Deep learning-based device fault diagnosis methods, storage media, and software products

Abstract

This invention discloses a deep learning-based equipment fault diagnosis method. It adopts an incremental learning architecture, which can acquire knowledge from old tasks, enabling the equipment fault diagnosis model to learn to diagnose new equipment operating data. At the same time, it retains the knowledge learned in previous tasks, which can avoid retraining the model parameters when new data arrives.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing, and in particular relates to a deep learning-based equipment fault diagnosis method, storage medium, and program product. Background Technology

[0002] Fault diagnosis of industrial equipment is a crucial link in ensuring the stable operation of production lines, especially the fault detection of critical components in complex mechanical equipment (such as belt conveyors for transmission). Bearings are core components of such equipment, and their operating status directly affects the overall performance and service life of the equipment. Traditional fault diagnosis methods typically rely on the analysis and modeling of offline data. However, with the development of industrial data acquisition and sensing technologies, the acquisition of real-time data has become more common. How to effectively utilize this real-time data to improve the performance of fault diagnosis models has become an important research topic in this field. In practical applications, equipment operation data is characterized by high dimensionality, nonlinearity, and dynamic changes. Traditional methods have significant shortcomings in handling these characteristics. For example, dimensionality reduction techniques are difficult to adapt to complex nonlinear features, classification models lack adaptability to real-time data, and cannot efficiently integrate historical and new data. Furthermore, in the learning process between new and old tasks, common methods require retraining the entire model, which wastes computational resources and may lead to the forgetting of knowledge from old tasks. Therefore, there is an urgent need for a diagnostic framework that can continuously learn from new data while retaining the diagnostic capability for old data, and efficiently extracting and classifying key features of high-dimensional vibration data.

[0003] Chinese invention patent application number CN202110667841.8 discloses a method and apparatus for diagnosing faults in electromechanical equipment, relating to the field of electromechanical equipment fault diagnosis technology. The method includes the following steps: acquiring the current signal of the secondary cable of a current transformer in the electromechanical equipment; preprocessing the current signal and then performing high-speed sampling to obtain current waveform sampling data; performing time-domain and frequency-domain conversion on the current waveform sampling data to obtain corresponding spectrum data; identifying corresponding fault characteristics in the spectrum data and / or time-domain waveform sampling data, and performing fault diagnosis. This invention installs a current sensor or transformer on the secondary cable of the current transformer in existing electromechanical equipment to acquire the current signal of the electromechanical equipment for fault diagnosis. Since the secondary current is relatively small, a small-sized current sensor or transformer can be selected, greatly reducing costs and making it suitable for batch installation. Furthermore, installation does not require stopping the electromechanical equipment, the installation method is simple, and it does not affect the normal use of the electromechanical equipment.

[0004] Chinese invention patent application number CN202110101052.8 proposes an equipment fault diagnosis system and method. The equipment fault diagnosis system includes a diagnosis center and several diagnosis units. Each diagnosis unit includes an industrial control computer, an infrared thermal imaging device, a visible light imaging device, and a communication module, all located at the target equipment. The diagnosis includes the following steps: S1: Initialize the equipment fault diagnosis system; S2: The diagnosis unit collects the current operating data of the target equipment; S3: Analyze the operating data to obtain the fault diagnosis result of the current target equipment; S4: Send the fault diagnosis result of the current target equipment to the diagnosis center. This invention solves the problems of high labor costs, low detection efficiency, and low fault analysis accuracy in existing technologies.

[0005] The existing technology still has the following problems that require further solutions:

[0006] 1. In equipment fault diagnosis tasks, traditional equipment operation data dimensionality reduction methods such as principal component analysis or linear discriminant analysis are difficult to dynamically adapt to the complex nonlinear characteristics of equipment vibration data, resulting in limited dimensionality reduction effects and affecting the feature extraction and classification performance of equipment fault diagnosis models.

[0007] 2. In equipment fault diagnosis tasks, traditional fault diagnosis models such as random forest classifiers often rely on a single feature when splitting nodes, ignoring the synergistic effect between features. This makes it impossible to fully utilize the key information in equipment vibration data, thus limiting classification accuracy.

[0008] 3. In equipment fault diagnosis tasks, the classification problem of boundary regions has not been effectively solved in traditional methods. Equipment fault diagnosis models are prone to misjudgment in these regions, which reduces the accuracy of fault diagnosis. Summary of the Invention

[0009] This invention relates to a device fault diagnosis method, storage medium, and program product based on deep learning. It inputs real-time collected data into a learning model, enabling the fault diagnosis model to learn the features in the real-time collected data. While maintaining the original classification effect, it integrates and improves the classification ability from the real-time data, thereby realizing incremental learning of the fault diagnosis model.

