Equipment fault diagnosis method based on deep learning, storage medium and program product

Abstract

The invention discloses an equipment fault diagnosis method based on deep learning, and the method employs an incremental learning architecture, can obtain knowledge from an old task, enables an equipment fault diagnosis model to learn and diagnose new equipment operation data, reserves the knowledge learned in a previous task, and can avoid the re-training of model parameters when new data comes.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent manufacturing, and in particular relates to a device fault diagnosis method, storage medium, and program product based on deep learning. Background Art

[0002] Fault diagnosis of industrial equipment is crucial for ensuring stable production line operation, particularly for detecting faults in critical components of complex mechanical equipment, such as conveyor belts. Bearings are core components of such equipment, and their operating condition directly impacts their overall performance and service life. Traditional fault diagnosis methods typically rely on offline data analysis and modeling. However, with the advancement of industrial data acquisition and sensing technologies, the acquisition of real-time data has become increasingly common. Effectively utilizing this real-time data to improve the performance of fault diagnosis models has become a key research topic in this field. In practical applications, equipment operating data is characterized by high dimensionality, nonlinearity, and dynamic changes. Traditional methods have significant shortcomings in handling these characteristics. For example, dimensionality reduction techniques struggle to adapt to complex nonlinear characteristics, classification models lack the adaptability to real-time data, and cannot efficiently integrate historical and new data. Furthermore, common methods require retraining the entire model when learning new and old tasks, which wastes computational resources and can lead to the forgetting of knowledge from previous tasks. Therefore, a diagnostic framework with continuous learning is urgently needed that can learn from new data in real time while retaining diagnostic capabilities for old data and efficiently extracting and classifying key features from high-dimensional vibration data.

[0003] The Chinese invention patent with the prior art application number CN202110667841.8 proposes a method and device for diagnosing electromechanical equipment faults, which relates to the technical field of electromechanical equipment fault diagnosis. The method includes the following steps: obtaining the current signal of the secondary side cable of the electromechanical equipment current transformer; 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 the corresponding fault characteristics in the spectrum data and / or time-domain waveform sampling data to perform fault diagnosis. This invention installs a current sensor or transformer on the secondary side cable of the existing electromechanical equipment current transformer to obtain the current signal of the electromechanical equipment for fault diagnosis; the secondary side current is small, and a small-sized current sensor or transformer can be selected, which greatly reduces the cost and is suitable for batch installation; in addition, the electromechanical equipment does not need to be shut down during installation, and the installation method is simple and 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 system includes a diagnostic center and several diagnostic units. Each diagnostic unit includes an industrial computer, an infrared thermal imaging device, a visible light camera, and a communication module located at the target device. The diagnosis includes the following steps: S1: Initializing the equipment fault diagnosis system; S2: The diagnostic unit collects operating data of the current target device; S3: Analyzing the operating data to obtain the fault diagnosis results of the current target device; S4: Sending the fault diagnosis results of the current target device to the diagnostic center. This invention solves the problems of high labor costs, low detection efficiency, and low fault analysis accuracy in the existing technology.

[0005] The existing technology currently has the following problems that still need to be further solved:

[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 the equipment fault diagnosis model.

[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 synergy between features, and are unable to fully utilize the key information in the equipment vibration data, resulting in limited classification accuracy.

[0008] 3. In the equipment fault diagnosis task, the classification problem of boundary areas has not been effectively solved in traditional methods. The equipment fault diagnosis model is prone to misjudgment in these areas, which reduces the accuracy of fault diagnosis. Summary of the Invention

[0009] The present invention provides a device fault diagnosis method, storage medium, and program product based on deep learning. The data collected in real time is input into a learning model, so that the fault diagnosis model can learn the features in the real-time collected data. On the basis of maintaining the original classification effect, the classification ability is improved by integrating from the real-time data, thereby realizing incremental learning of the fault diagnosis model.

