Multi-condition semantic-guided bearing fault sample diffusion generation method under missing working condition combination

Abstract

The application discloses a bearing fault sample diffusion generation method under multi-condition semantic guidance of variable working condition combination loss, and belongs to the technical field of mechanical equipment health monitoring and intelligent fault diagnosis. Firstly, a fault type, a fault size and a rotating speed combination loss scene is constructed to form a structured condition vector, time-frequency domain features are extracted, and task prompt words are organized. Then, GPT-2 is used as a text semantic encoder, parameters are fine-tuned by using LoRA, and the prompt words are mapped into continuous semantic vectors. Finally, the structured conditions and the semantic conditions are fused and input into a condition denoising diffusion model to guide the model to generate bearing vibration samples under the loss of the fault size-rotating speed combination. The application can alleviate the diagnostic information vacancy caused by the loss of the variable working condition combination, improve the matching degree between the generated samples and the target fault state and the target working condition, enhance the condition expression ability of the bearing fault generation task while reducing the fine-tuning overhead of the semantic encoder, and has good generation distribution stability and diagnosis enhancement effect.

Description

Technical Field

[0001] This invention belongs to the field of mechanical equipment health monitoring and intelligent fault diagnosis technology, and relates to a bearing fault sample diffusion generation method under the condition of missing variable working condition combinations. Specifically, it relates to a bearing fault sample generation and diagnosis enhancement method for scenarios with missing variable working condition combinations; including bearing vibration signal processing, time-frequency domain feature extraction, semantic prompt word construction, GPT-2 semantic encoding, efficient fine-tuning of LoRA parameters, conditional diffusion model, fault sample generation, and bearing fault diagnosis enhancement under strong noise conditions. Background Technology

[0002] Rolling bearings are critical support components in rotating machinery, and their operating condition directly affects the safety and stability of the equipment. During long-term operation, bearings are susceptible to alternating loads, speed fluctuations, impact vibrations, and environmental noise, leading to localized faults in the inner ring, outer ring, or rolling elements. If these faults are not identified in a timely manner, they can result in increased equipment vibration, decreased transmission accuracy, and even shutdown failure.

[0003] Existing bearing fault diagnosis methods mostly employ data-driven models, such as convolutional neural networks, recurrent neural networks, Transformers, and their improved structures. These methods can automatically learn fault features from vibration signals and achieve high diagnostic accuracy when there are sufficient samples and complete operating condition coverage. However, in practical engineering, bearing fault samples often suffer from insufficient quantity, imbalanced categories, and missing operating condition combinations. For example, the training set may lack samples for a specific fault size and speed combination, making it difficult for the diagnostic model to learn the complete fault distribution and reducing its generalization ability under unseen combinations.

[0004] To address the problem of insufficient samples, existing technologies often employ transfer learning, domain adaptation, few-shot learning, or data augmentation methods. Transfer learning and domain adaptation methods typically require a small number of samples in the target operating condition, making it difficult to directly solve the sample completion problem under completely missing combinations. Generative methods such as VAEs and GANs have also been used for fault sample augmentation, but VAE-generated samples tend to be overly smoothed, making it difficult to retain the impact characteristics in the vibration signal; GANs are prone to training instability, pattern collapse, and fluctuations in generation quality under conditions of small samples and strong noise.

[0005] Conditional diffusion models learn data distribution through a process of progressive noise addition and reverse denoising, offering advantages such as stable training, high-quality generation, and strong conditional control capabilities. They are increasingly being used for time series generation and fault sample augmentation tasks. However, existing conditional diffusion models often use simple category labels, speed labels, or load labels as conditional inputs, making it difficult to fully express the complex relationships between fault type, fault size, speed, and signal time-frequency characteristics.

[0006] In recent years, large language models have provided new approaches to expressing complex conditions. By organizing fault type, fault size, rotational speed, and time-frequency domain statistical features into textual prompts and using language models for semantic encoding, richer conditional information can be provided to generative models. However, how to effectively integrate structural chemical condition conditions with semantic conditions and guide diffusion models to generate high-quality bearing vibration samples under missing combinations remains a problem that needs to be solved in existing technologies. Summary of the Invention

[0007] This invention addresses the problems of insufficient bearing fault samples, missing combined operating conditions, unstable sample quality, and insufficient generalization ability of diagnostic models under varying operating conditions and high noise. It proposes a multi-condition semantic-guided bearing fault sample diffusion generation method under missing combined operating conditions. Using the CWRU bearing dataset as the research object, this invention constructs scenarios with missing combinations of fault type, fault size, and rotational speed. The fault type, fault size, and rotational speed are used to form a structured conditional vector. Simultaneously, time-frequency domain features such as mean, standard deviation, peak value, kurtosis, root mean square value, power spectral density, main peak frequency, root mean square frequency, and center frequency are extracted and organized into task prompt words. Then, GPT-2 is used as a text semantic encoder, and LoRA is used for efficient parameter fine-tuning to map the prompt words into continuous semantic vectors. Finally, the structured and semantic conditions are fused and input into a denoising diffusion probabilistic model (DDPM) to guide the model in generating bearing vibration samples under missing fault size-rotational speed combinations, which are then used for diagnostic model enhancement. The main contributions of this invention include: proposing a conditional diffusion generation framework that integrates structured conditions with GPT-2 semantic conditions; introducing time-frequency domain feature prompts and LoRA fine-tuning to enhance semantic conditional expressions; constructing a diagnostic scenario for missing fault size-speed combination; and verifying the effectiveness of the method through generation quality evaluation under multiple noise conditions, diagnostic enhancement experiments, ablation experiments, and comparative experiments.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for generating bearing fault samples with multi-condition semantic guidance under missing variable operating condition combinations includes the following steps: Step 1: Collect and preprocess continuous bearing vibration signals labeled with fault type, fault size, and rotational speed. Divide the samples into bearing vibration samples using a fixed window. Based on the fault type, fault size, and rotational speed labels, classify the samples into existing and missing combinations. Construct training data for the generation model, training data for the diagnostic model, and independent test data to obtain the training sample set for existing combinations, the target condition set for missing combinations, and the test sample set for all combinations. Specifically: Step 1.1: Obtain the continuous vibration signal of the bearing and its corresponding fault type, fault size and speed label; wherein, the fault type is used to determine the sub-fault type generation model to be trained or called, the fault size is used to describe the fault degree, and the speed is used to describe the operating condition.