[0010] To address the aforementioned problems, the present invention provides a device fault diagnosis method, storage medium, and program product based on deep learning.

[0011] Among them, deep learning-based equipment fault diagnosis methods include,

[0012] S1. Obtain the old dataset and train the old dataset offline to obtain the initial model; wherein, the old dataset is the equipment vibration data collected by the sensor;

[0013] S2. Acquire the dataset in real time, whereby the real-time acquired data is referred to as a new dataset; train the initial model using the new dataset; the new dataset is real-time device vibration data collected by sensors;

[0014] S3. Use the old dataset to test the accuracy of the initial model trained on the new dataset. If the accuracy of the initial model is improved, the new dataset is usable; if the accuracy of the initial model is not improved, the new dataset is not usable.

[0015] S4. Merge the available new dataset and the old dataset to form a merged dataset;

[0016] S5. Train the current initial model using the fused dataset to generate the current initial model. Use the new dataset to test the accuracy of the current initial model trained with the fused dataset. If the accuracy of the current initial model is improved, the fused dataset training is usable. Then return to S4.

[0017] S6. When all data fails to improve the accuracy of the current initial model, training is terminated, and the parameters of the equipment fault diagnosis model are obtained to obtain the equipment fault diagnosis model. The equipment fault diagnosis model adopts a random forest classifier model based on sparse cooperation. A deep autoencoder based on transformation projection is used to reduce the dimensionality of the equipment vibration data features. Traditional dimensionality reduction methods usually rely on fixed dimensionality reduction strategies, such as principal component analysis or linear discriminant analysis. These methods are difficult to handle the nonlinear relationships of complex equipment vibration data. The transformation projection can dynamically adjust the direction and method of projection according to the different characteristics of the equipment vibration data, so that the dimensionality reduction process can better adapt to the complex structure that may exist in the equipment vibration data and overcome the limitations of traditional dimensionality reduction methods in handling the nonlinear relationships of equipment vibration data.

[0018] S7. Input the dimensionality-reduced equipment vibration data into a random forest classifier model based on sparse cooperation to perform equipment fault diagnosis and obtain the diagnosis results.

[0019] Furthermore, a transformation projection-based depth autoencoder is used to reduce the dimensionality of the device vibration data. The device vibration data input to the transformation projection-based depth autoencoder is standardized, and an adaptive projection matrix is ​​dynamically generated to project the device vibration data into a low-dimensional space, represented as:

[0020]

[0021] In the formula, The adaptive projection matrix is ​​adjusted adaptively during training to ensure that key information of the input device vibration data is preserved as much as possible in the low-dimensional space after projection. Vibration data of the device projected into a low-dimensional space; This is the standardized data of the device vibration data input from the depth autoencoder.

[0022] Furthermore, the adaptive projection matrix is ​​dynamically optimized during the adjustment process using gradient descent, and is expressed as follows:

[0023]

[0024] In the formula, For the first The adaptive projection matrix for the next iteration; For the first The adaptive projection matrix for the next iteration; The learning rate is the adaptive projection matrix; It is the gradient of the loss function of the deep autoencoder with respect to the adaptive projection matrix; It is an adaptive adjustment term; The loss function for a deep autoencoder; It is the first The dynamic adjustment factor is adjusted through gradient local sensitivity in each iteration. During the adjustment process, the adaptive projection matrix aims to retain the most effective information from the equipment vibration data, reduce redundant features, and avoid losing important information. Figure 4 As shown, the dynamic adjustment process of the adaptive projection matrix is ​​analyzed, and the adaptive evolution mechanism of the projection matrix during training is visualized. Six heatmaps are used to show the changes in the 8×8 projection matrix from 0 to 1000 iterations. The horizontal axis represents the input dimension (original vibration feature dimension), and the vertical axis represents the projection dimension (feature dimension after dimensionality reduction). Color mapping represents weight values ​​(blue for negative weights, red for positive weights). Experiments show that the initial matrix (0 iterations) exhibits a random noise pattern, diagonal reinforcement features appear at 200 iterations, and a distinct blocky structure forms at 600 iterations (e.g., ...). The strongly correlated region in rows 3-5 and columns 2-4), after 1000 iterations, the matrix presents a stable cross-grid pattern. Experimental results show that: 1) the projection matrix successfully captures the coupling relationship between the rotational speed characteristics (diagonal) and fault harmonic characteristics (blocky region) in the bearing vibration signal; 2) the dynamic adjustment factor enables the model to quickly strengthen key dimensions (red area expands by 40%) in the middle of training (400 iterations); 3) the final pattern is fundamentally different from the fixed orthogonal pattern of traditional principal component analysis, verifying the effectiveness of dynamically adjusting the projection direction according to data characteristics.

