenario.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A vibration-based intelligent fault diagnosis modeling method for rotating machinery includes: at the data level, a time-frequency structure alignment method for monitoring big data is proposed, which aligns the time-frequency information of the data; at the model level, a multi-scale state collaborative intelligent diagnostic network is proposed; and at the adaptation level, an adaptive fine-tuning adaptation method for the target entity model is proposed.

[0010] A vibration-based intelligent fault diagnosis large-scale modeling method for rotating machinery includes the following steps:

[0011] Step 1: For the known multi-entity sampling signal and the single-target entity sampling signal, define them as follows: and i = 1, 2, ..., N Dataset ,in X represents the acquired signal of the i-th known entity; i,j Y represents the j-th sampled signal of the i-th known entity; i,j To be with X i,j Corresponding health status; N represents the number of health states of the i-th known entity; Dataset X represents the total number of known entities; k Represents target entity D Target The sampled signal; N Target Represents the number of health states of the target entity; sampling signals from known multiple entities. and the single target entity sampling signal D Target Data preprocessing was performed to create samples, and the corresponding datasets are represented as follows: and in Given a dataset of known entities; y represents the m-th sample in the i-th known entity dataset; i,m For x i,m Corresponding health status; d represents the number of samples in a single known entity dataset; Target For the target entity dataset; y represents the nth sample in the target entity dataset; n For x n Corresponding health status; L represents the number of samples in the target entity dataset; Sample Indicates the length of a single sample;

[0012] Step 2: Normalization of cross-entity data: First, perform mean removal processing on each sampled signal; then, calculate the root mean square (RMS) index for the health data of different entities; finally, select the health signal of one entity from the known multi-entity sampled signals as the baseline data D. Base Using benchmark data D Base The ratio of the root mean square (RMS) of other entities is used to scale the magnitude of the data for other entities; the specific normalization formula is as follows:

[0013]

[0014] In the formula, L represents the length of signal X; X i,Healthy With X Base,Healthy μ represents the health signal of the i-th entity and the baseline data, respectively; i,Healthy With μ Base,Healthy Corresponding to X respectively i,Healthy and X Base,Healthy The mean of R; i,Base μ represents the magnitude scaling factor of the i-th entity. i,j X represents the j-th health state of the i-th entity. i,j The mean of j = 1, 2, ..., N i ;X i,j,Norm X representsi,j Normalized signal; N i R represents the number of health states of the i-th entity; Scale This is the scaling factor;

[0015] Step 3: Perform cross-entity data time-frequency alignment on multi-entity sampled signals: First, pre-set the reference sampling rate f. s,Base and reference frequency f n,Base Then, compare the actual sampling rate of each entity with the set baseline sampling rate f. s,Base To address the differences, linear interpolation resampling and low-pass filtering were performed on the data of different entities to achieve time information alignment; finally, the actual rotational speed of each entity was compared with the set reference rotational speed f. n,Base The scaling factor is calculated based on the differences, and the corresponding stretching and compression processing is performed on each entity data according to the scaling factor; the specific calculation formula is as follows:

[0016]

[0017] In the formula, Represents the actual rotational speed of entity i; I(·) represents the linear interpolation algorithm, based on the rotational speed ratio. Returns the signal after the specified signal speed is aligned. t i,j For the time axis of the signal before rotational speed alignment; The time axis is aligned with the rotational speed; finally, the known multi-entity sampling signals are obtained respectively. and the single target entity sampling signal D Target The preprocessed data, from which known entity datasets are obtained respectively. and target entity dataset d Target ;

[0018] Step 4: Set multi-scale state labels for the known entity dataset: The multi-scale state label consists of 3 sibling labels, namely, multi-entity state independent label, multi-entity state fusion label, and multi-entity state binary label.

[0019] Step 5: Construct a multi-scale state collaborative intelligent diagnostic network: The backbone of the intelligent diagnostic network structure uses the Transformer encoder. The overall structure of the multi-scale state collaborative intelligent diagnostic network model is divided into three parts: (1) data block embedding module; (2) feature extraction module based on multi-head attention mechanism; (3) multi-scale state collaborative classification module.