[0009] Step 1.2: The continuous bearing vibration signal obtained in Step 1.1 is segmented according to the preset window length and sliding step size, so that each sample is a one-dimensional vibration signal of fixed length, and each sample inherits the corresponding fault type, fault size and speed label to obtain a fixed length sample.

[0010] Step 1.3: Based on the fixed-length samples obtained in Step 1.2, classify them according to "fault type-fault size-speed". Under each fault type, define the fault size-speed combination that participates in the training of the generative model as the existing combination, and define the fault size-speed combination that does not participate in the training of the generative model as the missing combination, so as to obtain the set of existing combinations and the set of target conditions for missing combinations.

[0011] Step 1.4: Based on the existing combination set obtained in Step 1.3, construct the training sample set for the generative model and the training sample set for the benchmark diagnostic model, and separately retain real samples covering all fault size-speed combinations as an independent test sample set; the missing combination target condition set does not participate in the training of the generative model, and is only used for quality evaluation in experimental verification scenarios with corresponding real samples.

[0012] Step 2: Based on the existing combined training sample set and the missing combined target condition set obtained in Step 1, the fault size and rotational speed are normalized and encoded. Conditional denoising and diffusion probability models corresponding to inner race faults, rolling element faults, and outer race faults are constructed and trained respectively, resulting in structured numerical condition vectors. And the fault type model selection condition g; specifically: Step 2.1: Denote the fault type of each sample as g, the fault size as d, and the rotational speed as v; use the existing combination set of the same fault type to train the conditional DDPM corresponding to g, and use the same g to select the corresponding conditional DDPM when generating missing combination samples.

[0013] Step 2.2: Normalize the fault size d and the rotational speed v according to their respective value ranges to obtain the normalized fault size d. norm and normalized rotational speed v norm And construct structured numerical condition vectors : Step 2.3, based on the fault type model selection condition g determined in Step 2.1 and the structured numerical condition vector c obtained in Step 2.2.num This yields explicit operating conditions for subsequent semantic condition fusion; where g is used to constrain the fault category. Used to constrain the fault level and rotational speed conditions of the target sample.

[0014] Step 3: Based on the existing combined training samples obtained in Step 1, extract time-domain and frequency-domain statistical features, and organize these statistical features with the target fault type, target fault size, and target rotational speed into task cue words p, obtaining the target condition description used for semantic encoding; specifically: Step 3.1: Extract time-domain statistical features from the existing combined training samples. The time-domain statistical features include mean, standard deviation, peak value, kurtosis, root mean square value, and kurtosis factor.

[0015] Step 3.2: Based on the existing combined training samples corresponding to Step 3.1, further extract frequency domain statistical features, including power spectral density, main peak frequency, root mean square frequency, and center frequency.

[0016] Step 3.3: Organize the target fault type, target fault size, target rotation speed, and the time-domain and frequency-domain statistical features obtained in steps 3.1 and 3.2 into English task prompt words p according to a fixed field order; For missing combination samples, the target time-domain and frequency-domain statistical features are determined by the statistical features of existing combination samples of the same fault type, specifically by using the statistical values ​​of existing combination features or estimating based on existing combination features, without using the missing combination real samples to participate in the generative model training.

[0017] For each sample to be generated, an English task cue word p is first constructed. This cue word p includes the fault type, fault size, rotational speed, and the time-domain and frequency-domain statistical characteristics of the target signal, enabling the generative model to obtain more fine-grained signal distribution constraints.

[0018] Step 4: Based on the task prompt word p obtained in Step 3, introduce LoRA low-rank adaptation parameters to the GPT-2 semantic encoder and perform efficient parameter fine-tuning. Then, perform word segmentation, semantic encoding, average pooling, and linear projection on the task prompt word to obtain a fixed-dimensional text semantic conditional vector. Specifically: Step 4.1, let p be the English task prompt word corresponding to the sample to be generated, and use a token segmenter to convert it into a token sequence of length L: Where p represents the task prompt word input to the GPT-2 semantic encoder; z represents the token sequence obtained after the task prompt word is segmented. L represents the i-th token; L represents the length of the token sequence.

[0019] Step 4.2: Input the token sequence into the LoRA-tuned GPT-2 semantic encoder to obtain the last hidden state matrix: Where H represents the hidden state matrix output by the GPT-2 semantic encoder; This represents the hidden state vector corresponding to the i-th token; Indicates the dimension of the GPT-2 hidden layer; To hide the state space, This represents the number of Transformer layers.

[0020] Step 4.3: To obtain a fixed-length text semantic condition vector, the hidden states of all token sequences are averaged and mapped to the DDPM condition space via linear projection. in, This represents the text semantic condition vector extracted by the GPT-2 semantic encoder; and These represent the weight matrix and bias term of the linear projection layer, respectively; Indicates the first i The hidden state vector corresponding to each token; This indicates the index number of the token in the sequence.

[0021] Textual semantic condition vectors are used to express the time-domain and frequency-domain features of the target signal contained in the task prompts, enabling the generative model to obtain more fine-grained signal distribution constraints.