[0025] Furthermore, the encoding part of the depth autoencoder based on transformation projection performs dimensionality reduction processing on the projected device vibration data features, gradually compressing them into a low-dimensional space. The encoding process is implemented through neural network layers, and the output of each layer is:

[0026]

[0027] in, This represents the encoder digit in a depth autoencoder. Layer activation output; This represents the encoder digit in a depth autoencoder. Layer activation output; Use the Sigmoid activation function; This underestimates the adjustment factor.

[0028] Furthermore, to avoid underestimating key information, the underestimation adjustment factor is dynamically updated through error feedback to correct the output. The calculation method is expressed as follows:

[0029]

[0030] In the formula, For the first The adjustment factor was underestimated in the next iteration; For the first The adjustment factor was underestimated in the next iteration; To underestimate the learning rate of the adjustment factor.

[0031] Furthermore, by optimizing the reconstruction error through backpropagation and adjusting the parameters of the transformation projection-based depth autoencoder, the dimensionality reduction mapping and the representation of equipment vibration data features are continuously optimized. The loss function of the transformation projection-based depth autoencoder is expressed as:

[0032]

[0033] In the formula, The weights of the last layer of the decoder in a deep autoencoder. It is an L2 norm.

[0034] Based on the loss function of the depth autoencoder based on transformation projection, the weights and biases of the depth autoencoder are updated, and the update method is expressed as follows:

[0035]

[0036]

[0037] In the formula, This indicates a parameter update; is the learning rate of the deep autoencoder; The loss function of a depth autoencoder with respect to the th The gradient of the layer's weight parameters; The loss function of a depth autoencoder with respect to the th The gradient of the layer's bias parameters; Represents the depth autoencoder's first... Layer weights Represents the depth autoencoder's first... Layer bias.

[0038] Furthermore, the training process of the sparse cooperative random forest algorithm includes,

[0039] Initialize the decision tree structure in the random forest. A sparse cooperation strategy is used when constructing the decision trees. The initial structure of each decision tree in the random forest is defined by the input dimensionality-reduced device vibration dataset, the randomly selected feature set, and the splitting parameters.

[0040] Feature selection is performed based on a sparse cooperation strategy, and a subset of features is selected through an optimization function; this is expressed as:

[0041]

[0042] In the formula, Indicates the first The weights of each feature are related to its sparsity and synergy; The number of features input into the random forest; For the first A feature subset, and, From the dimensionality-reduced equipment vibration dataset The selected feature subset; It is the first The degree of cooperation between the feature subset and the node splitting parameter measures the degree of cooperation between the first feature subset and the node splitting parameter. The contribution of each feature subset to node splitting;

[0043] Feature selection and sparsification are performed. During the construction of each decision tree structure, features are dynamically selected through a sparse cooperation mechanism.

[0044] After the decision tree structure is trained, the classification path is optimized by fuzzy logic, and the boundary data points in the dimensionality-reduced equipment vibration dataset are fuzzy processed.

[0045] After all decision trees are trained, the models are integrated, and the final classification result is obtained through weighted voting.

[0046] Furthermore, the degree of cooperation between the feature subset and the node splitting parameters is determined by the interaction between the features and the parameters, and the calculation method is expressed as follows:

[0047]

[0048] In the formula, For the first The first feature subset One feature; The loss function of random forest is the first... The first feature subset The gradient of each feature It is the first The first feature subset The correlation between each feature and the splitting parameter.

[0049] In a second aspect, a computer-readable storage medium storing instructions, wherein the storage medium stores a computer program or instructions that, when executed by an image processing device, implement a device fault diagnosis method such as one based on deep learning.

[0050] Thirdly, a computer program product, the computer program product comprising: computer program code, which, when the computer program code is executed, causes a processor to execute a deep learning-based device fault diagnosis method.

[0051] The beneficial effects of this invention are:

[0052] 1. In equipment fault diagnosis tasks, a depth autoencoder based on transformation projection is used to reduce the feature dimension of equipment vibration signal data. By dynamically generating an adaptive projection matrix, redundant features in high-dimensional equipment vibration signal data are reduced, overcoming the problem that traditional dimension reduction methods cannot handle the nonlinear relationship of equipment vibration signal data.

[0053] 2. In the equipment fault diagnosis task, a random forest classifier based on sparse cooperation is used to classify the dimensionality-reduced equipment vibration data to achieve equipment fault diagnosis. The selection of node splitting features is optimized through the sparse cooperation mechanism. Combined with the synergistic effect between features, the classification accuracy of fault categories is improved, which solves the limitation of traditional random forest classifiers that rely on single feature selection.