[0020] (1) The data block embedding module consists of convolution operations and positional encoding. The positional encoding vector is defined using sine and cosine functions, and its calculation formula is as follows:

[0021]

[0022]

[0023] In the formula, pos = 0, 1, ..., R represents the row index of the position code; R is the number of rows of the position code; 2p and 2p+1 are the even-numbered column and odd-numbered column indices of the position code, respectively; p = 0, 1, ..., d / 2; d represents the dimension of the position code vector; PE(pos, 2p) represents the value at the even-numbered index position in the position code vector, and PE(pos, 2p+1) represents the value at the odd-numbered index position.

[0024] (2) The feature extraction module based on the multi-head attention mechanism consists of multiple stacked sub-units. Each sub-unit can be regarded as being composed of two residual connections connected in series: 1) The first residual connection consists of two signals: the original input signal; the signal after the original input signal is standardized and then passes through the multi-head self-attention mechanism; 2) The second residual connection consists of two signals: the input signal; the signal after the input signal is standardized and then passes through the MLP; the specific structure of the MLP from top to bottom is a fully connected layer plus a GELU activation function and then a fully connected layer connected in series.

[0025] (3) The multi-scale state collaborative classification module consists of three classifiers of the same level, namely, the multi-entity state independent classifier, the multi-entity state fusion classifier, and the multi-entity state binary classifier.

[0026] Step 6: Pre-train the intelligent diagnostic network for multi-scale state collaboration: First, use the known entity dataset after monitoring the big data time-frequency structure alignment preprocessing in Steps 2 and 3. As model input; then, using the model output multi-scale state labels and the true multi-scale state labels, the pre-training loss L is calculated using cross-entropy loss. Pre_train Then, gradient descent is used to update the model parameter weights; finally, after reaching the maximum number of iterations, pre-training ends and the model parameter weights are saved; where the pre-training loss L Pre_train The calculation formula is as follows:

[0027]

[0028] In the formula, L1, L2 and L3 represent the loss values ​​of the multi-entity state independent classifier, the multi-entity state fusion classifier and the multi-entity state binary classifier, respectively. Let x represent the i-th input data respectively. i The neuron values ​​of the k1, k2, and k3 neurons in the fully connected layers fc1, fc2, and fc3 of the above three classifiers; y i For x iThe corresponding health status label; N represents the number of input samples; ε1, ε2, and ε3 are the weight values ​​of loss L1, L2, and L3, respectively;

[0029] The calculation formula for model update is as follows:

[0030]

[0031] In the formula, θ f,i These are the parameter weights of the i-th layer of model f; η is the learning rate.

[0032] Step 7: Adapt and fine-tune the multi-scale state collaborative intelligent diagnostic network: First, in the fine-tuning stage, the overall model structure remains consistent with the model in the pre-training stage, and the parameter weights saved in the pre-training stage are imported; then, based on the target entity dataset d... Target The health status information is used to adjust the multi-entity state independent classifier fc1 of the multi-scale state collaborative intelligent diagnostic network. The adjusted structure is represented as fc. 1,New Then, select the structure θ for model fine-tuning. f,Adapt Then, the parameter values ​​of the remaining modules are frozen to prevent incorrect parameter updates during fine-tuning; finally, the target entity dataset d is used. Target As input, the fine-tuning loss value L is calculated according to equation (7). Fine_tune And according to equation (8) θ f,Adapt The weights are updated to obtain a model that adapts to the target entity;

[0033]

[0034]

[0035] In the formula, L 1,New The loss of the adjusted multi-entity state independent classifier; For the i-th input data x i In the fully connected layer fc 1,New The neuron value of the k1th neuron.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] This invention proposes a vibration-based intelligent fault diagnosis modeling method for rotating machinery. Based on massive pre-training data, the method further fine-tunes the model using a small amount of target domain data, achieving rapid and efficient adaptation to the target domain. At the data level, the proposed method aligns the time-frequency structure of large monitoring data by narrowing the amplitude range of health status data and aligning the time-frequency structure, thus solving the problems of large amplitude differences and misaligned time-frequency data information encountered by existing methods when dealing with multi-entity data. At the model level, addressing the problem of joint modeling of multi-entity data, this invention proposes a multi-scale state collaborative intelligent diagnostic network, thus solving the complexity of multi-entity and multi-health status modeling at the model level. At the adaptation level, this invention proposes an adaptive fine-tuning method for the target entity model. By pre-training the model with multi-entity data and then fine-tuning it using a small amount of target entity data, this method addresses the difficulty of generalizing models built on a single entity to other devices. Attached Figure Description

[0038] Figure 1 This is a flowchart of an embodiment of the present invention.