[0022] Step 4.4: To reduce the number of parameters in GPT-2 fine-tuning, this invention employs the LoRA method for efficient parameter fine-tuning of the GPT-2 semantic encoder. LoRA does not directly update the original weights of the GPT-2 semantic encoder, but instead introduces low-rank trainable increments in the attention module. For the original weight matrix... The equivalent weights after LoRA fine-tuning are expressed as: in, This represents the original frozen weights of the GPT-2 semantic encoder; This represents the low-rank weight increment introduced by LoRA; and For trainable low-rank matrices, For the first low-rank matrix, It is the second low-rank matrix; Indicates the LoRA rank; This represents the scaling factor.

[0023] During training, only the LoRA adaptation parameters and subsequent conditional projection layer parameters are updated, thereby obtaining semantic encoding capabilities for bearing fault generation tasks with low computational overhead.

[0024] Step 5, based on the structured numerical condition vector c obtained in Step 2 num The fault type model selection condition g and the text semantic condition vector c obtained in step 4 text A complete multi-conditional guidance vector c is constructed, and the conditional DDPM corresponding to g is trained with forward noise addition and denoising with backward guidance to generate bearing vibration samples with target missing fault size-speed combination; specifically: The structured conditions consist of fault size and rotational speed, and are expressed as follows: in, Represents a structured numerical condition vector; This represents the normalized fault size. This represents the normalized rotational speed.

[0025] Combining structured numerical condition vectors with textual semantic condition vectors By concatenating the vectors, we obtain the complete multi-condition guided vector of the input condition DDPM: in, This represents the final condition guidance vector. This final condition guidance vector simultaneously contains fault type, fault severity, speed and operating conditions, fault characteristic frequencies, and time-frequency domain statistical feature semantic information.

[0026] The inverse denoising network of the final conditional DDPM is represented as follows: in, This represents the noisy vibration signal at step t; t represents the diffusion time step. Represents a multi-condition guiding vector; This represents the noise in the conditional DDPM prediction.

[0027] In this way, the Conditional Denoising Diffusion Model (DDPM) is not only subject to explicit conditional constraints such as fault type, fault size, and rotational speed during the reverse denoising process, but also to time-frequency domain semantic information encoded by the GPT-2 semantic encoder, thereby enabling the generation of controllable fault samples for missing fault size-rotational speed combinations.

[0028] Step 6: Based on the existing combined training sample set obtained in Step 1 and the bearing vibration generation sample of the target missing fault size-speed combination obtained in Step 5, construct an enhanced training set, train the fault diagnosis model, and perform diagnostic verification on the full combination test sample set obtained in Step 1. This yields a missing combination bearing fault generation sample set and an enhanced fault diagnosis model with verified diagnostic effectiveness, thus completing the multi-condition semantic-guided bearing fault sample diffusion generation under varying operating condition combinations. Specifically: Step 6.1: Use the existing combined training sample set obtained in Step 1 to train the baseline fault diagnosis model and obtain the baseline diagnosis results.

[0029] Step 6.2: Add the bearing vibration generation sample with the target missing fault size-speed combination obtained in Step 5 to the existing combination real diagnostic training sample used in Step 6.1 to construct an enhanced training set, and train the enhanced fault diagnosis model under the same diagnostic network structure and training settings.

[0030] Step 6.3: Use the full combination test sample set obtained in Step 1 to test the Baseline diagnostic model and the enhanced fault diagnosis model respectively. Evaluate the supplementary role of the generated samples in missing combination diagnostic information by the diagnostic accuracy. When there are real samples with missing combinations in the experimental scenario, the generation quality can be further evaluated by the temporal similarity and distribution difference between the generated samples and the real samples.

[0031] The beneficial effects of this invention are as follows: (1) In step 1, the present invention explicitly constructs an existing combination training sample set and a missing combination target condition set, and in step 5, it generates missing combination samples after training the condition denoising diffusion model DDPM using only the existing combination training sample set, so that the training data can cover the fault size-speed combination that did not appear in the original training set, thereby alleviating the gap in diagnostic information caused by the missing variable working condition combination from the data level.

[0032] (2) The present invention combines fault type conditions, structured numerical conditions consisting of fault size and rotation speed, and text semantic conditions consisting of time-frequency domain statistical features for diffusion generation. Among them, the fault type is constrained by fault type model and task prompt words, the structured numerical conditions provide clear working condition constraints, and the text semantic conditions provide fine-grained distribution constraints. Therefore, it can improve the matching degree between the generated sample and the target fault state and the target working condition.

[0033] (3) The present invention introduces LoRA low-rank adaptation parameters in the GPT-2 semantic encoder. Only the low-rank adaptation parameters and the conditional projection layer parameters are updated, which can map the fault type, fault size, speed and time-frequency domain statistical features in the task prompt words into fixed-dimensional semantic condition vectors. This reduces the fine-tuning overhead of the semantic encoder while enhancing the conditional expression capability for bearing fault generation tasks.

[0034] (4) The present invention adopts the forward noise addition and reverse noise reduction mechanism of the conditional denoising diffusion model DDPM, and corrects the noise prediction results by conditional discarding in the training stage and conditional / unconditional joint guidance in the generation stage, so that the generation process still has stable distribution learning ability under small sample and strong noise conditions; combined with the comparative experiment of this embodiment, the generation distribution stability and diagnostic enhancement effect of the method under strong noise conditions are better than GAN and VAE baselines. Attached Figure Description

[0035] Figure 1 Flowchart of a bearing fault sample diffusion generation method guided by multi-condition semantics under missing variable operating condition combinations; Figure 2 The training loss curves for DDPM under fault type conditions are IR, B, and OR at 5dB; where IR, B, and OR are inner race fault, rolling element fault, and outer race fault, respectively. Figure 3 The training loss curves for DDPM under fault type conditions (IR, B, and OR) under 0dB conditions; Figure 4 The training loss curves for DDPM under fault type conditions (IR, B, and OR) under -2dB conditions; Figure 5 The training loss curves for DDPM under fault type conditions (IR, B, and OR) under -4dB conditions are shown.