[0054] 3. In equipment fault diagnosis tasks, the classification path is optimized through fuzzy logic during the classification process, and the boundary samples are fuzzy, which improves the fault diagnosis model's ability to distinguish fuzzy region data. Attached Figure Description

[0055] Figure 1 For incremental learning framework;

[0056] Figure 2 This is a flowchart illustrating the device fault diagnosis process based on incremental learning according to an embodiment of the present invention.

[0057] Figure 3 This is an experimental graph comparing the information retention rates of different dimensionality reduction methods in embodiments of the present invention.

[0058] Figure 4This is an experimental analysis diagram of the dynamic adjustment of the adaptive projection matrix during the training process according to an embodiment of the present invention;

[0059] Figure 5 This is an experimental diagram comparing the feature space separation degree of different dimensionality reduction methods in embodiments of the present invention. Detailed Implementation

[0060] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0061] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0062] This invention proposes a deep learning-based equipment fault diagnosis method. It adopts an incremental learning architecture, which can acquire knowledge from old tasks, enabling the equipment fault diagnosis model to learn to diagnose new equipment operating data. At the same time, it retains the knowledge learned from previous tasks, which can avoid retraining the model parameters when new data arrives.

[0063] Figure 1 As shown, based on the incremental learning concept, this invention proposes an incremental learning model training framework. Real-time collected data is input into the learning model, enabling the fault diagnosis model to learn features from the real-time collected data. While maintaining the original classification performance, it integrates and improves the classification ability from the real-time data, thus achieving incremental learning of the fault diagnosis model.

[0064] Incremental learning is a continuous learning process. During training, it is assumed that the model has already learned the previous... One task: In the face of new tasks and its corresponding data In this case, models trained on historical equipment operation data can utilize prior knowledge learned from old tasks to help... The learning process involves acquiring new knowledge and then updating the model using that knowledge. Incremental learning, on the other hand, retains most of the previously learned knowledge while learning new knowledge, thus enabling the equipment fault diagnosis model to perform well on both old and new tasks.

[0065] During the training process, For new unlabeled device operation data collected at different times, a pseudo-labeling method is used to supplement the original labeled data with the unlabeled data. Training is performed, specifically, using an equipment fault diagnosis model on unlabeled data. Make a prediction and output the prediction results Added as a pseudo-label to the original training set This constitutes a new training set. The network is then retrained to gradually learn new equipment fault categories and improve the performance of the equipment fault diagnosis model.

[0066] like Figure 2 As shown, the deep learning-based equipment fault diagnosis method proposed in this invention incorporates the trained fault diagnosis model into an incremental learning framework to perform sample incremental learning.

[0067] Specifically, the training steps for the equipment fault diagnosis model include:

[0068] 1. Import a basic equipment fault diagnosis model obtained after offline training on old data. At the same time, determine whether the new dataset is available at the next moment. The basis for the determination is whether the accuracy of the model in the original data has improved after training on the new equipment operation data. If there is an improvement, the new dataset is available and proceed to the next step. If there is no improvement, the new dataset is not available and training is stopped.

[0069] 2. In the new dataset, select m samples for each equipment failure category and save them. Specifically, obtain feature centers by clustering the data of the same category in the new and old datasets. New data points that fall inside the two feature center points with the distance between the feature centers as the radius of the sphere are selected sample points. Then, add the selected sample points to the same category in the original dataset. Send the saved fused dataset to the new task and proceed to the next step.

[0070] 3. Use the equipment fault diagnosis model to start training the fused dataset after adding samples, generate the model corresponding to the next task, and determine the availability of the dataset corresponding to the next task. If it is available, return to step 2; otherwise, proceed to the next step.

[0071] 4. Terminate training, update the equipment fault diagnosis model parameters, and output the current version of the equipment fault diagnosis model.

[0072] Furthermore, the offline training equipment fault diagnosis model uses an initially constructed training dataset. Specifically, the data originates from a conveyor belt used for transmission. This conveyor belt has bearing components, and failures in these bearing components can easily lead to failures in the conveyor belt itself. Therefore, this invention focuses on fault diagnosis of the bearing components of the conveyor belt. Vibration signals from the conveyor belt are collected using sensors. The format of these vibration signals is primarily time-domain vibration signal data, which can be represented as follows: ,Right now, For the first Vibration signal at a given moment.

[0073] In one specific embodiment, time was acquired. arrive The vibration signal data are used to calculate the mean and variance of the signal, which are expressed as follows:

[0074]

[0075]

[0076] In the formula, This represents the average value of the equipment vibration signal. Let Variance be the variance of the equipment vibration signal. The time length of the vibration signal sequence. For the first Vibration signal at a given moment.

[0077] Furthermore, if the mean and variance of the equipment vibration signal are within a preset threshold range, they can be used as training samples for model training.