[0010] To solve the above problems, the technical solution provided by the present invention is a device fault diagnosis method, storage medium and program product based on deep learning.

[0011] Among them, the equipment fault diagnosis method based on deep learning includes:

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

[0013] S2. Acquire a data set in real time, wherein the real-time acquired data is represented as a new data set; and train the initial model using the new data set; wherein the new data set is real-time equipment vibration data collected by a sensor;

[0014] S3. Using the old data set to test the accuracy of the initial model after training with the new data set, if the accuracy of the initial model is improved, the new data set is usable; if the accuracy of the initial model is not improved, the new data set is unusable;

[0015] S4, fusing the available new dataset and the old dataset to form a fused dataset;

[0016] S5. Use the fused data set to train the current initial model to generate the current initial model, and use the new data set to verify the accuracy of the current initial model trained with the fused data set. If the accuracy of the current initial model is improved, the fused data set training is usable, and then return to S4.

[0017] S6. When all the data cannot improve the accuracy of the current initial model, the training is terminated, the equipment fault diagnosis model parameters are obtained, and the equipment fault diagnosis model is obtained; the equipment fault diagnosis model adopts a random forest classifier model based on sparse collaboration; 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 relationship of complex equipment vibration data. The transformation projection can dynamically adjust the direction and mode 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 relationship of equipment vibration data.

[0018] S7. Input the reduced-dimensional equipment vibration data into a sparse collaborative random forest classifier model to perform equipment fault diagnosis and obtain a diagnosis result.

[0019] Furthermore, a deep autoencoder based on transformation projection is used to reduce the dimension of the device vibration data, the device vibration data input into the deep autoencoder based on transformation projection is standardized, and an adaptive projection matrix is ​​dynamically generated to project the device vibration data into a low-dimensional space, which is expressed as:

[0020]

[0021] Where, The adaptive projection matrix is ​​adaptively adjusted during the training process to ensure that the key information of the input device vibration data is retained as much as possible in the low-dimensional space after projection; Equipment vibration data projected into low-dimensional space; This is the normalized data for the device vibration data input into the deep autoencoder.

[0022] Furthermore, the adaptive projection matrix is ​​dynamically optimized by the gradient descent method during the adjustment process, which is expressed as:

[0023]

[0024] Where, For the Adaptive projection matrix of iterations; For the Adaptive projection matrix of iterations; is the learning rate of the adaptive projection matrix; is the gradient of the loss function of the deep autoencoder with respect to the adaptive projection matrix; It is an adaptive adjustment item; is the loss function of the deep autoencoder; It is The dynamic adjustment factor is adjusted by the gradient local sensitivity in the iteration. During the adjustment process of the adaptive projection matrix, the learning goal is to maximize the retention of effective information of the equipment vibration data, reduce the redundant features of the equipment vibration data and avoid losing important information. Figure 4 As shown in the figure, 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 sets of heat maps are used to show the changes of the 8×8 projection matrix from iteration 0 to 1000. The horizontal axis is the input dimension (original vibration feature dimension), the vertical axis is the projection dimension (feature dimension after dimensionality reduction), and the color mapping represents the weight value (blue negative weight, red positive weight). The experiment shows that the initial matrix (0 iterations) presents a random noise pattern, diagonal reinforcement features appear at 200 iterations, and an obvious block structure is formed at 600 iterations (such as The projection matrix shows a strong correlation between rows 3-5 and columns 2-4). After 1000 iterations, the matrix exhibits a stable cross-grid pattern. Experimental results demonstrate that: 1) the projection matrix successfully captures the coupling relationship between the speed characteristics (diagonal lines) and the fault harmonic characteristics (block-shaped areas) in the bearing vibration signal; 2) the dynamic adjustment factor enables the model to rapidly strengthen key dimensions (red areas expand by 40%) during the mid-training period (400 iterations); and 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 based on data characteristics.