[0036] Figure 6 This is a graph showing the trend of Euclidean similarity as a function of signal-to-noise ratio. Figure 7 This is a graph showing the trend of the maximum mean difference as a function of the signal-to-noise ratio. Figure 8 A comparison of the baseline diagnostic accuracy under 5dB, 0dB, -2dB and -4dB conditions with the diagnostic accuracy after adding generated samples; Figure 9 The baseline diagnostic confusion matrix under 5dB conditions; Figure 10 The diagnostic confusion matrix after adding generated samples under 5dB conditions; Figure 11 The baseline diagnostic confusion moment is given under 0dB conditions; Figure 12Add the diagnostic confusion matrix after generating samples to the 0dB condition; Figure 13 This is the baseline diagnostic confusion matrix under -2dB conditions; Figure 14 This is the diagnostic confusion matrix after adding generated samples under the -2dB condition; Figure 15 This is the baseline diagnostic confusion matrix under -4dB conditions; Figure 16 The diagnostic confusion matrix after adding generated samples under the -4dB condition. Detailed Implementation

[0037] The following describes the invention in gold and silver coin form with specific implementation examples.

[0038] Figures 2-5 Training loss curves for DDPM under different noise conditions and fault type conditions; Figures 6-7 This invention generates quality index variation trends under different noise conditions. Addressing the problem of sparse and discrete samples of fault size and speed combinations in bearing fault diagnosis, and the difficulty in diagnosing missing combinations, this invention constructs a multi-condition semantic-guided diffusion generation method. It provides a multi-condition semantic-guided diffusion generation method for bearing fault samples under missing combinations of varying operating conditions, enabling controllable generation and data augmentation of bearing fault samples with missing fault size-speed combinations. The overall framework of the method includes three core stages: model training, missing combination sample generation, and fault diagnosis enhancement. The flowchart is shown below. Figure 1 As shown, it includes the following steps: Step 1: Collect and preprocess continuous bearing vibration signals labeled with fault type, fault size, and rotational speed. Divide the samples into bearing vibration samples using a fixed window. Based on the fault type, fault size, and rotational speed labels, classify the samples into existing and missing combinations. Construct training data for the generation model, training data for the diagnostic model, and independent test data to obtain the training sample set for existing combinations, the target condition set for missing combinations, and the test sample set for all combinations. Specifically: Step 1.1: In this embodiment, 12kHz drive-end bearing vibration data from the CWRU bearing fault dataset is selected. Fault types include inner ring fault, rolling element fault, and outer ring fault, denoted as IR, B, and OR, respectively; fault sizes include 0.007 inch, 0.014 inch, and 0.021 inch; and operating speeds include 1797 r / min, 1772 r / min, 1750 r / min, and 1730 r / min. Different speeds correspond to different load conditions. In this embodiment, speed is used as the primary operating condition variable, and fault size is used as the fault severity variable.

[0039] Step 1.2: Based on the continuous vibration signal obtained in Step 1.1, a sliding window with a length of 4096 sampling points is used for segmentation, with a sliding step size of 2048 sampling points, so that each sample is a single-channel one-dimensional vibration signal. To examine the strong noise adaptability, noisy samples are constructed under signal-to-noise ratio conditions of 5dB, 0dB, -2dB, and -4dB, respectively. As the signal-to-noise ratio decreases, the amplitude of random noise increases relatively, and the impact structure in the original vibration signal is gradually masked by noise.

[0040] Step 1.3: Based on the fixed-length samples obtained in Step 1.2, the combinations are divided according to "fault type-fault size-speed". For the three fault types IR, B, and OR, each fault type contains 3 fault sizes and 4 speeds, for a total of 12 fault size-speed subcategories. Only 4 subcategories are retained as existing combinations for each fault type, and the remaining 8 subcategories are as missing combinations. Therefore, a total of 12 existing combinations and 24 missing combinations are obtained for the three fault types. The specific combination division is shown in Table 1, where "0.014@1797" indicates a fault size of 0.014 inches and a speed of 1797 r / min.

[0041] Table 1: Existing Combinations and Missing Combinations Settings

[0042] Step 1.4: Based on the combination partitioning obtained in Step 1.3, 80 real samples are selected from each subclass of existing combinations for generating model training and baseline diagnostic model training, totaling 960 real training samples for existing combinations; 80 samples are generated from each subclass of missing combinations in Step 5, totaling 1920 missing combination samples. During the testing phase, 20 real samples that did not participate in training are selected from each of the 36 subclasses, totaling 720 real test samples for all combinations; among them, the 480 real samples corresponding to the missing combinations are only used for generating quality assessments. This embodiment conducts experiments according to the above fixed parameters and does not use test samples for model training or parameter updates.

[0043] Step 2: Based on the existing combined training samples and the set of missing combined target conditions obtained in Step 1, the fault size and rotational speed are normalized and encoded. Conditional denoising and diffusion probability models corresponding to inner race faults, rolling element faults, and outer race faults are constructed and trained respectively, resulting in a structured numerical condition vector c. num And fault type model selection condition g: Step 2.1: Use IR, B, and OR as the three fault type model selection conditions g; for any target sample, first determine the condition DDPM corresponding to IR, B, or OR by g, so as to avoid mutual interference between the distributions of different fault types.

[0044] Step 2.2: Denote the target fault size as d and the target rotational speed as v, and normalize them according to the fault size range and rotational speed range of this embodiment to obtain d_norm and v_norm; the structured numerical condition vector c_num is represented as: in, Represents a structured numerical condition vector; This represents the normalized fault size. This represents the normalized rotational speed. The structured numerical condition vector is used to explicitly constrain the fault level and rotational speed conditions of the target sample.

