Class imbalance data driven variable working condition equipment key component fault diagnosis method

Abstract

The application discloses a kind of variable working condition equipment key component fault diagnosis method of class disequilibrium data driving. For the key rotating component in electromechanical complex equipment, obtain the vibration signal and its fault type under the normal and fault state operation of component, construct diagnostic data set and train model, so as to obtain the diagnostic result of the sensor data to be predicted using the trained fault diagnosis model. The application considers the complex situation that the current industrial monitoring data is under variable working condition and class disequilibrium, firstly applies channel-by-channel convolution to generate local feature enhancement weight matrix, and realizes feature enhancement in local feature space using fewer training parameters;Design and introduce wide-range convolution kernel considering the inherent physical characteristics of rotating components, fully and stably capture effective feature information under variable working condition, realize end-to-end fault diagnosis of key rotating components of electromechanical equipment. The application is simple and efficient, not only has theoretical innovation, but also has high practicability.

Description

Technical Field

[0001] This invention relates to a method for intelligent diagnosis of faults in key rotating components of complex electromechanical equipment, specifically a method for diagnosing faults in key components of equipment under varying operating conditions driven by a type of unbalanced data. Background Technology

[0002] Rotating components in electromechanical equipment play a crucial role in fields such as advanced rail transit equipment, aero-engines, and industrial and agricultural equipment. These components are highly susceptible to wear and surface damage. Timely and real-time fault diagnosis of rotating components in electromechanical equipment can significantly reduce maintenance and downtime costs for production lines and equipment. Under complex operating conditions, such as non-constant operating conditions, the hidden features contained in the data are more complex. Shallow networks struggle to uncover sufficient effective features, while deep networks, with their dramatically increased number of network parameters, are prone to overfitting and increased training time. Furthermore, when the monitored data types are unevenly distributed, the model learns prior information including the proportion of samples in the training set, leading to biased predictions across multiple categories and hindering the model's learning of more fundamental features. Therefore, purely data-driven fault diagnosis methods that do not consider common real-world engineering scenarios often lack robustness and generalization. End-to-end fault diagnosis research for rotating components in electromechanical equipment under complex operating conditions is essential. This ensures the healthy operation of electromechanical equipment, enables early fault detection, and prevents potential problems.

[0003] In the field of fault diagnosis, convolutional neural networks (CNNs) have gained widespread use due to their advantages of parallel computing, low hardware requirements, and excellent image data feature extraction capabilities. Traditional CNNs achieve stronger analytical capabilities by increasing the number of layers; however, with increasing depth, gradient explosion or vanishing becomes increasingly severe, leading to network degradation and significantly impacting training performance. Meanwhile, the vibration signals of rotating components in faulty electromechanical equipment contain a series of impulse responses. Collecting fault information and capturing long-term dependencies within the signals over a long time span is crucial for fault diagnosis. CNNs using smaller kernels have limited receptive fields, resulting in deficiencies in learning long-term dependencies. Simply increasing the kernel size introduces too many parameters, significantly increasing computational load and time. Another common method to increase the receptive field is dilated convolution, which can expand the receptive field, but the presence of zero elements in the kernel may lead to information loss. Finally, existing diagnostic methods often require preprocessing of monitoring signals for further feature extraction, failing to achieve end-to-end diagnosis, which is detrimental to the trend towards real-time monitoring in diagnostics.

[0004] Uneven distribution of sample categories is a widespread problem in industrial data. Solutions to this problem primarily focus on two aspects: data and algorithms. From a data perspective, methods include expanding the dataset and implementing balanced sampling. However, in practical engineering applications, fundamentally increasing the number of samples leads to increased monitoring costs during production. On the other hand, balanced sampling methods suffer from insufficient overall sample size due to the presence of very few samples from different categories, failing to provide enough data for data-driven fault diagnosis models. Therefore, reducing the robustness of models caused by data imbalance at the algorithmic level can lower fault monitoring costs and make fault diagnosis more convenient and efficient.

[0005] Considering the above, this paper proposes a fault diagnosis method for key components of equipment under varying operating conditions driven by unbalanced data. It is very important to solve the problems of low efficiency and large accuracy deviation in the classification of common unbalanced data in the fault diagnosis of rotating parts of electromechanical equipment in practical engineering. Summary of the Invention

[0006] To address the problems mentioned in the background art, this invention proposes a fault diagnosis method for key components of equipment under varying operating conditions driven by unbalanced data. This method enables end-to-end fault diagnosis of rotating components of electromechanical equipment under varying operating conditions and with unbalanced data monitoring data. It can be widely applied to fault diagnosis of rotating components of electromechanical equipment in practical engineering.