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

[0026]

[0027] in, represents the encoder in the deep autoencoder The activation output of the layer; represents the encoder in the deep autoencoder The activation output of the layer; is the Sigmoid activation function; is the underestimation adjustment factor.

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

[0029]

[0030] Where, For the The underestimate adjustment factor for the iteration; For the The underestimate adjustment factor for the iteration; is the learning rate for the underestimation adjustment factor.

[0031] Furthermore, by back-propagation optimization of the reconstruction error, the parameters of the deep autoencoder based on transformation projection are adjusted to continuously optimize the dimensionality reduction mapping and the feature representation of the equipment vibration data. The loss function of the deep autoencoder based on transformation projection is expressed as:

[0032]

[0033] Where, is the weight of the last layer of the decoder of the deep autoencoder, is the L2 norm.

[0034] According to the loss function of the deep autoencoder based on transformation projection, the weights and biases of the deep autoencoder are updated. The update method is expressed as follows:

[0035]

[0036]

[0037] Where, Indicates parameter update; is the learning rate of the deep autoencoder; is the loss function of the deep autoencoder with respect to the The gradient of the layer's weight parameters; is the loss function of the deep autoencoder with respect to the The gradient of the layer's bias parameters; Represents the deep autoencoder The weights of the layers, Represents the deep autoencoder The bias of the layer.

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

[0039] Initialize the decision tree structure in the random forest. A sparse collaborative strategy is used to construct the decision tree. At the same time, the initial structure of each decision tree in the random forest is defined by the input reduced-dimensional equipment vibration dataset, the randomly selected feature set, and the splitting parameter.

[0040] Feature selection is performed based on the sparse collaborative strategy, and feature subsets are selected by optimizing the function; it can be expressed as:

[0041]

[0042] Where, Indicates the The weight of each feature is related to its sparsity and synergy; is the number of features input to the random forest; For the feature subsets, and The equipment vibration dataset after dimensionality reduction The feature subset selected in It is The degree of cooperation between the feature subset and the node splitting parameter is used to measure the The contribution of a feature subset to the node split;

[0043] Perform feature selection and sparse processing. In the process of building each decision tree structure, the features are dynamically screened through the sparse collaboration mechanism.

[0044] After the decision tree structure training is completed, the classification path is optimized through fuzzy logic, and the boundary data points in the equipment vibration data set after dimensionality reduction are fuzzy processed;

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

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

[0047]

[0048] Where, For the The first in the feature subset Features is the loss function of random forest for the The first in the feature subset The gradient of a feature, It is The first in the feature subset The correlation between the features and the splitting parameter.

[0049] In a second aspect, a computer-readable storage medium storing instructions is provided, wherein the storage medium stores a computer program or instructions. When the computer program or instructions are executed by an image processing device, a device fault diagnosis method based on deep learning is implemented.

[0050] In a third aspect, a computer program product is provided, comprising: a computer program code, which, when executed, causes a processor to execute a device fault diagnosis method based on deep learning.

[0051] Beneficial effects of the present invention:

[0052] 1. In the equipment fault diagnosis task, a deep autoencoder based on transformation projection is used to reduce the feature dimensionality of the equipment vibration signal of the equipment operation data. By dynamically generating an adaptive projection matrix, the redundant features in the high-dimensional equipment vibration signal data are reduced, overcoming the problem that traditional dimensionality reduction methods are difficult to handle the nonlinear relationship of equipment vibration signal data.

[0053] 2. In the equipment fault diagnosis task, a sparse collaboration-based random forest classifier is used to classify the reduced-dimensional equipment vibration data to achieve equipment fault diagnosis. The sparse collaboration mechanism optimizes the selection of node splitting features and combines the synergy between features to improve the classification accuracy of fault categories, thus overcoming the limitation of the traditional random forest classifier that relies on single feature selection.