[0078] Furthermore, the collected data was labeled, and based on the different types of bearing failures and their locations, it was labeled into the following 5 categories:

[0079] Label 1 - Normal: The bearing is in normal working condition with no obvious faults;

[0080] Label 2 - Inner Ring Failure: The failure occurs in the inner ring of the bearing;

[0081] Label 3 - Outer Ring Failure: The failure occurs in the outer ring of the bearing;

[0082] Tag 4 - Ball bearing failure: The failure occurs in the balls of the bearing;

[0083] Tag 5 - Composite Fault: Multiple types of faults exist simultaneously.

[0084] The collected equipment vibration signals have a high dimension, typically 8192 dimensions, which leads to a large amount of information redundancy during the training of the model.

[0085] This invention employs a depth autoencoder based on transform projection to reduce the dimensionality of equipment vibration data features. Traditional dimensionality reduction methods typically rely on fixed dimensionality reduction strategies, such as principal component analysis or linear discriminant analysis. These methods struggle to handle the complex nonlinear relationships in equipment vibration data. The transform projection method can dynamically adjust the direction and manner of projection according to the different characteristics of the equipment vibration data, enabling the dimensionality reduction process to better adapt to the complex structures that may exist in the equipment vibration data and overcome the limitations of traditional dimensionality reduction methods in handling the nonlinear relationships in equipment vibration data.

[0086] Specifically, the training process for a depth autoencoder based on transformation projection is as follows:

[0087] 1. Initialize the parameters of the depth autoencoder. The initialization method for the weights and biases of the depth autoencoder is as follows:

[0088]

[0089]

[0090] In the formula, Represents the depth autoencoder's first... Layer weights Represents the depth autoencoder's first... Layer bias; This represents a normal distribution with a mean of 0 and a variance of 0.01. This indicates that it conforms to a specific distribution.

[0091] 2. Standardize and preprocess the equipment vibration data input to the depth autoencoder to eliminate the influence of dimensions and scale, as follows:

[0092]

[0093] In the formula, For inputting device vibration data from a depth autoencoder; The mean vector of the device vibration data input from the depth autoencoder; The standard deviation vector of the device vibration data input from the depth autoencoder; This is the standardized data of the device vibration data input from the depth autoencoder.

[0094] 3. Perform spatial transformation projection calculation. The standardized equipment vibration data from the input depth autoencoder is dynamically transformed by the spatial transformation module to generate an adaptive projection matrix, projecting the equipment vibration data into a low-dimensional space, represented as:

[0095]

[0096] In the formula, The adaptive projection matrix is ​​adjusted adaptively during training to ensure that key information of the input device vibration data is preserved as much as possible in the low-dimensional space after projection. Vibration data of the device projected into a low-dimensional space.

[0097] Furthermore, during the adjustment process of the adaptive projection matrix, the learning objective is to maximize the retention of effective information from the equipment vibration data, reduce redundant features, and avoid losing important information. This is specifically achieved through dynamic optimization using the gradient descent method, as follows:

[0098]

[0099] In the formula, For the first The adaptive projection matrix for the next iteration; For the first The adaptive projection matrix for the next iteration; The learning rate is the adaptive projection matrix; It is the gradient of the loss function of the deep autoencoder with respect to the adaptive projection matrix; It is an adaptive adjustment term; The loss function for a deep autoencoder; It is the first The dynamic adjustment factor adjusted by gradient local sensitivity in the next iteration. Preferably, Set as identity matrix, Set to 0.01.

[0100] Furthermore, the dynamic adjustment factor adjusts the learning pace of different device vibration data feature dimensions in each training iteration, enabling the model to adaptively enhance itself when facing device vibration data features with varying noise or high sensitivity. To achieve gradient local sensitivity adjustment, the dynamic adjustment factor is related to the gradient changes of each device vibration data feature dimension and can be dynamically adjusted according to the local errors of the device vibration training data, expressed as:

[0101]

[0102] In the formula, For the first The dynamic adjustment factor after gradient local sensitivity adjustment in the next iteration; To adjust the learning rate of the factor; It is an L2 norm; This is the gradient of the loss function of the deep autoencoder with respect to the adaptive projection matrix. Preferably, Set it to 0.1.

[0103] 4. The projected equipment vibration data features are dimensionality-reduced using the encoding part of a deep autoencoder, gradually compressed into a low-dimensional space. The encoding process is implemented through a series of neural network layers, with the output of each layer being:

[0104]

[0105] in, This represents the encoder digit in a depth autoencoder. Layer activation output; This represents the encoder digit in a depth autoencoder. Layer activation output; Use the Sigmoid activation function; This underestimates the adjustment factor.

[0106] Furthermore, to avoid underestimating key information, the underestimation adjustment factor is dynamically updated through error feedback to correct the output. The calculation method is expressed as follows:

[0107]

[0108] In the formula, For the first The adjustment factor was underestimated in the next iteration; For the first The adjustment factor was underestimated in the next iteration; To underestimate the learning rate of the adjustment factor. Preferably, Set to 0.01.