[0045] Step 2.3: Based on the fault type model selection condition g obtained in Step 2.1 and c_num obtained in Step 2.2, the explicit conditions required for subsequent diffusion generation are formed, where g is responsible for fault category constraints and c_num is responsible for fault size and speed constraints.

[0046] Step 3: Based on the existing combined training samples obtained in Step 1, extract time-domain and frequency-domain statistical features, and organize these statistical features with the target fault type, target fault size, and target rotational speed into task cue words p, obtaining the target condition description used for semantic encoding; specifically: Step 3.1: Extract six time-domain statistical features from the existing combined real training samples: mean, standard deviation, peak value, kurtosis, root mean square value, and kurtosis factor.

[0047] Step 3.2: Based on the samples corresponding to Step 3.1, further extract four frequency domain statistical features: power spectral density, main peak frequency, root mean square frequency, and center frequency.

[0048] Step 3.3: Organize the target fault type, target fault size, target rotational speed, and the time-frequency domain statistical features obtained in steps 3.1 and 3.2 into a fixed-format English task prompt word p. For missing combinations, the target time-frequency domain features in the task prompt word are determined by the statistical features of existing combination samples of the same fault type; real samples of missing combinations do not participate in the training of the generative model. For example, consider a task prompt word with the target conditions of inner ring fault, fault size of 0.007 inches, and rotational speed of 1797 r / min: Generate a bearing vibration signal under the following targetconditions. Fault type: inner race fault. Fault size: 0.007 inch. Shaft speed: 1797 rpm. Target time-domain features: mean 0.0046, standard deviation 0.9982, peak value 6.7339, kurtosis 8.8438, root mean square 0.9984, kurtosis factor 8.8576. Target frequency-domain features: power spectral density 0.9965, dominant peak frequency 2519.3573 Hz, root mean square frequency 3104.3626 Hz, center frequency 2905.1786 Hz. Step 4: Based on the task prompt word p obtained in Step 3, introduce LoRA low-rank adaptation parameters to the GPT-2 semantic encoder and perform efficient parameter fine-tuning. Then, perform word segmentation, semantic encoding, average pooling, and linear projection on the task prompt word to obtain a fixed-dimensional text semantic conditional vector. : Step 4.1, let p be the English task prompt word corresponding to the sample to be generated, and use a token segmenter to convert it into a token sequence of length L: Where p represents the task prompt word input to the GPT-2 semantic encoder; z represents the token sequence obtained after the task prompt word is segmented. L represents the i-th token; L represents the length of the token sequence.

[0049] Step 4.2: Input the token sequence into the LoRA-tuned GPT-2 semantic encoder to obtain the last hidden state matrix: Where H represents the hidden state matrix output by the GPT-2 semantic encoder; The hidden state vector corresponding to the i-th token is represented by dh; dh represents the dimension of the GPT-2 hidden layer. To hide the state space, For the number of Transformer layers, The dimension of each hidden unit.

[0050] Step 4.3: To obtain a fixed-length text semantic condition vector, the hidden states of all token sequences are averaged and mapped to the DDPM condition space via linear projection. in, This represents the text semantic condition vector extracted by the GPT-2 semantic encoder; and These represent the weight matrix and bias term of the linear projection layer, respectively; This represents the hidden state vector corresponding to the i-th token; This indicates the index number of the token in the sequence.

[0051] Textual semantic condition vectors are used to express the time-domain and frequency-domain features of the target signal contained in the task prompts, enabling the generative model to obtain more fine-grained signal distribution constraints.

[0052] Step 4.4: To reduce the number of parameters in GPT-2 fine-tuning, this invention employs the LoRA method for efficient parameter fine-tuning of the GPT-2 semantic encoder. LoRA does not directly update the original weights of the GPT-2 semantic encoder, but instead introduces low-rank trainable increments in the attention module. For the original weight matrix... The equivalent weights after LoRA fine-tuning are expressed as: in, This represents the original frozen weights of the GPT-2 semantic encoder; This represents the low-rank weight increment introduced by LoRA; and For trainable low-rank matrices, For the first low-rank matrix, It is the second low-rank matrix; Indicates the LoRA rank; This represents the scaling factor.

[0053] During training, only the LoRA adaptation parameters and subsequent conditional projection layer parameters are updated, thereby obtaining semantic encoding capabilities for bearing fault generation tasks with low computational overhead.

[0054] Step 5, based on the structured numerical condition vector c obtained in Step 2 num The fault type model selection condition g and the text semantic condition vector c obtained in step 4 text Construct a complete multi-conditional guidance vector c, and perform forward noise-adding training and backward guidance denoising generation on the conditional DDPM corresponding to g to obtain bearing vibration generation samples with target missing fault size-speed combination: The structured conditions consist of fault size and rotational speed, and are expressed as follows: in, Represents a structured numerical condition vector; This represents the normalized fault size; This represents the normalized rotational speed.

[0055] Combining structured numerical condition vectors with textual semantic condition vectors By concatenating the vectors, we obtain the complete multi-condition guided vector of the input condition DDPM: in, This represents the final condition guidance vector. This final condition guidance vector simultaneously contains fault type, fault severity, speed and operating conditions, fault characteristic frequencies, and time-frequency domain statistical feature semantic information.

[0056] in, This represents the noisy vibration signal at step t; t represents the diffusion time step. Represents a multi-condition guiding vector; This represents the noise in the conditional DDPM prediction.

[0057] The inverse denoising network of the final conditional DDPM is represented as follows: In equation (7), x_t represents the noisy bearing vibration sample at the t-th diffusion time step, t represents the current diffusion time step, c represents the multi-conditional guidance vector obtained from equation (6), θ represents the parameters of the conditional DDPM noise prediction network, and ε_θ represents the conditional DDPM prediction noise. The model input and output are both (1,4096)-dimensional one-dimensional signals. The model structure is shown in Table 2.