[0054] 3. In the equipment fault diagnosis task, the classification path is optimized through fuzzy logic during the classification process, and the boundary samples are fuzzy processed, which improves the fault diagnosis model's ability to distinguish fuzzy area data. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is an incremental learning framework;

[0056] Figure 2 This is a workflow diagram for device fault diagnosis based on incremental learning according to an embodiment of the present invention;

[0057] Figure 3 This is a graph comparing the information retention rates of different dimensionality reduction methods according to an embodiment 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 a comparative experimental diagram of the feature space separation of different dimensionality reduction methods in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0061] It should be noted that the terms "first," "second," and the like in the specification and claims of the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having," as well as any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatus.

[0062] The present invention proposes a device fault diagnosis method based on deep learning. It adopts an incremental learning architecture, which can acquire knowledge from old tasks, enabling the device fault diagnosis model to learn to diagnose new device operation data while retaining the knowledge learned in previous tasks, avoiding the need to retrain model parameters when new data arrives.

[0063] Figure 1 As shown in the figure, based on the idea of ​​incremental learning, the present invention proposes a model training framework for incremental learning. The real-time collected data is input into the learning model, so that the fault diagnosis model can learn the features of the real-time collected data. On the basis of maintaining the original classification effect, the classification ability is improved by integrating the real-time data, thereby realizing incremental learning of the fault diagnosis model.

[0064] Incremental learning is a continuous learning process. During the training process, it is assumed that the model has learned the previous Tasks: , in the face of new tasks and its corresponding data When the model is trained with historical equipment operation data, it can use the prior knowledge learned from the old tasks to help Incremental learning is a process of learning new knowledge while retaining most of the previously learned knowledge, which enables the equipment fault diagnosis model to perform well on both old and new tasks.

[0065] During training, For new unlabeled equipment operation data collected at different times, the pseudo-label method is used to use unlabeled data to assist the original labeled data To train, specifically, use the equipment fault diagnosis model to train unlabeled data Make predictions and report the prediction results Added to the original training set as pseudo labels Construct a new training set , and train the network again, so that new equipment fault categories can be gradually learned and the performance of the equipment fault diagnosis model can be improved.

[0066] like Figure 2 As shown in FIG, the device fault diagnosis method based on deep learning proposed in the present invention places the trained fault diagnosis model into the incremental learning framework to complete the function of sample incremental learning.

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

[0068] 1. Import the old data after offline training to obtain a basic equipment fault diagnosis model. At the same time, determine whether the new data set is available at the next moment. The basis for judgment is whether the accuracy of the model in the original data is improved after training with the new equipment operation data. If it is improved, the new data set is available and the next step is entered. If not, the new data set is not available and training is stopped.

[0069] 2. Select m sample data for each equipment fault category in the new data set and save them. Specifically, the data in the same category of the new and old data are clustered to obtain the feature center. The new data points falling inside the two spheres with the distance between the feature centers as the radius are selected as sample points. The selected sample points are then added to the same category of the original data set. The saved fused data set is sent to the new task and enters the next step at the same time.

[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 available, return to step 2; if not, 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 device fault diagnosis model adopts an initially constructed training data set. Specifically, the data comes from a belt conveyor for transmission purposes. The belt conveyor for transmission purposes has bearing components, and the failure of the bearing components is likely to cause the failure of the belt conveyor for transmission purposes. Therefore, the present invention performs fault diagnosis on the bearing components of the belt conveyor for transmission purposes, and collects the vibration signal of the belt conveyor for transmission purposes through a sensor. The format of the vibration signal is mainly vibration signal data in the time domain, which can be expressed as ,Right now, For the Vibration signal at the moment.

[0073] In a specific embodiment, the acquisition time arrive The vibration signal data is used to calculate the mean and variance of the signal, which can be expressed as:

[0074]

[0075]

[0076] Where, is the mean value of the equipment vibration signal, is the variance of the equipment vibration signal, is the time series length of the vibration signal sequence, For the Vibration signal at the moment.