[0039] Figure 2 This is a schematic diagram of a multi-scale state-coordinated intelligent diagnostic network according to an embodiment of the present invention.

[0040] Figure 3 The figures show the comparison results of different input and output modes in the embodiments of the present invention; where (a) is the experimental result (1 sample per class) and (b) is the experimental result (20 samples per class). Detailed Implementation

[0041] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings.

[0042] like Figure 1 As shown, a large-scale modeling method for intelligent fault diagnosis of rotating machinery based on vibration includes the following steps:

[0043] Step 1: For known multi-entity sampling signals and single-target entity sampling signals, they can be defined as follows: and i = 1, 2, ..., N Dataset ,in X represents the acquired signal of the i-th known entity; i,j Y represents the j-th sampled signal of the i-th known entity; i,j To be with X i,j Corresponding health status; N represents the number of health states of the i-th known entity; Dataset X represents the total number of known entities; k Represents target entity DTarget The sampled signal; N Target Represents the number of health states of the target entity; sampling signals from known multiple entities. and the single target entity sampling signal D Target Data preprocessing is performed to create samples, and the corresponding datasets can be represented as follows: and in Given a dataset of known entities; y represents the m-th sample in the i-th known entity dataset; i,m For x i,m Corresponding health status; d represents the number of samples in a single known entity dataset; Target For the target entity dataset; y represents the nth sample in the target entity dataset; n For x n Corresponding health status; L represents the number of samples in the target entity dataset; Sample It represents the length of a single sample; by constructing massive amounts of known entity data with varying distributions, the learning and generalization capabilities of the multi-scale state-coordinated intelligent diagnostic network can be enhanced during the pre-training stage.

[0044] Step 2: Normalization of cross-entity data: First, perform mean removal processing on each sampled signal; then, calculate the root mean square (RMS) index for the health data of different entities; finally, select the health signal of one entity from the known multi-entity sampled signals as the baseline data D. Base Using benchmark data D Base The ratio of the root mean square (RMS) of other entities is used to scale the magnitude of the data for other entities; the specific normalization formula is as follows:

[0045]

[0046] In the formula, L represents the length of signal X; X i,Healthy With X Base,Healthy μ represents the health signal of the i-th entity and the baseline data, respectively; i,Healthy With μ Base,Healthy Corresponding to X respectively i,Healthy and X Base,Healthy The mean of R; i,Base μ represents the magnitude scaling factor of the i-th entity. i,j X represents the j-th health state of the i-th entity. i,j The mean of j = 1, 2, ..., N i ;X i,j,Norm X represents i,j Normalized signal; Ni R represents the number of health states of the i-th entity; Scale This is the scaling factor;

[0047] Using a cross-entity data time-frequency alignment method alleviates the problems of model overfitting and unit inconsistency caused by large differences in signal amplitude across cross-entity datasets. On the other hand, it does not interfere with the health and fault data of the same entity, thus preserving the amplitude difference information between the original health and fault data signals in a single entity. In addition, a scaling factor is introduced to avoid the phenomenon that the signal amplitude is too low after standardization, which makes it difficult for the model to fit.