[0007] To achieve the above functions, the technical solution of the present invention is as follows:

[0008] S1: Collect vibration signals from key components of electromechanical equipment;

[0009] S2: After preprocessing the collected vibration signals, a vibration data sample set is obtained;

[0010] S3: Construct a multi-scale wide-domain convolutional residual network fault diagnosis model, and train the multi-scale wide-domain convolutional residual network fault diagnosis model using vibration data sample set to obtain a trained multi-scale wide-domain convolutional residual network fault diagnosis model.

[0011] S4: Input the vibration data samples to be tested into the trained multi-scale wide-domain convolutional residual network fault diagnosis model to obtain the corresponding diagnosis results.

[0012] Specifically, S1 is:

[0013] Accelerometers are installed on rotating parts of electromechanical equipment, and the corresponding vibration signals are collected by the accelerometers under normal and fault conditions of the rotating parts.

[0014] In S1, the vibration signal includes vibration signals in both the horizontal and vertical directions.

[0015] The multi-scale wide-area convolutional residual network fault diagnosis model includes a first convolutional layer, a first batch of normalization layers, a first activation layer, a flattening layer, and multiple multi-scale wide-area convolutional residual blocks. The input of the multi-scale wide-area convolutional residual network fault diagnosis model is used as the input of the first convolutional layer. The first convolutional layer is connected to the first activation layer after passing through the first batch of normalization layers. The first activation layer is connected to the flattening layer after passing through multiple cascaded multi-scale wide-area convolutional residual blocks. The output of the flattening layer is used as the output of the multi-scale wide-area convolutional residual network fault diagnosis model.

[0016] The multiple multi-scale wide-domain convolutional residual blocks have the same structure. Each multi-scale wide-domain convolutional residual block includes a wide-domain channel-wise convolutional local feature enhancement block, a second convolutional layer, a second batch normalization layer, a second activation function layer, a channel-wise convolutional layer, a third batch normalization layer, and a third activation function layer. The input of the multi-scale wide-domain convolutional residual block is used as the input of the wide-domain channel-wise convolutional local feature enhancement block. The wide-domain channel-wise convolutional local feature enhancement block is sequentially passed through the second convolutional layer, the second batch normalization layer, the second activation function layer, and the channel-wise convolutional layer before being connected to the third batch normalization layer. The output of the third batch normalization layer is added to the input of the multi-scale wide-domain convolutional residual block and then input into the third activation function layer. The output of the third activation function layer is used as the output of the multi-scale wide-domain convolutional residual block.

[0017] The wide-domain channel-wise convolutional local feature enhancement block includes a first wide-domain channel-wise convolutional layer, a fourth activation function layer, a second wide-domain channel-wise convolutional layer, and a fifth activation function layer. The input of the wide-domain channel-wise convolutional local feature enhancement block is used as the input of the first wide-domain channel-wise convolutional layer. The first wide-domain channel-wise convolutional layer is connected to the fifth activation function layer after passing through the fourth activation function layer and the second wide-domain channel-wise convolutional layer in sequence. The output of the fifth activation function layer is used as the output of the wide-domain channel-wise convolutional local feature enhancement block after performing element-wise multiplication with the input of the wide-domain channel-wise convolutional local feature enhancement block.

[0018] In the first wide-area channel-wise convolutional layer and the second wide-area channel-wise convolutional layer, the kernel size k is greater than twice the reciprocal of the minimum rotating component failure frequency under varying operating conditions, as shown in the following formula:

[0019]

[0020] Among them, f c To monitor the data acquisition frequency, t c To monitor the duration of data collection; f failure This represents the failure frequency of rotating components.

[0021] In S3, the Adam optimizer is used during the training process of the multi-scale wide-domain convolutional residual network fault diagnosis model.

[0022] In S3, during the training process of the multi-scale wide-domain convolutional residual network fault diagnosis model, for vibration data sample sets with imbalanced classes, the training loss function uses the weighted cross-entropy function.

[0023] In S3, during the training process of the multi-scale wide-domain convolutional residual network fault diagnosis model, for multi-classification tasks, the softmax function is selected as the activation function of the output layer.