[0058] Table 2: Model Structure

[0059] In this embodiment, T=500, batch size is 8, training epochs are 100, the optimizer is Adam, and the initial learning rate is 10. -4 Independent conditional DDPMs were trained for IR, B, and OR respectively.

[0060] Figures 2-5Training loss curves for DDPM under fault type conditions of IR, B, and OR are presented at 5dB, 0dB, -2dB, and -4dB respectively. The overall loss under all four noise conditions decreases with increasing training epochs and tends to stabilize at approximately 100 epochs. The curve fluctuates more significantly under the -4dB condition, but still maintains a decreasing and converging trend, indicating that strong noise does not disrupt the model's learning process of the main distribution characteristics of existing combinations.

[0061] To quantify the closeness between generated samples and real samples with missing combinations, this embodiment uses Euclidean Similarity (EUS) and Maximum Mean Discrepancy (MMD) for evaluation. A larger EUS indicates a smaller temporal Euclidean distance between the real and generated samples. A smaller MMD indicates a smaller difference between the distributions of the real and generated samples. The overall evaluation results under different noise conditions are shown in Table 3.

[0062] Table 3: Overall Evaluation Results of Generation Quality under Different Noise Conditions

[0063] Figure 6 This demonstrates the trend of EUS as a function of signal-to-noise ratio. Figure 7 The trend of MMD with signal-to-noise ratio is shown in Table 3. Figure 6 , Figure 7 It can be seen that when the SNR decreases from 5dB to -4dB, the EUS decreases from 0.3773 to about 0.3449 and then stabilizes; the MMD decreases from 0.2897 to 0.2031.

[0064] Step 6: Based on the existing combined training sample set obtained in Step 1 and the bearing vibration generation sample of the target missing fault size-speed combination obtained in Step 5, construct an enhanced training set, train the fault diagnosis model, and perform diagnostic verification on the full combination test sample set obtained in Step 1. This yields a missing combination bearing fault generation sample set and an enhanced fault diagnosis model with verified diagnostic effectiveness, thus completing the multi-condition semantic-guided bearing fault sample diffusion generation under varying operating condition combinations of missing components. Step 6.1: A three-class one-dimensional CNN is used as the fault diagnosis model, with output categories of IR, B, and OR. The baseline is trained using only the 960 existing combined real training samples obtained in Step 1.

[0065] Step 6.2: Combine the 1920 missing combination samples generated in Step 5 with the 960 existing combination real training samples in Step 6.1 to form an enhanced training set, and train the enhanced fault diagnosis model under the same one-dimensional CNN structure and training settings as the baseline.

[0066] Step 6.3: Use the 720 fully combined real test samples retained in Step 1 to test the Baseline and Enhanced Fault Diagnosis Models respectively. All test samples are not used in the training of the Generative or Diagnostic Models. The test results are shown in Table 4.

[0067] Table 4: Diagnostic Enhancement Results of the Entire Real Test Set

[0068] Figure 8 A visual comparison of the diagnostic accuracy under the four noise conditions in Table 4 was performed. The accuracy improved from 0.9958 to 1.0000 under 5dB noise; from 0.9542 to 0.9931 under 0dB noise; from 0.9319 to 0.9847 under -2dB noise; and from 0.7458 to 0.9236 under -4dB noise. As the noise increased, the performance of the baseline decreased more significantly on missing combinations, while the augmented model with generated samples maintained a high accuracy, indicating that the generated samples provided the missing operating condition information from the original training set.

[0069] Figure 9 and Figure 10 The values ​​represent the diagnostic confusion matrices at the baseline and after adding generated samples under the 5dB condition, respectively, corresponding to overall accuracies of 0.9958 and 1.0000. Figure 11 and Figure 12 The two diagnostic results represent the results under 0 dB conditions, with overall accuracies of 0.9542 and 0.9931, respectively. Figure 13 and Figure 14 The figures represent two diagnostic results under a -2dB condition, corresponding to overall accuracies of 0.9319 and 0.9847, respectively. Figure 15 and Figure 16 The figures represent two diagnostic results under a -4dB condition, corresponding to overall accuracies of 0.7458 and 0.9236, respectively. All four sets of results collectively demonstrate that the generated samples provide the most significant diagnostic enhancement in scenarios with strong noise and severe combinational missing values.

[0070] As can be seen from steps 6.1-6.3, the missing combined bearing vibration samples generated in step 5 can supplement diagnostic training information without using the missing combined real samples for generating model training. Therefore, a missing combined bearing fault generation sample set that has been verified as effective by the full combined real test set is obtained, and an enhanced fault diagnosis model is obtained at the same time.

[0071] Ablation experiment and comparative experiment: (1) Ablation experiment. To verify the role of each component in the multi-condition semantic guidance, this embodiment sets up three models: M1, M2 and M3. M1 only uses structured numerical conditions, namely fault size and rotational speed; M2 adds GPT-2 ordinary text semantics to the structured numerical conditions, but does not include time-frequency domain statistical features; M3 is the present invention, which adds task prompt words containing time-domain and frequency-domain statistical features to the structured numerical conditions, and uses LoRA to efficiently fine-tune the parameters of GPT-2.

[0072] The ablation experiment results under the -2dB condition are shown in Table 5. The diagnostic accuracy of M3 is 0.9847, which is higher than that of M1 (0.9833) and M2 (0.9792). The differences in EUS and MMD among the three models are small, indicating that the overall statistical similarity is close at this noise level, but the complete time-frequency semantic condition can still bring diagnostic enhancement advantages.

[0073] Table 5: Ablation Experiment Results under -2dB Condition

[0074] The ablation experiment results under the -4dB condition are shown in Table 6. M1 had the lowest diagnostic accuracy and EUS, and the highest MMD; M2 improved diagnostic accuracy after introducing basic text semantics; M3 further introduced time-frequency domain statistical features and used LoRA fine-tuning, achieving a diagnostic accuracy of 0.9569, while also obtaining the highest EUS and the lowest MMD, indicating that under strong noise conditions, complete multi-condition semantic constraints have a more significant effect on the quality of generated samples and diagnostic performance.