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

[0078] Furthermore, the collected data is labeled and divided into the following five categories based on the different bearing fault types and their locations:

[0079] Label 1-Normal: The bearing is in normal working condition and has 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] Label 4-Ball Failure: The failure occurs in the ball of the bearing;

[0083] Label 5 - Composite fault: Multiple types of faults exist simultaneously.

[0084] The collected equipment vibration signal has a high dimension, usually 8192 dimensions, which brings a lot of information redundancy in the model training process.

[0085] The present invention adopts a deep autoencoder based on transformation projection to reduce the dimensionality of 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 relationship of complex equipment vibration data. The transformation projection can dynamically adjust the direction and mode of projection according to the different characteristics of the equipment vibration data, so that the dimensionality reduction process can better adapt to the complex structures that may exist in the equipment vibration data, overcoming the limitations of traditional dimensionality reduction methods in handling the nonlinear relationship of equipment vibration data.

[0086] Specifically, the training process of the deep autoencoder based on transformation projection is as follows:

[0087] 1. Initialize the parameters of the deep autoencoder. The initialization method of the weight and bias of the deep autoencoder is expressed as:

[0088]

[0089]

[0090] Where, Represents the deep autoencoder The weights of the layers, Represents the deep autoencoder Layer bias; represents a normal distribution with a mean of 0 and a variance of 0.01; Indicates that it obeys a specific distribution.

[0091] 2. The vibration data of the equipment input to the deep autoencoder is standardized and preprocessed to eliminate the influence of dimension and scale, which is expressed as:

[0092]

[0093] Where, The device vibration data is input to the deep autoencoder; is the mean vector of the device vibration data input to the deep autoencoder; is the standard deviation vector of the device vibration data input to the deep autoencoder; This is the normalized data for the device vibration data input into the deep autoencoder.

[0094] 3. Perform spatial transformation projection calculations. After the device vibration data of the input deep autoencoder is standardized, the spatial transformation module dynamically generates an adaptive projection matrix to project the device vibration data into a low-dimensional space, which is expressed as:

[0095]

[0096] Where, The adaptive projection matrix is ​​adaptively adjusted during the training process to ensure that the key information of the input device vibration data is retained as much as possible in the low-dimensional space after projection; Equipment vibration data projected into low-dimensional space.

[0097] Furthermore, during the adjustment process of the adaptive projection matrix, the learning goal is to maximize the retention of effective information of the equipment vibration data, reduce the redundant features of the equipment vibration data, and avoid losing important information. Specifically, dynamic optimization is performed through the gradient descent method, which is expressed as:

[0098]

[0099] Where, For the Adaptive projection matrix of iterations; For the Adaptive projection matrix of iterations; is the learning rate of the adaptive projection matrix; is the gradient of the loss function of the deep autoencoder with respect to the adaptive projection matrix; It is an adaptive adjustment item; is the loss function of the deep autoencoder; It is The dynamic adjustment factor after the gradient local sensitivity adjustment of the iteration. Preferably, Set to the 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, so that the model can adaptively enhance when facing different noise or high-sensitivity device vibration data features. In order to achieve gradient local sensitivity adjustment, the dynamic adjustment factor is related to the gradient change of each device vibration data feature dimension and can be dynamically adjusted according to the local error of the device vibration training data. It is expressed as:

[0101]

[0102] Where, For the The dynamic adjustment factor after the local sensitivity adjustment of the gradient at the iteration; is the learning rate of the adjustment factor; is the L2 norm; is the gradient of the loss function of the deep autoencoder with respect to the adaptive projection matrix. Preferably, Set to 0.1.

[0103] 4. The encoding part of the deep autoencoder performs dimensionality reduction processing on the projected equipment vibration data features, gradually compressing them into a low-dimensional space. The encoding process is implemented through a series of neural network layers, and the output of each layer is:

[0104]

[0105] in, represents the encoder in the deep autoencoder The activation output of the layer; represents the encoder in the deep autoencoder The activation output of the layer; is the Sigmoid activation function; is the underestimation adjustment factor.