[0109] 5. By optimizing the reconstruction error through backpropagation and adjusting the parameters of the spatial transformation module and the deep autoencoder, the dimensionality reduction mapping and the representation of equipment vibration data features are continuously optimized. The loss function of the deep autoencoder is expressed as:

[0110]

[0111] In the formula, The weights of the last layer of the decoder in a deep autoencoder. It is an L2 norm.

[0112] Furthermore, based on the loss function of the deep autoencoder, the weights and biases of the deep autoencoder are updated, and the update method is expressed as follows:

[0113]

[0114]

[0115] In the formula, This indicates a parameter update; is the learning rate of the deep autoencoder; The loss function of a depth autoencoder with respect to the th The gradient of the layer's weight parameters; The loss function of a depth autoencoder with respect to the th The gradient of the layer's bias parameters. Preferably, Set to 0.01.

[0116] 6. Repeat the above steps until a preset stopping iteration condition is met, indicating that the model training is complete. In one embodiment, the preset stopping iteration condition is reaching a preset maximum number of iterations; preferably, the preset maximum number of iterations is set to 1000. Figure 3 The study compared the information retention rates of different dimensionality reduction methods to verify their information retention capabilities when processing vibration data from complex equipment. A three-dimensional coordinate system was constructed, with the horizontal axis representing data complexity (dimensionless, higher values ​​indicate more complex data structures), the vertical axis representing noise level (dimensionless, higher values ​​indicate stronger interference), and the vertical axis representing information retention rate (percentage). The performance of principal component analysis, linear discriminant analysis, traditional autoencoders, and the proposed technique were compared. The experiments showed that the proposed technique (green surface) exhibited significant advantages in three key regions. In the low-complexity region (horizontal axis 0-3), the information retention rate remained stable above 95%, and in the high-noise region (vertical axis 7-10), it maintained a retention rate above 85%. Particularly noteworthy was the significant performance spike in the central region (coordinate 5,5), where the retention rate reached 98%, demonstrating the enhancing effect of the dynamic adjustment factor on core features. In contrast, the traditional method's surface showed an overall downward trend, especially when the noise level exceeded 5, with the information retention rate of principal component analysis plummeting to below 65%, proving its difficulty in processing nonlinear vibration data.

[0117] like Figure 4As shown, a comparison of feature space separation is performed to evaluate the inter-class separability of the feature space after dimensionality reduction. The vertical axis represents feature separation (0-1.0, the higher the value, the clearer the class distinction), and the horizontal axis represents four dimensionality reduction methods. This technique (green on the far right) exhibits three major advantages: 1) The median separation is 0.89 (surpassing the traditional autoencoder's 0.24), proving that adaptive projection better preserves discriminative features; 2) The violin-shaped data structure is narrow and tall (width only 0.15), indicating stable results across different experimental batches; 3) The upper edge extends to 0.98, indicating the existence of some ideally separable samples. In contrast, the traditional method: Principal component analysis produces a flat violin-shaped data structure (width 0.35) and has a large number of low-separation samples (the bottom extends to 0.2); although the traditional autoencoder has a median of 0.65, the upper quartile range is 0.3 (25% of samples have a separation below 0.0), verifying the protective effect of the underestimation adjustment factor of this technique on key information, avoiding excessive smoothing of fault features due to feature compression.

[0118] Furthermore, the dimensionality-reduced equipment vibration data is input into the equipment fault diagnosis model for equipment fault diagnosis. The equipment fault diagnosis model adopts a random forest classifier model based on sparse cooperation. On the basis of traditional random forest, a sparse cooperation strategy is adopted to optimize the construction of decision trees by utilizing the synergistic effect between features. The sparse cooperation strategy not only relies on a single feature in the selection of nodes in each decision tree, but also selects the optimal splitting feature through the synergistic effect between features, thereby improving the accuracy of random forest in diagnosing equipment faults.

[0119] Specifically, the training process for the sparse cooperative random forest algorithm is as follows:

[0120] 1. Initialize the decision tree structure in the random forest. Each decision tree is trained by randomly selecting a subset of features from the input dimensionality-reduced equipment vibration data. To improve the effectiveness of feature selection, a sparse cooperation strategy is adopted when constructing the decision tree, so that each node split not only depends on the current feature, but also combines the features of other nodes to optimize the split decision. Let the initial structure of each decision tree in the random forest be the input dimensionality-reduced equipment vibration dataset. Randomly selected feature set And splitting parameters (referring to the splitting conditions of a decision tree) The common definition is expressed as:

[0121]

[0122] In the formula, This represents the initial structure set for each decision tree in a random forest.