[0075] Table 6: Ablation Experiment Results under -4dB Condition

[0076] (2) Comparative experiment. GAN and VAE were selected as the baseline of the generative model and compared with M3. The three methods used the same existing combination / missing combination division and full combination real test method. Missing combination generated the same number of samples for diagnostic enhancement.

[0077] The results under the -2dB condition are shown in Table 7. The diagnostic accuracy of M3 is 0.9847, which is higher than that of GAN (0.9347) and VAE (0.7917). The MMD of M3 is 0.2053, which is lower than that of GAN and VAE, indicating that the distribution of diffusion-generated samples is closer to the true missing combination distribution.

[0078] Table 7: Comparison of different generative models under -2dB conditions

[0079] The results under the -4dB condition are shown in Table 8. The diagnostic accuracy after GAN enhancement is 0.3861, VAE is 0.7125, and M3 is 0.9236. Meanwhile, the MMD of M3 is 0.2031, which is lower than that of GAN (2.0700) and VAE (0.3747). This indicates that the method of the present invention has a more stable generation distribution and a more obvious diagnostic enhancement effect when small samples, strong noise, and combined deletions coexist.

[0080] Table 8: Comparison Results of Different Generative Models under -4dB Condition

[0081] The above embodiments are merely illustrative of the implementation methods of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for generating bearing fault samples using multi-condition semantic guidance under missing variable operating condition combinations, characterized in that, The bearing fault sample diffusion generation method includes the following steps: Step 1: Collect and preprocess the continuous vibration signal of the bearing with labels of fault type, fault size and speed. Divide the bearing vibration sample according to a fixed window. Divide the existing combination and missing combination according to the labels of fault type, fault size and speed. Construct the training data of the generation model, the training data of the diagnostic model and the independent test data to obtain the training sample set of the existing combination, the target condition set of the missing combination and the test sample set of the whole combination. Step 2: Based on the existing combined training sample set and the missing combined target condition set obtained in Step 1, the fault size and rotational speed are normalized and encoded. Conditional denoising and diffusion probability models corresponding to inner race faults, rolling element faults, and outer race faults are constructed and trained respectively, resulting in structured numerical condition vectors. And fault type model selection condition g; Step 3: Extract time-domain and frequency-domain statistical features based on the existing combined training sample set obtained in Step 1, and organize the statistical features with the target fault type, target fault size and target rotation speed into task prompt words p to obtain the target condition description for semantic encoding; Step 4: Based on the task prompt word p obtained in Step 3, introduce LoRA low-rank adaptation parameters into the GPT-2 semantic encoder and perform efficient parameter fine-tuning. Then, perform word segmentation, semantic encoding, average pooling, and linear projection on the task prompt word to obtain a fixed-dimensional text semantic condition vector. ; Step 5, based on the structured numerical condition vector c num Fault type model selection condition g and text semantic condition vector c text Construct a complete multi-condition guidance vector c, and perform forward noise-adding training and reverse guidance denoising generation on the condition DDPM corresponding to g to obtain bearing vibration generation samples with target missing fault size-speed combination. Step 6: Based on the existing combined training sample set in Step 1 and the bearing vibration generation sample of the target missing fault size-speed combination in Step 5, construct an enhanced training set, train the fault diagnosis model, and perform diagnostic verification on the full combination test sample set obtained in Step 1. This yields a missing combination bearing fault generation sample set and an enhanced fault diagnosis model that have been verified for diagnostic effectiveness, thereby completing the multi-condition semantic-guided bearing fault sample diffusion generation under the missing combination of variable working conditions.

2. The bearing fault sample diffusion generation method under multi-condition semantic guidance under missing variable operating condition combinations as described in claim 1, characterized in that, Specifically, step 1 is as follows: Step 1.1: Obtain the continuous vibration signal of the bearing and its corresponding fault type, fault size and speed label; where the fault type is used to determine the sub-fault type generation model to be trained or called, the fault size is used to describe the fault degree, and the speed is used to describe the operating condition. Step 1.2: The continuous bearing vibration signal obtained in Step 1.1 is segmented according to the preset window length and sliding step size, so that each sample is a one-dimensional vibration signal of fixed length, and each sample inherits the corresponding fault type, fault size and speed label to obtain a fixed length sample. Step 1.3: Based on the fixed-length samples obtained in Step 1.2, classify them according to "fault type-fault size-speed". Under each fault type, define the fault size-speed combination that participates in the training of the generative model as the existing combination, and define the fault size-speed combination that does not participate in the training of the generative model as the missing combination, so as to obtain the set of existing combinations and the set of target conditions for missing combinations. Step 1.4: Based on the existing combination set obtained in Step 1.3, construct the training sample set for the generative model and the training sample set for the benchmark diagnostic model, and separately retain real samples covering all fault size-speed combinations as an independent test sample set; the missing combination target condition set does not participate in the training of the generative model, and is only used for quality evaluation in experimental verification scenarios with corresponding real samples.

3. The bearing fault sample diffusion generation method under multi-condition semantic guidance under missing variable operating condition combinations as described in claim 2, characterized in that, Step 2 specifically includes: Step 2.1: Denote the fault type of each sample as g, the fault size as d, and the rotational speed as v; use the existing combination set of the same fault type to train the conditional DDPM corresponding to g, and use the same g to select the corresponding conditional DDPM when generating missing combination samples; Step 2.2: Normalize the fault size d and the rotational speed v according to their respective value ranges to obtain the normalized fault size d. norm and normalized rotational speed v norm And construct structured numerical condition vectors : Step 2.3, based on the fault type model selection condition g determined in Step 2.1 and the structured numerical condition vector c obtained in Step 2.