[0024] This invention is the first to apply channel-wise convolution to generate a local feature enhancement weight matrix. Under varying operating conditions, it enhances the features in the local feature space of the original monitoring data with fewer training parameters, making deep fault features more prominent and providing a foundation for end-to-end fault diagnosis without feature pre-extraction. It uses wide-domain convolution kernels, combined with the inherent physical characteristics of the faulty rotating component, to stably encompass the information contained in adjacent pulses under varying operating conditions, achieving stable capture of effective feature information under these conditions. It uses a weighted cross-entropy function to lightweightly address the accuracy bias caused by data class imbalance at the algorithm level. It uses stacked multi-scale wide-domain convolution residual blocks to increase network depth, deeply mine the fault feature information of the rotating component, and map the extracted features onto the diagnostic results, achieving robust and efficient fault diagnosis of key components of electromechanical equipment under varying operating conditions.

[0025] Compared with existing technologies, the present invention has the following advantages:

[0026] Compared to traditional methods, the proposed method utilizes channel-wise convolution to generate local feature enhancement weight matrices. Compared to the "contextual relationship modeling" method of attention models, channel-wise convolution, with its data inductive bias, can extract the model's potential with fewer raw samples, resulting in more stable training. In terms of reducing computational costs, channel-wise convolution reduces the number of network parameters. Simultaneously, the residual structure in the overall network model and the inherent parallel computing characteristics of convolution accelerate network training, reducing computation time and enabling more timely and rapid fault diagnosis of rotating parts in electromechanical equipment. Under varying operating conditions, purely data-driven fault diagnosis methods for rotating parts of electromechanical equipment, which are sensitive to training samples, often exhibit poor generalization and robustness. However, by designing and introducing a wide-domain convolution kernel that considers the inherent physical characteristics of the faulty rotating part, the method stably encompasses the information contained between adjacent pulses of the monitoring signal, achieving stable capture of effective feature information under varying operating conditions. Using a weighted cross-entropy function as the training loss function reduces the problem of insufficient model robustness caused by imbalanced data samples at the algorithmic level, lowering fault monitoring costs and making fault diagnosis more convenient and efficient.

[0027] The method of this invention takes into account the complex situation where actual monitoring data in the current industry is under varying operating conditions and is uneven in category. It uses a method that combines algorithm characteristics with signal physical characteristics to realize end-to-end fault diagnosis of key components of electromechanical equipment. It not only has theoretical innovation, but also has high practicality. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the method of the present invention.

[0029] Figure 2 This invention presents a multi-scale residual network model structure with local feature enhancement for fault diagnosis of key components of electromechanical equipment under variable operating conditions driven by unbalanced data.

[0030] Figure 3 This is a graph showing the trend of training error during the training process in an embodiment of the present invention.

[0031] Figure 4 This is the confusion matrix of classification results on the test set in this embodiment of the invention. Detailed Implementation

[0032] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0033] Specific embodiments of the present invention are as follows:

[0034] like Figure 1 As shown, this embodiment uses a set of fault data from healthy bearings running naturally to failure under varying operating conditions to experimentally verify the fault diagnosis method for key components of equipment under varying operating conditions proposed in this invention, which is driven by a kind of unbalanced data.

[0035] First, the verification dataset used in this invention is obtained by collecting vibration signals of the corresponding failed bearing after an accelerated degradation experiment was conducted on a bearing degradation test platform until the bearing failed. This bearing degradation test platform mainly includes a rotating section, a loading section, and a measurement section. The rotating section is driven by a motor, keeping the rolling bearing in a continuous working state. The motor speed is set to 25Hz, and the actual measured bearing speed is 23.4Hz. To accelerate bearing degradation, the loading section applies a controllable radial load to the running rolling bearing; the initial radial load is 1470N (~0.15T). The loading device uses a hand-held hydraulic pump, which, during operation, irresistibly and naturally removes part of the load, providing a varying load during the experiment. The measurement section uses an accelerometer fixed to the bearing housing to measure the vertical vibration signal of the rolling bearing. When collecting vibration signals from faulty bearings, the accelerometer sampled every 1 minute for 20 seconds at a frequency of 12.8 kHz, resulting in 256,000 data sets per sampling. These 256,000 data sets were compiled into a single file. The tested bearing was a deep groove ball bearing (Type 6006). Through the above experiment, vibration signals were collected from one of ten bearings of the same type and from the same manufacturer, while the remaining bearings were allowed to run naturally until failure. This generated fault monitoring data for nine different bearing failure states. The specific experimental conditions and corresponding fault types are shown in Table 1. The distribution of fault types among the bearings in this dataset is uneven.