[0123] Furthermore, feature selection is performed based on sparse cooperation, and a subset of features is selected through an optimization function, expressed as:

[0124]

[0125] In the formula, Indicates the first The weights of each feature are related to its sparsity and synergy; The number of features input into the random forest; For the first A feature subset, and, From the dimensionality-reduced equipment vibration dataset The selected feature subset; It is the first The degree of cooperation between the feature subset and the node splitting parameter measures the degree of cooperation between the first feature subset and the node splitting parameter. The contribution of each feature subset to node splitting.

[0126] Furthermore, the degree of cooperation between features and node splitting parameters is determined by the interaction between features and parameters, and the calculation method is expressed as follows:

[0127]

[0128] In the formula, For the first The first feature subset One feature; The loss function of random forest is the first... The first feature subset The gradient of each feature It is the first The first feature subset The correlation between each feature and the splitting parameter.

[0129] Furthermore, the feature weights are calculated based on the sparsity of the features and dynamically adjusted according to their sparsity and correlation with the decision tree parameters to ensure that the most informative features are selected, expressed as:

[0130]

[0131] In the formula, Parameters to control the impact of sparsity on weights; The L2 norm measures the sparsity of features. Preferably, Set it to 0.2.

[0132] Furthermore, the correlation of features is calculated using the Pearson correlation coefficient, expressed as:

[0133]

[0134] In the formula, For the first The feature mean in each feature subset; The mean of the splitting parameters; For the first The parameters of the split nodes corresponding to each feature.

[0135] 2. Feature selection and sparsity processing are performed. During the construction of each decision tree, features are dynamically selected through a sparse cooperation mechanism. At each node split, feature selection considers not only information gain but also sparsity and cooperation. The objective function for feature selection is:

[0136]

[0137] In the formula, The objective function selected for the features; For the first The number of features in each feature subset; For the first The first feature subset One feature; It is the first The first feature subset Information gain of each feature The input is the dimensionality-reduced dataset of device vibrations; It is a characteristic sparse factor.

[0138] Furthermore, the information gain of a feature is calculated as follows:

[0139]

[0140] In the formula, The entropy of the input dimensionality-reduced device vibration dataset; The entropy of the input dimensionality-reduced device vibration dataset is the first... The entropy of a feature subset; The entropy of the input dimensionality-reduced device vibration dataset is the first... A subset of features; The preset number of feature subsets; The average value of the features of the input dimensionality-reduced device vibration dataset; The entropy of the input dimensionality-reduced device vibration dataset is the first... The feature average of a subset of features.

[0141] Furthermore, the feature sparsity factor controls the intensity of the sparsity processing. Through a sparse cooperation mechanism, it effectively selects features with high information content, reduces redundancy, and improves the model's generalization ability. The calculation method is expressed as follows:

[0142]

[0143] In the formula, This is a factor that modulates the effect of sparsity.

[0144] 3. After the decision tree training is completed, the classification path is optimized using fuzzy logic. The boundary data points in the dimensionality-reduced equipment vibration dataset are then fuzzified. Specifically, the predicted values ​​of the dimensionality-reduced equipment vibration data samples are calculated using fuzzy membership degrees, as follows:

[0145]

[0146] In the formula, The input is a dimensionality-reduced sample of device vibration data from a random forest. For the first The first feature subset The weights of each feature are training parameters, obtained through gradient descent. The first sample of dimensionality-reduced device vibration data input to the random forest. One feature; It is a fuzzy membership degree, which represents the membership relationship between the dimension-reduced equipment vibration data sample and the category center; For the first The category center of a subset of features.

[0147] Furthermore, the fuzzy membership degree is calculated using a Gaussian kernel function. Fuzzy logic processing can smooth the classification of boundary regions and enhance the model's discriminative ability in fuzzy regions, expressed as:

[0148]

[0149] In the formula, It is the first The Gaussian kernel width of each feature subset controls the expansion range of the membership function. Preferably, Set to 0.01.

[0150] 4. After all decision trees are trained, the models are ensembled, and the final classification result is obtained through weighted voting. The weight of each decision tree is determined by its training accuracy, expressed as:

[0151]

[0152] In the formula, It is the first The weights of each decision tree; It is the first The classification accuracy of the trees on the training set; Indicates the first The structure of a tree; Indicates the first The structure of a tree; The number of decision trees in the random forest.