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

[0107]

[0108] Where, For the The underestimate adjustment factor for the iteration; For the The underestimate adjustment factor for the iteration; is the learning rate that underestimates the adjustment factor. Preferably, Set to 0.01.

[0109] 5. By back-propagation, the reconstruction error is optimized and the parameters of the spatial transformation module and the deep autoencoder are adjusted to continuously optimize the dimensionality reduction mapping and the feature representation of the equipment vibration data. The loss function of the deep autoencoder is expressed as:

[0110]

[0111] Where, is the weight of the last layer of the decoder of the deep autoencoder, is the L2 norm.

[0112] Furthermore, according to the loss function of the deep autoencoder, the weights and biases of the deep autoencoder are updated. The update method is expressed as:

[0113]

[0114]

[0115] Where, Indicates parameter update; is the learning rate of the deep autoencoder; is the loss function of the deep autoencoder with respect to the The gradient of the layer's weight parameters; is the loss function of the deep autoencoder with respect to the The gradient of the bias parameter of the layer. Preferably, Set to 0.01.

[0116] 6. Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times. Figure 3 As described above, the information retention rates of different dimensionality reduction methods were compared to verify their information retention capabilities when processing complex equipment vibration data. A three-dimensional coordinate system was constructed, with the horizontal axis representing data complexity (dimensionless, larger 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 experiment compared the performance of principal component analysis, linear discriminant analysis, traditional autoencoders, and this technology. The experiment showed that this technology (green surface) exhibited significant advantages in three key areas. In the low-complexity area (horizontal axis 0-3), the information retention rate was stable at above 95%, and in the high-noise area (vertical axis 7-10), the retention rate remained above 85%. In particular, a significant performance bump was formed in the central area (coordinate 5,5), with a retention rate as high as 98%, indicating that the dynamic adjustment factor enhanced the core features. The surface of the traditional method showed an overall sinking trend. In particular, when the noise level exceeded 5, the information retention rate of principal component analysis dropped sharply to below 65%, proving that it was difficult to process nonlinear vibration data.

[0117] like Figure 4As shown, a comparison of feature space separability was performed to assess the inter-class separability of the feature space after dimensionality reduction. The vertical axis represents feature separability (0-1.0, with higher values ​​indicating clearer class distinctions), while the horizontal axis represents four dimensionality reduction methods. This technique (green on the far right) demonstrates three key advantages: 1) a median separability of 0.89 (surpassing the 0.24 of a traditional autoencoder), demonstrating that adaptive projection better preserves discriminative features; 2) a narrow and tall violin body (only 0.15 in width), indicating stable results across different test batches; and 3) an upper margin extending to 0.98, indicating the presence of some ideally separable samples. In contrast, the traditional method exhibits a flat violin body (0.35 in width) and a large number of low-separability samples (with the bottom extending to 0.2). While the traditional autoencoder achieves a median of 0.65, the upper interquartile range spans 0.3 (25% of samples have a separability below 0.0). This demonstrates the effectiveness of this technique's underestimation adjustment factor in protecting critical information, preventing oversmoothing 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 collaboration. On the basis of the traditional random forest, a sparse collaboration strategy is adopted to optimize the construction of the decision tree by utilizing the synergy between features. The sparse collaboration strategy not only relies on a single feature in the node selection of each decision tree, but also selects the optimal split feature through the synergy between features, thereby improving the accuracy of random forest in fault diagnosis of equipment data.

[0119] Specifically, the training process of the random forest algorithm based on sparse collaboration is as follows:

[0120] 1. Initialize the decision tree structure in the random forest. Each decision tree randomly selects a feature subset from the input reduced-dimensional equipment vibration data for training. In order to improve the effectiveness of feature selection, a sparse collaborative strategy is adopted when constructing the decision tree. Each node split not only relies on the current feature, but also combines the features of other nodes to optimize the split decision. The initial structure of each decision tree in the random forest is composed of the input reduced-dimensional equipment vibration data set. , randomly selected feature set and splitting parameters (referring to the splitting conditions of the decision tree) Common definition, expressed as:

[0121]

[0122] Where, Represents the initial structure set of each decision tree in the random forest.