[0048] Step 3: Perform cross-entity data time-frequency alignment on multi-entity sampled signals: First, pre-set the reference sampling rate f. s,Base and reference frequency f n,Base Then, compare the actual sampling rate of each entity with the set baseline sampling rate f. s,Base To align time information, linear interpolation resampling and low-pass filtering are performed on the data of different entities to address the differences. Finally, the actual rotational speed of each entity is compared with the set reference rotational speed f. n,Base The scaling factor is calculated based on the differences, and the corresponding stretching and compression processing is performed on each entity data according to the scaling factor; the specific calculation formula is as follows:

[0049]

[0050] In the formula, Represents the actual rotational speed of entity i; I(·) represents the linear interpolation algorithm, based on the rotational speed ratio. Returns the signal after the specified signal speed is aligned. t i,j For the time axis of the signal before rotational speed alignment; The time axis is aligned with the rotational speed; finally, the known multi-entity sampling signals are obtained respectively. and the single target entity sampling signal D Target The preprocessed data, from which known entity datasets are obtained respectively. and target entity dataset d Target ;

[0051] By performing linear interpolation resampling and low-pass filtering on different entity data, time information alignment is achieved; by stretching and compressing the data of each entity according to the difference between the actual rotation speed and the set reference rotation speed, the information inconsistency caused by rotation speed is resolved, and rotation speed information alignment is achieved; finally, the problem of inconsistency in time-frequency structure information is resolved.

[0052] Step 4: Set multi-scale state labels for the known entity dataset: The multi-scale state labels mainly consist of three sibling labels, namely, multi-entity state independent label, multi-entity state fusion label, and multi-entity state binary label.

[0053] By setting multi-level state labels for the dataset, the requirements for subsequent model structure are met, and the goal of multi-angle feature extraction is achieved.

[0054] Step 5: Construct a multi-scale state-coordinated intelligent diagnostic network: Refer to... Figure 2 The backbone of the intelligent diagnostic network structure uses the Transformer encoder. The overall structure of the multi-scale state collaborative intelligent diagnostic network model is divided into three parts: (1) data block embedding module; (2) feature extraction module based on multi-head attention mechanism; and (3) multi-scale state collaborative classification module.

[0055] (1) The data block embedding module consists of convolution operations and positional encoding. The positional encoding vector is defined using sine and cosine functions, and its calculation formula is as follows:

[0056]

[0057]

[0058] In the formula, pos = 0, 1, ..., R represents the row index of the position code; R is the number of rows of the position code; 2p and 2p+1 are the even-numbered column and odd-numbered column indices of the position code, respectively; p = 0, 1, ..., d / 2; d represents the dimension of the position code vector; PE(pos, 2p) represents the value at the even-numbered index position in the position code vector, and PE(pos, 2p+1) represents the value at the odd-numbered index position.

[0059] (2) The feature extraction module based on the multi-head attention mechanism is mainly composed of multiple stacked sub-units. A single sub-unit can be regarded as being composed of two residual connections connected in series: 1) The first residual connection consists of two signals: the original input signal; the signal after the original input signal is standardized and then passes through the multi-head self-attention mechanism; 2) The second residual connection consists of two signals: the input signal; the signal after the input signal is standardized and then passes through the MLP; the specific structure of the MLP from top to bottom is a fully connected layer (number of neurons: 256) plus the GELU activation function and then connected in series with another fully connected layer (number of neurons: 256);

[0060] (3) The multi-scale state collaborative classification module is mainly composed of three classifiers of the same level, namely, the multi-entity state independent classifier, the multi-entity state fusion classifier, and the multi-entity state binary classifier.

[0061] By constructing a multi-scale state-coordinated intelligent diagnostic network, we can improve the generalization performance of the final model in the target domain by extracting features from single entity data and paying attention to the potential shared features between cross-entity data.

[0062] Step 6: Pre-train the intelligent diagnostic network for multi-scale state collaboration: First, use the known entity dataset after monitoring the big data time-frequency structure alignment preprocessing in Steps 2 and 3. As model input; then, using the model output multi-scale state labels and the true multi-scale state labels, the pre-training loss L is calculated using cross-entropy loss. Pre_train Then, gradient descent is used to update the model parameter weights; finally, after reaching the maximum number of iterations, pre-training ends and the model parameter weights are saved; where the pre-training loss L Pre_train The calculation formula is as follows:

[0063]

[0064] In the formula, L1, L2 and L3 represent the loss values ​​of the multi-entity state independent classifier, the multi-entity state fusion classifier and the multi-entity state binary classifier, respectively. Let x represent the i-th input data respectively. i The neuron values ​​of the k1, k2, and k3 neurons in the fully connected layers fc1, fc2, and fc3 of the above three classifiers; y i For x i The corresponding health status label; N represents the number of input samples; ε1, ε2, and ε3 are the weight values ​​of loss L1, L2, and L3, respectively;

[0065] The calculation formula for model update is as follows:

[0066]

[0067] In the formula, θ f,i These are the parameter weights of the i-th layer of model f; η is the learning rate.