[0153] Furthermore, the ensemble decision-making of random forests considers the performance of each tree through weighted voting, assigning higher weights to trees with better performance. Additionally, the classification results output by the model undergo boundary adjustment using a fuzzy path enhancement strategy. Through fuzzification, the model's classification accuracy in boundary regions is further improved. The calculation method is expressed as follows:

[0154]

[0155] In the formula, It is the first decision trees for the first The predicted probabilities for each category, including possible predicted categories:

[0156] Label 1 - Normal: The bearing is in normal working condition with no obvious faults;

[0157] Label 2 - Inner Ring Failure: The failure occurs in the inner ring of the bearing;

[0158] Label 3 - Outer Ring Failure: The failure occurs in the outer ring of the bearing;

[0159] Tag 4 - Ball bearing failure: The failure occurs in the balls of the bearing;

[0160] Tag 5 - Composite Fault: Multiple types of faults exist simultaneously.

Claims

1. A device fault diagnosis method based on deep learning, characterized in that, S1. Obtain the old dataset and train the old dataset offline to obtain the initial model; wherein, the old dataset is the equipment vibration data collected by the sensor; S2. Acquire the dataset in real time, whereby the real-time acquired data is referred to as a new dataset; train the initial model using the new dataset; the new dataset is real-time device vibration data collected by sensors; S3. Use the old dataset to test the accuracy of the initial model trained on the new dataset. If the accuracy of the initial model is improved, the new dataset is usable; if the accuracy of the initial model is not improved, the new dataset is not usable. S4. Merge the available new dataset and the old dataset to form a merged dataset; S5. Train the current initial model using the fused dataset to generate the current initial model. Use the new dataset to test the accuracy of the current initial model trained with the fused dataset. If the accuracy of the current initial model is improved, the fused dataset training is usable. Then return to S4. S6. When all data cannot improve the accuracy of the current initial model, terminate the training, obtain the equipment fault diagnosis model parameters, and obtain the equipment fault diagnosis model; the equipment fault diagnosis model adopts a random forest classifier model based on sparse cooperation. S7. Input the dimensionality-reduced equipment vibration data into a sparse cooperative random forest classifier model for equipment fault diagnosis to obtain the diagnosis results. The training process of the sparse cooperative random forest classifier model includes... Initialize the decision tree structure in the random forest. A sparse cooperation strategy is used when constructing the decision trees. The initial structure of each decision tree in the random forest is defined by the input dimensionality-reduced device vibration dataset, the randomly selected feature set, and the splitting parameters. Feature selection and sparsification are performed. During the construction of each decision tree structure, features are dynamically selected through a sparse cooperation mechanism. After the decision tree structure is trained, the classification path is optimized by fuzzy logic, and the boundary data points in the dimensionality-reduced equipment vibration dataset are fuzzy processed. After all decision trees are trained, the models are integrated, and the final classification result is obtained through weighted voting.

2. The device fault diagnosis method based on deep learning according to claim 1, characterized in that, The device vibration data is reduced in dimensionality using a transformation projection-based depth autoencoder. The device vibration data input to the transformation projection-based depth autoencoder is standardized, and an adaptive projection matrix is ​​dynamically generated to project the device vibration data into a low-dimensional space.

3. The deep learning-based equipment fault diagnosis method according to claim 2, characterized in that... The adaptive projection matrix is ​​dynamically optimized during the adjustment process using gradient descent, and is expressed as follows: In the formula, For the first The adaptive projection matrix for the next iteration; For the first The adaptive projection matrix for the next iteration; The learning rate is the adaptive projection matrix; It is the gradient of the loss function of the deep autoencoder with respect to the adaptive projection matrix; It is an adaptive adjustment term; The loss function for a deep autoencoder; It is the first The dynamic adjustment factor after gradient local sensitivity adjustment in the next iteration.

4. The device fault diagnosis method based on deep learning according to claim 2 or 3, characterized in that, It also includes an encoding part based on a transformation projection-based depth autoencoder to reduce the dimensionality of the projected equipment vibration data features, gradually compressing them into a low-dimensional space. The encoding process is implemented through neural network layers, with the output of each layer being: in, This represents the encoder digit in a depth autoencoder. Layer activation output; This represents the encoder digit in a depth autoencoder. Layer activation output; Use the Sigmoid activation function; This underestimates the adjustment factor; Represents the depth autoencoder's first... Layer weights Represents the depth autoencoder's first... Layer bias.

5. The device fault diagnosis method based on deep learning according to claim 4, characterized in that, The underestimation adjustment factor is dynamically updated through error feedback to correct the output.

6. The deep learning-based device fault diagnosis method according to claim 2 or 3, characterized in that, The reconstruction error is optimized through backpropagation, and the parameters of the depth autoencoder based on transformation projection are adjusted.

7. A computer-readable storage medium for storing instructions, characterized in that, The storage medium stores a computer program or instructions that, when executed by an image processing device, implement the method as described in any one of claims 1-6.

8. A computer program product, characterized in that, The computer program product includes: computer program code, which, when executed, causes the processor to perform any one of the methods of claims 1-6.

Images