[0123] Furthermore, feature selection is performed based on sparse collaboration, and a feature subset is selected by optimizing the function, which is expressed as:

[0124]

[0125] Where, Indicates the The weight of each feature is related to its sparsity and synergy; is the number of features input to the random forest; For the feature subsets, and It is the equipment vibration dataset after dimensionality reduction The feature subset selected in It is The degree of cooperation between the feature subset and the node splitting parameter is used to measure the The contribution of a 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:

[0127]

[0128] Where, For the The first in the feature subset Features is the loss function of random forest for the The first in the feature subset The gradient of a feature, It is The first in the feature subset The correlation between the features and the splitting parameter.

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

[0130]

[0131] Where, Parameters to control the impact of sparsity on weights; is the L2 norm, which measures the sparsity of features. Preferably, Set to 0.2.

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

[0133]

[0134] Where, For the The feature mean in the feature subset; is the mean of the splitting parameter; For the The parameters of the splitting node corresponding to the feature.

[0135] 2. Perform feature selection and sparsification. During the construction of each decision tree, features are dynamically screened through a sparse collaboration mechanism. At each node split, feature selection not only considers information gain but also combines sparsity and collaboration. The objective function of feature selection is:

[0136]

[0137] Where, The objective function for feature selection; For the The number of features in a feature subset; For the The first in the feature subset Features It is The first in the feature subset The information gain of the features, is the input equipment vibration dataset after dimensionality reduction; is the feature sparsity factor.

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

[0139]

[0140] Where, is the entropy of the input equipment vibration dataset after dimensionality reduction; is the entropy of the input equipment vibration dataset after dimensionality reduction The entropy of a feature subset; is the entropy of the input equipment vibration dataset after dimensionality reduction feature subsets; is the number of preset feature subsets; is the feature average of the input equipment vibration data set after dimensionality reduction; is the entropy of the input equipment vibration dataset after dimensionality reduction The average value of features for a subset of features.

[0141] Furthermore, the feature sparsity factor controls the intensity of the sparsification process. Through the sparse cooperation mechanism, it effectively filters out features with large amounts of information, reduces redundancy, and improves the generalization ability of the model. The calculation method is expressed as:

[0142]

[0143] Where, is a factor to adjust the effect of sparsity.

[0144] 3. After the decision tree training is completed, the classification path is optimized through fuzzy logic, and the boundary data points in the reduced-dimensional equipment vibration data set are fuzzy processed. Specifically, the predicted value of the reduced-dimensional equipment vibration data sample is calculated through fuzzy membership, which is expressed as:

[0145]

[0146] Where, The equipment vibration data samples after dimensionality reduction are input into the random forest; For the The first feature subset The weight of each feature is a training parameter, which is obtained by gradient descent training; The first sample of the equipment vibration data after dimensionality reduction is input into the random forest. Features is the fuzzy membership, which indicates the membership relationship between the equipment vibration data sample after dimensionality reduction and the category center; For the The category center of the feature subset.

[0147] Furthermore, the fuzzy membership is calculated by the Gaussian kernel function. Fuzzy logic processing can smooth the classification of boundary areas and enhance the discrimination ability of the model in fuzzy areas, which is expressed as:

[0148]

[0149] Where, It is The Gaussian kernel width of the feature subset controls the expansion range of the membership function. Set to 0.01.