[0036] Table 1 shows the experimental conditions and corresponding data failure types.

[0037]

[0038] After obtaining the faulty bearing dataset, 60% of the samples in the dataset are used as the training set and 40% of the samples are used as the test set.

[0039] Next, a multi-scale wide-domain convolutional residual network fault diagnosis model with local feature enhancement capabilities is constructed, such as... Figure 2 As shown, the multi-scale wide-domain convolutional residual network fault diagnosis model is trained, its parameters are adjusted, and it is evaluated using a vibration data sample set to obtain a well-trained multi-scale wide-domain convolutional residual network fault diagnosis model.

[0040] The multi-scale wide-domain convolutional residual network fault diagnosis model includes a first convolutional layer, a first batch of normalization layers, a first activation layer, a flattening layer, and multiple multi-scale wide-domain convolutional residual blocks. The input to the multi-scale wide-domain convolutional residual network fault diagnosis model is used as the input to the first convolutional layer. The first convolutional layer passes through the first batch of normalization layers and is then connected to the first activation layer. The first activation layer passes through multiple cascaded multi-scale wide-domain convolutional residual blocks and is then connected to the flattening layer. The output of the flattening layer is used as the output of the multi-scale wide-domain convolutional residual network fault diagnosis model. In this embodiment, the number of multi-scale wide-domain convolutional residual blocks is 5.

[0041] Multiple multi-scale wide-domain convolutional residual blocks have the same structure. Each multi-scale wide-domain convolutional residual block includes a wide-domain channel-wise convolutional local feature enhancement block, a second convolutional layer, a second batch normalization layer, a second activation function layer, a channel-wise convolutional layer, a third batch normalization layer, and a third activation function layer. The input of the multi-scale wide-domain convolutional residual block is used as the input of the wide-domain channel-wise convolutional local feature enhancement block. The wide-domain channel-wise convolutional local feature enhancement block is sequentially passed through the second convolutional layer, the second batch normalization layer, the second activation function layer, and the channel-wise convolutional layer before being connected to the third batch normalization layer. The output of the third batch normalization layer is added to the input of the multi-scale wide-domain convolutional residual block and then input into the third activation function layer. The output of the third activation function layer is used as the output of the multi-scale wide-domain convolutional residual block, realizing the effective mining of deep features of rotating parts of electromechanical equipment under varying operating conditions.

[0042] The wide-domain channel-wise convolutional local feature enhancement block includes a first wide-domain channel-wise convolutional layer, a fourth activation function layer, a second wide-domain channel-wise convolutional layer, and a fifth activation function layer. The input of the wide-domain channel-wise convolutional local feature enhancement block is used as the input of the first wide-domain channel-wise convolutional layer. The first wide-domain channel-wise convolutional layer is then connected to the fifth activation function layer after passing through the fourth and second activation function layers in sequence. The output of the fifth activation function layer is multiplied element-wise with the input of the wide-domain channel-wise convolutional local feature enhancement block, and the output is used as the output of the wide-domain channel-wise convolutional local feature enhancement block, thereby achieving local feature enhancement of the input data.

[0043] In the first and second wide-domain channel-wise convolutional layers, "wide-domain" specifically means using a larger convolutional kernel to cover the positions of adjacent pulses. The kernel size k is greater than twice the reciprocal of the minimum rotating component failure frequency under varying operating conditions, allowing each neuron in the convolutional layer to interact with the information contained in adjacent pulses. The wide-domain convolutional kernel size k should not deviate too much from its minimum value; a larger convolutional kernel will introduce excessive parameters, increasing the network's computational burden and time. The specific formula is shown below:

[0044]

[0045] Among them, fc To monitor the data acquisition frequency (Hz), t c For monitoring data collection duration (s); f failure The empirical failure frequency of rotating components is determined by the specific type of rotating component.

[0046] Taking rolling bearings as an example, the experienced failure frequency f of the bearing failure The formula is: n is the bearing operating speed (rad / s), P is the bearing relative load (N), and z is the number of bearing rollers.