[0068] By learning from massive amounts of data, the model gains the ability to extract features across entities.

[0069] Step 7: Adapt and fine-tune the multi-scale state collaborative intelligent diagnostic network: First, in the fine-tuning stage, the overall model structure remains consistent with the model in the pre-training stage, and the parameter weights saved in the pre-training stage are imported; then, based on the target entity dataset d... Target The health status information is used to adjust the multi-entity state independent classifier fc1 of the multi-scale state collaborative intelligent diagnostic network. The adjusted structure is represented as fc. 1,New Then, select the structure θ for model fine-tuning. f,AdaptThen, the parameter values ​​of the remaining modules are frozen to prevent incorrect parameter updates during fine-tuning; finally, the target entity dataset d is used. Target As input, the fine-tuning loss value L is calculated according to equation (7). Fine_tune And according to equation (8) θ f,Adapt The weights are updated to obtain a model that adapts to the target entity;

[0070]

[0071]

[0072] In the formula, L 1,New The loss of the adjusted multi-entity state independent classifier; For the i-th input data x i In the fully connected layer fc 1,New The neuron value of the k1th neuron;

[0073] This step allows for fine-tuning of the pre-trained model using only a small amount of target entity data, based on the pre-trained model, to achieve faster and more accurate application to target entity scenarios.

[0074] The embodiments use key components of rotating equipment such as bearings, gears, and motors in mechanical equipment as examples, and verify the effectiveness of the present invention based on publicly available datasets and unpublished laboratory datasets.

[0075] The dataset includes multi-condition and multi-sensor data from various types of components such as bearings, gears, and motors. Partial information about the dataset is shown in the table below. Some parameters and their values ​​involved in the experiment are also shown in the table below. Among them, N... Batch The sample size for each batch; N Epoch N represents the total number of training epochs. Patch Size of each block.

[0076] Table 1 Experimental Dataset and Related Information

[0077]

[0078]

[0079] Table 2. Some parameters used in the experiment and their corresponding values.

[0080]

[0081] Comparison methods: Directly train and test the model; fine-tune the structure: only fine-tune the classification module of the model; fine-tune the overall structure and fine-tune the individual model.

[0082] Experimental results are as follows Figure 3As shown, by Figure 3 As can be seen in (a), with an extremely small sample size (1 per class), directly training the model easily leads to overfitting. The proposed method shows a significant improvement in diagnostic accuracy compared to the comparison method. Figure 3 As shown in Figure (b), when the sample size is sufficient, the proposed method also improves the diagnostic accuracy compared to the comparison method. Experimental results show that the model constructed in the pre-training stage can extract general health status feature information and can respond to the target device in the fine-tuning stage, ultimately achieving a good adaptation effect.