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

[0151]

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

[0153] Furthermore, the ensemble decision of the random forest considers the performance of each tree through weighted voting, giving higher weights to trees with better performance. In addition, the classification results output by the model are adjusted at the boundary through the fuzzy path enhancement strategy. Through fuzzification processing, the classification accuracy of the model in the boundary area is further improved. The calculation method is expressed as:

[0154]

[0155] Where, It is The decision tree for The predicted probability of each category. Possible predicted categories include:

[0156] Label 1-Normal: The bearing is in normal working condition and has 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] Label 4-Ball Failure: The failure occurs in the ball of the bearing;

[0160] Label 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 an old data set, and train the old data set offline to obtain an initial model; wherein the old data set is equipment vibration data collected by a sensor; S2. Acquire a data set in real time, wherein the real-time acquired data is represented as a new data set; and train the initial model using the new data set; wherein the new data set is real-time equipment vibration data collected by a sensor; S3. Using the old data set to test the accuracy of the initial model after training with the new data set, if the accuracy of the initial model is improved, the new data set is usable; if the accuracy of the initial model is not improved, the new data set is unusable; S4, fusing the available new dataset and the old dataset to form a fused dataset; S5. Use the fused data set to train the current initial model to generate the current initial model, and use the new data set to verify the accuracy of the current initial model trained with the fused data set. If the accuracy of the current initial model is improved, the fused data set training is usable, and then return to S4. S6. When all the 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 collaboration; S7. Input the reduced-dimensional equipment vibration data into a sparse collaborative random forest classifier model to perform equipment fault diagnosis and obtain a diagnosis result.

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

3. The device fault diagnosis method based on deep learning according to claim 2, characterized in that: During the adjustment process, the adaptive projection matrix is ​​dynamically optimized by the gradient descent method, which is expressed as: Where, For the Adaptive projection matrix of the iteration; For the Adaptive projection matrix of the iteration; is the learning rate of the adaptive projection matrix; is the gradient of the loss function of the deep autoencoder with respect to the adaptive projection matrix; It is an adaptive adjustment item; is the loss function of the deep autoencoder; It is The dynamic adjustment factor after adjusting the local sensitivity of the gradient for each iteration.

4. The device fault diagnosis method based on deep learning according to claim 2 or 3, characterized in that: It also includes the encoding part of the deep autoencoder based on transformation projection to reduce the dimension of the projected equipment vibration data features and gradually compress them into a low-dimensional space. The encoding process is implemented through neural network layers, and the output of each layer is: in, represents the encoder in the deep autoencoder The activation output of the layer; represents the encoder in the deep autoencoder The activation output of the layer; is the Sigmoid activation function; is the underestimation adjustment factor; Represents the deep autoencoder The weights of the layers, Represents the deep autoencoder The bias of the layer.

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

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

7. The deep learning-based device fault diagnosis method according to claim 1, characterized in that: The training process of the random forest algorithm based on sparse collaboration includes: Initialize the decision tree structure in the random forest. A sparse collaborative strategy is used to construct the decision tree. At the same time, the initial structure of each decision tree in the random forest is defined by the input reduced-dimensional equipment vibration dataset, the randomly selected feature set, and the splitting parameter. Perform feature selection and sparse processing. In the process of building each decision tree structure, the features are dynamically screened through the sparse collaboration mechanism. After the decision tree structure training is completed, the classification path is optimized through fuzzy logic, and the boundary data points in the equipment vibration data set after dimensionality reduction are fuzzy processed; After all decision trees are trained, the models are integrated and the final classification results are obtained through weighted voting.

8. The device fault diagnosis method based on deep learning according to claim 7, characterized in that: The degree of cooperation between the feature subset and the node splitting parameter is determined by the interaction between the features and the parameters.

9. A computer-readable storage medium storing instructions, characterized in that: The storage medium stores a computer program or instruction, and when the computer program or instruction is executed by the image processing device, the method according to any one of claims 1 to 8 is implemented.

10. A computer program product, characterized in that The computer program product comprises: a computer program code, and when the computer program code is executed, a processor is caused to execute any one of the methods of claims 1-8.

Images