[0047] During the training of the multi-scale wide-domain convolutional residual network fault diagnosis model, the Adam optimizer is used. For imbalanced vibration data sample sets, the weighted cross-entropy function (Weighted CE) is used as the training loss function; for multi-class classification tasks, the softmax function is selected as the activation function for the output layer. The mathematical expression of the weighted cross-entropy function is: Weighted CE softmax =-α c (1-P c ) γ log(P c ), where α c P represents the weight of the c-th class sample. c This represents the probability value of the c-th class sample output by softmax, and the hyperparameter γ is the decay parameter. In this embodiment, the sample weight vector α c After fine-tuning based on the amount of data in each category of the sample, the values ​​are taken as [1, 3, 6, 6]. The decay parameter γ is set to 0 after parameter tuning. The training process is as follows: Figure 3 As shown.

[0048] In this embodiment, the size of the first-layer wide-area convolution kernel is calculated based on the experimental conditions and the inherent parameters of the corresponding bearing type. The smallest integer value that the convolution kernel size can take is 21.

[0049] The channel count and kernel size settings for the five multi-scale wide-area convolutional residual blocks are shown in Table 2.

[0050] Table 2 shows the parameters for the number of channels and kernel size of the five multi-scale wide-domain convolutional residual blocks.

[0051]

[0052] Finally, the vibration data samples to be tested are input into the trained multi-scale wide-domain convolutional residual network fault diagnosis model to obtain the corresponding diagnosis results.

[0053] In this embodiment, 40% of the samples in the dataset are used as the test set to test the trained model. Due to the imbalanced nature of the dataset, in order to accurately evaluate the prediction performance, this example uses three commonly used evaluation metrics in imbalanced data processing to evaluate the prediction results: precision, recall, and the harmonic mean (F1) based on precision and recall.

[0054] To avoid the influence of random factors, six experiments were conducted for the four unevenly distributed fault types, and the precision, recall, and harmonic mean of the classification task test set were recorded. The mean and standard deviation of the above three indicators were calculated as the final evaluation parameters. In this embodiment, the mean and standard deviation of the three evaluation parameters are shown in Table 3:

[0055] Table 3 shows the average and standard deviation of the three evaluation parameters in this embodiment.

[0056]

[0057]

[0058] As shown in Table 3, for each category of the classification experiment, the harmonic mean of precision and recall is greater than 99.9%. Meanwhile, based on... Figure 4 The results of the confusion matrix on the test set demonstrate that the proposed scheme has high classification accuracy and good classification performance. Furthermore, the small standard deviations of the three experimental parameters indicate minimal fluctuations and reliable experimental results. These experimental results demonstrate that the proposed method is feasible and effective for fault diagnosis of rotating components in electromechanical equipment with imbalanced data under varying operating conditions.

[0059] This invention is the first to apply channel-wise convolution to generate a local feature enhancement weight matrix. Under varying operating conditions, it achieves feature enhancement in the local feature space using fewer training parameters on the original monitoring data. The data inductive bias of channel-wise convolution allows for the extraction of model potential with fewer original samples, making training more stable and providing a foundation for end-to-end fault diagnosis without feature pre-extraction. Under varying operating conditions, purely data-driven fault diagnosis methods for rotating electromechanical equipment components, which are sensitive to training samples, often exhibit poor generalization and robustness. This invention designs and introduces a wide-domain convolution kernel that considers the inherent physical characteristics of the faulty rotating components. Combined with these inherent physical characteristics, it stably encompasses the information contained between adjacent pulses of the monitoring signal under varying operating conditions, thus achieving stable extraction of effective feature information under these conditions. Stacked multi-scale wide-domain convolutional residual blocks are used to increase network depth and deeply mine the fault feature information of the rotating components. The extracted features are mapped onto the diagnostic results, achieving robust and efficient fault diagnosis of key components of electromechanical equipment under varying operating conditions.