2. num This yields explicit operating conditions for subsequent semantic condition fusion; where g is used to constrain the fault category. Used to constrain the fault level and rotational speed conditions of the target sample.

4. The bearing fault sample diffusion generation method under multi-condition semantic guidance under missing variable operating condition combinations as described in claim 3, characterized in that, Step 3 specifically includes: Step 3.1: Extract time-domain statistical features from the existing combined training sample set. The time-domain statistical features include mean, standard deviation, peak value, kurtosis, root mean square value, and kurtosis factor. Step 3.2: Based on the existing combined training samples corresponding to Step 3.1, further extract frequency domain statistical features, including power spectral density, main peak frequency, root mean square frequency, and center frequency. Step 3.3: Organize the target fault type, target fault size, target rotation speed, and the time-domain and frequency-domain statistical features obtained in steps 3.1 and 3.2 into English task prompt words p according to a fixed field order; For missing combination samples, the target time-domain and frequency-domain statistical features are determined by the statistical features of existing combination samples of the same fault type, specifically by using the statistical values ​​of existing combination features or estimating based on existing combination features, without using real samples of missing combinations to participate in the training of the generative model.

5. The bearing fault sample diffusion generation method under multi-condition semantic guidance under missing variable operating condition combinations as described in claim 4, characterized in that, In step 3, for each sample to be generated, an English task prompt word p is first constructed. p includes the fault type, fault size, rotation speed, and time-domain and frequency-domain statistical features of the target signal, so that the generation model can obtain more fine-grained signal distribution constraints.

6. The bearing fault sample diffusion generation method under multi-condition semantic guidance in the absence of variable operating condition combinations according to claim 5, characterized in that, Step 4 specifically includes: Step 4.1, let p be the English task prompt word corresponding to the sample to be generated, and use a token segmenter to convert it into a token sequence of length L: Where p represents the task prompt word input to the GPT-2 semantic encoder; z represents the token sequence obtained after the task prompt word is segmented. L represents the i-th token; L represents the length of the token sequence. Step 4.2: Input the token sequence into the LoRA-tuned GPT-2 semantic encoder to obtain the last hidden state matrix: Where H represents the hidden state matrix output by the GPT-2 semantic encoder; This represents the hidden state vector corresponding to the i-th token; Indicates the dimension of the GPT-2 hidden layer; To hide the state space, The number of Transformer layers; Step 4.3: To obtain a fixed-length text semantic condition vector, the hidden states of all token sequences are averaged and mapped to the DDPM condition space via linear projection. in, This represents the text semantic condition vector extracted by the GPT-2 semantic encoder, used to express the time-domain and frequency-domain features of the target signal contained in the task prompt words, enabling the generative model to obtain more fine-grained signal distribution constraints; and These represent the weight matrix and bias term of the linear projection layer, respectively; Indicates the first i The hidden state vector corresponding to each token; This represents the index number of the token in the sequence; Step 4.4: The LoRA method is used for efficient parameter fine-tuning of the GPT-2 semantic encoder; a low-rank trainable increment is introduced into the attention module; for the original weight matrix... The equivalent weights after LoRA fine-tuning are expressed as: in, This represents the original frozen weights of the GPT-2 semantic encoder; This represents the low-rank weight increment introduced by LoRA; and For trainable low-rank matrices, For the first low-rank matrix, It is the second low-rank matrix; Indicates the LoRA rank; Indicates the scaling factor; During training, only the LoRA adaptation parameters and subsequent conditional projection layer parameters are updated to ultimately obtain the semantic encoding capability for bearing fault generation tasks.

7. The bearing fault sample diffusion generation method under multi-condition semantic guidance under missing variable operating condition combinations as described in claim 6, characterized in that, Step 5 specifically includes: The structured conditions consist of fault size and rotational speed, and are expressed as follows: in, Represents a structured numerical condition vector; This represents the normalized fault size; This represents the normalized rotational speed; Combining structured numerical condition vectors with textual semantic condition vectors By concatenating the vectors, we obtain the complete multi-condition guided vector of the input condition DDPM: in, This represents the final condition guidance vector; the final condition guidance vector simultaneously contains fault type, fault severity, speed condition, fault characteristic frequency, and time-frequency domain statistical feature semantic information. The inverse denoising network of the final conditional DDPM is represented as follows: in, This represents the noisy vibration signal at step t; t represents the diffusion time step. Represents a multi-condition guiding vector; This represents the noise predicted by conditional DDPM. The Conditional Denoising Diffusion Model (DDPM) is subject to explicit constraints such as fault type, fault size, and rotational speed during the reverse denoising process. It is also constrained by time-frequency domain semantic information encoded by the GPT-2 semantic encoder, thereby enabling the generation of controllable fault samples for missing fault size-rotational speed combinations.

8. The bearing fault sample diffusion generation method under multi-condition semantic guidance under missing variable operating condition combinations as described in claim 7, characterized in that, Specifically, step 6 includes: Step 6.1: Use the existing combined training sample set obtained in Step 1 to train the baseline fault diagnosis model and obtain the baseline diagnosis results; Step 6.2: Add the bearing vibration generation sample of the target missing fault size-speed combination obtained in Step 5 to the existing combination real diagnostic training sample used in Step 6.1 to construct an enhanced training set, and train the enhanced fault diagnosis model under the same diagnostic network structure and training settings. Step 6.3: Use the full combination test sample set obtained in Step 1 to test the Baseline diagnostic model and the enhanced fault diagnosis model respectively. Evaluate the supplementary role of the generated samples in missing combination diagnostic information by the diagnostic accuracy. When there are real samples with missing combinations in the experimental scenario, further evaluate the generation quality by the temporal similarity and distribution difference between the generated samples and the real samples.

Images