Claims

1. A large-scale modeling method for intelligent fault diagnosis of rotating machinery based on vibration, characterized in that, include: At the data level, a time-frequency structure alignment method for monitoring big data is proposed. This method aligns the data with time-frequency information, including known entity datasets. Multi-scale state labels are then defined for these datasets: each multi-scale state label consists of three sibling labels: independent multi-entity state label, fused multi-entity state label, and binary multi-entity state label. At the model level, a multi-scale state collaborative intelligent diagnostic network is proposed. The backbone of this network uses a Transformer-based encoder. The overall structure of the multi-scale state collaborative intelligent diagnostic network model consists of three parts: a data block embedding module; a feature extraction module based on a multi-head attention mechanism; and a multi-head attention-based feature extraction module. Multi-scale state collaborative classification module; The data includes known multi-entity sampled signals and single-target entity sampled signals. Cross-entity data time-frequency alignment is performed on the multi-entity sampled signals: First, a reference sampling rate is pre-set. and reference frequency ; Then, compare the actual sampling rate of each entity with the set baseline sampling rate. To address the differences, linear interpolation resampling and low-pass filtering were performed on the data of different entities to achieve time information alignment; finally, the actual rotational speed of each entity was compared with the set reference rotational speed. The difference is used to calculate the scaling factor, and the corresponding stretching and compression processing is performed on each entity data according to the scaling factor; The data block embedding module consists of convolution operations and positional encoding. The positional encoding vector is defined using sine and cosine functions, and its calculation formula is as follows: (3) (4) In the formula, R represents the row index of the position code; R is the number of rows in the position code. and These are the indexes of the even-numbered and odd-numbered columns, respectively, for position encoding. ; Indicates the dimension of the position encoding vector; This represents the value at an even index in the positional encoding vector. Indicates the value at the odd index position; The feature extraction module based on the multi-head attention mechanism consists of multiple stacked sub-units. Each sub-unit can be considered as two residual connections connected in series: 1) The first residual connection consists of two signals: the original input signal and the signal after the original input signal is standardized and then passes through the multi-head self-attention mechanism; 2) The second residual connection consists of two signals: the input signal and the signal after the input signal is standardized and passes through the MLP; The specific structure of the MLP, from top to bottom, is a fully connected layer plus a GELU activation function and then another fully connected layer. The multi-scale state collaborative classification module consists of three peer classifiers, namely, a multi-entity state independent classifier, a multi-entity state fusion classifier, and a multi-entity state binary classifier. At the adaptation level, an adaptive fine-tuning method for the target entity model is proposed. The target entity model is obtained by pre-training a multi-scale state-coordinated intelligent diagnostic network. First, the known entity dataset after preprocessing by aligning the time-frequency structure of the monitoring big data in steps 2 and 3 is used. As input to the model; then, the pre-training loss is calculated using cross-entropy loss with the model's output multi-scale state labels and the true multi-scale state labels. Then, gradient descent is used to update the model parameter weights; finally, the pre-training ends and the model parameter weights are saved after the maximum number of iterations is reached.