[0060] The above embodiments represent an optimization result of the bearing dataset used in this embodiment of the present invention, but the specific implementation of the present invention is not limited to the above embodiments. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention; at the same time, for those skilled in the art, there will be changes in specific implementation methods and application scope based on the ideas of the present invention. In summary, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A fault diagnosis method for key components of equipment under variable operating conditions driven by unbalanced data, characterized in that, Includes the following steps: S1: Collect vibration signals from key components of electromechanical equipment; S2: After preprocessing the collected vibration signals, a vibration data sample set is obtained; S3: Construct a multi-scale wide-domain convolutional residual network fault diagnosis model, and train the multi-scale wide-domain convolutional residual network fault diagnosis model using vibration data sample set to obtain a trained multi-scale wide-domain convolutional residual network fault diagnosis model. The multi-scale wide-domain convolutional residual network fault diagnosis model includes a first convolutional layer, a first batch of normalization layers, a first activation layer, a flattening layer, and multiple multi-scale wide-domain convolutional residual blocks. The input of the multi-scale wide-domain convolutional residual network fault diagnosis model is used as the input of the first convolutional layer. The first convolutional layer is connected to the first activation layer after passing through the first batch of normalization layers. The first activation layer is connected to the flattening layer after passing through multiple cascaded multi-scale wide-domain convolutional residual blocks. The output of the flattening layer is used as the output of the multi-scale wide-domain convolutional residual network fault diagnosis model. The multiple multi-scale wide-domain convolutional residual blocks have the same structure. Each multi-scale wide-domain convolutional residual block includes a wide-domain channel-wise convolutional local feature enhancement block, a second convolutional layer, a second batch normalization layer, a second activation function layer, a channel-wise convolutional layer, a third batch normalization layer, and a third activation function layer. The input of the multi-scale wide-domain convolutional residual block is used as the input of the wide-domain channel-wise convolutional local feature enhancement block. The wide-domain channel-wise convolutional local feature enhancement block is sequentially passed through the second convolutional layer, the second batch normalization layer, the second activation function layer, and the channel-wise convolutional layer before being connected to the third batch normalization layer. The output of the third batch normalization layer is added to the input of the multi-scale wide-domain convolutional residual block and then input into the third activation function layer. The output of the third activation function layer is used as the output of the multi-scale wide-domain convolutional residual block. The wide-domain channel-wise convolutional local feature enhancement block includes a first wide-domain channel-wise convolutional layer, a fourth activation function layer, a second wide-domain channel-wise convolutional layer, and a fifth activation function layer. The input of the wide-domain channel-wise convolutional local feature enhancement block is used as the input of the first wide-domain channel-wise convolutional layer. The first wide-domain channel-wise convolutional layer is connected to the fifth activation function layer after passing through the fourth activation function layer and the second wide-domain channel-wise convolutional layer in sequence. The output of the fifth activation function layer is used as the output of the wide-domain channel-wise convolutional local feature enhancement block after performing element-wise multiplication with the input of the wide-domain channel-wise convolutional local feature enhancement block. S4: Input the vibration data samples to be tested into the trained multi-scale wide-domain convolutional residual network fault diagnosis model to obtain the corresponding diagnosis results.

2. The method for fault diagnosis of key components of equipment under variable operating conditions driven by unbalanced data as described in claim 1, characterized in that, Specifically, S1 is: Accelerometers are installed on rotating parts of electromechanical equipment, and the corresponding vibration signals are collected by the accelerometers under normal and fault conditions of the rotating parts.

3. A method for fault diagnosis of key components of equipment under variable operating conditions driven by unbalanced data as described in claim 1 or 2, characterized in that, In S1, the vibration signal includes vibration signals in both the horizontal and vertical directions.

4. The method for fault diagnosis of key components of equipment under variable operating conditions driven by unbalanced data as described in claim 1, characterized in that, In the first wide-area channel-wise convolutional layer and the second wide-area channel-wise convolutional layer, the kernel size k is greater than twice the reciprocal of the minimum rotating component failure frequency under varying operating conditions, as shown in the following formula: in, To monitor the data collection frequency, To monitor the duration of data collection; This represents the failure frequency of rotating components.

5. The method for fault diagnosis of key components of equipment under variable operating conditions driven by unbalanced data according to claim 1, characterized in that, In S3, the Adam optimizer is used during the training process of the multi-scale wide-domain convolutional residual network fault diagnosis model.

6. The method for fault diagnosis of key components of equipment under variable operating conditions driven by unbalanced data according to claim 1, characterized in that, In S3, during the training process of the multi-scale wide-domain convolutional residual network fault diagnosis model, for vibration data sample sets with imbalanced classes, the training loss function uses the weighted cross-entropy function.

7. The method for fault diagnosis of key components of equipment under variable operating conditions driven by unbalanced data according to claim 1, characterized in that, In S3, during the training process of the multi-scale wide-domain convolutional residual network fault diagnosis model, for multi-classification tasks, the softmax function is selected as the activation function of the output layer.

Images