2. A large-scale modeling method for intelligent fault diagnosis of rotating machinery based on vibration, characterized in that, Includes the following steps: Step 1: For the known multi-entity sampling signal and the single-target entity sampling signal, define them as follows: and , ,in Indicates the first The acquired signal of a known entity; Indicates the first The first known entity A state of health; To and Corresponding health status; Indicates the first Number of health statuses of known entities; Indicates the total number of known entities; Represents the target entity The sampled signal; Represents the number of health states of the target entity; sampling signals from known multiple entities. and single target entity sampling signal Data preprocessing was performed to create samples, and the corresponding datasets are represented as follows: and ,in Given a dataset of known entities; Indicates the first The first known entity dataset One sample; To and Corresponding health status; This represents the number of samples in a single known entity dataset. For the target entity dataset; Represents the first element of the target entity dataset. One sample; To and Corresponding health status; Indicates the number of samples in the target entity dataset; Indicates the length of a single sample; Step 2: Normalization of cross-entity data: First, perform mean removal processing on each sampled signal; then, calculate the root mean square index for the health data of different entities; finally, select the health signal of one entity from the known multi-entity sampled signals as the baseline data. Using benchmark data The ratio of the root mean square (RMS) of other entities is used to scale the magnitude of the data for other entities; the specific normalization formula is as follows: (1) In the formula, Indicates signal Length; and They represent the first Health signals of individual entities and benchmark data; and Corresponding to and The mean; Indicates the first The magnitude scaling factor for each entity; Indicates the first The entity health status The mean, ; express The normalized signal; Indicates the first The number of health statuses of each entity; This is the scaling factor; Step 3: Perform cross-entity data time-frequency alignment on multi-entity sampled signals: First, pre-set the reference sampling rate. and reference frequency ; Then, compare the actual sampling rate of each entity with the set baseline sampling rate. To address the differences, linear interpolation resampling and low-pass filtering were performed on the data of different entities to achieve time information alignment; finally, the actual rotational speed of each entity was compared with the set reference rotational speed. The scaling factor is calculated based on the differences, and the corresponding stretching and compression processing is performed on each entity data according to the scaling factor; the specific calculation formula is as follows: (2) In the formula, Representing entities The actual rotational speed; This represents a linear interpolation algorithm, based on the rotational speed ratio. Returns the signal after the specified signal speed is aligned. ; For the time axis of the signal before rotational speed alignment; The time axis is aligned with the rotational speed; finally, the known multi-entity sampling signals are obtained respectively. and single target entity sampling signal The preprocessed data, from which known entity datasets are obtained respectively. and target entity dataset ; Step 4: Set multi-scale state labels for the known entity dataset: The multi-scale state label consists of 3 sibling labels, namely, multi-entity state independent label, multi-entity state fusion label, and multi-entity state binary label. Step 5: Construct a multi-scale state collaborative intelligent diagnostic network: The backbone of the intelligent diagnostic network structure uses the Transformer encoder part. The overall structure of the multi-scale state collaborative intelligent diagnostic network model is divided into three parts: (1) data block embedding module; (2) feature extraction module based on multi-head attention mechanism; (3) multi-scale state collaborative classification module. (1) The data block embedding module consists of convolution operations and positional encoding. The positional encoding vector is defined using sine and cosine functions, and its calculation formula is as follows: (3) (4) In the formula, R represents the row index of the position code; R is the number of rows in the position code. and These are the indexes of the even-numbered and odd-numbered columns, respectively, for position encoding. ; Indicates the dimension of the position encoding vector; This represents the value at an even index in the positional encoding vector. Indicates the value at the odd index position; (2) The feature extraction module based on the multi-head attention mechanism consists of multiple stacked sub-units. Each sub-unit can be regarded as being composed of two residual connections connected in series: 1) The first residual connection consists of two signals: the original input signal; the signal after the original input signal is standardized and then passes through the multi-head self-attention mechanism; 2) The second residual connection consists of two signals: the input signal; the signal after the input signal is standardized and then passes through the MLP; the specific structure of the MLP from top to bottom is a fully connected layer plus a GELU activation function and then a fully connected layer connected in series. (3) The multi-scale state collaborative classification module consists of three classifiers of the same level, namely, multi-entity state independent classifier, multi-entity state fusion classifier, and multi-entity state binary classifier. Step 6: Pre-train the intelligent diagnostic network for multi-scale state collaboration: First, use the known entity dataset after monitoring the big data time-frequency structure alignment preprocessing in Steps 2 and 3. As input to the model; then, the pre-training loss is calculated using cross-entropy loss with the model's output multi-scale state labels and the true multi-scale state labels. Then, gradient descent is used to update the model parameter weights; finally, after reaching the maximum number of iterations, pre-training ends and the model parameter weights are saved; where, the pre-training loss... The calculation formula is as follows: (5) In the formula, , and These represent the loss values ​​of the multi-entity state independent classifier, the multi-entity state fusion classifier, and the multi-entity state binary classifier, respectively. , , They represent the first Input data The above three classifiers are in the fully connected layer. , and The , and Neuron values ​​of one neuron; To and Corresponding health status labels; Indicates the number of input samples; , , Losses , and The weight value; The calculation formula for model update is as follows: (6) In the formula, It is a model No. Layer parameter weight values; The learning rate; Step 7: Adapt and fine-tune the multi-scale state collaborative intelligent diagnostic network: First, in the fine-tuning stage, the overall model structure remains consistent with the model in the pre-training stage, and the parameter weights saved in the pre-training stage are imported; then, based on the target entity dataset... The health status information is used by a multi-entity state independent classifier in a multi-scale state collaborative intelligent diagnostic network. Make the corresponding adjustments, and the adjusted structure is represented as follows: Then, select the structure for model fine-tuning. Then, the parameter values ​​of the remaining modules are frozen to prevent incorrect parameter updates during fine-tuning; finally, the target entity dataset is used. As input, the fine-tuning loss value is calculated according to equation (7). And according to equation (8) The weights are updated to obtain a model that adapts to the target entity; (7) (8) In the formula, The loss of the adjusted multi-entity state independent classifier; For the first Input data In the fully connected layer The Neuron value of each neuron.

Images