Knowledge-driven Equipment Predictive Maintenance Method and Device

Abstract

The present invention provides a knowledge-driven device predictive maintenance method and device, which relates to the technical field of data processing. The working parameters of industrial devices are subjected to feature extraction through a pre-constructed feature extraction model, and after data dynamic reshaping operation on the target features, dimensionality reduction processing is carried out. Then, a pre-constructed classifier model is used to classify and identify the dimensionality-reduced feature representation to determine the device fault phenomenon indicated by the current working parameters of the industrial device. Among them, the feature extraction model updates network parameters based on the feature non-linear dependence relationship corresponding to a preset training sample set, and can mine more complex feature relationships. The encoding layer of the autoencoder is obtained by sparse optimization based on a non-linear coupling optimization strategy, and can, on the premise of strengthening the internal connection between features, adaptively adjust its dimensionality reduction process according to the changes and patterns of data, maximize the discriminability of feature representation, so as to accurately capture the potential fault phenomenon of the device under the current working parameters.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular, to a method and device for predictive maintenance of equipment based on knowledge-driven. Background Art

[0002] With the improvement of the level of industrial automation and intelligence, the way of equipment maintenance has gradually changed from traditional periodic maintenance to condition-based maintenance, and then developed into predictive maintenance. Predictive maintenance monitors the equipment status in real time, analyzes the equipment data, predicts equipment failures, and thus performs maintenance in advance, reducing unexpected downtime, lowering maintenance costs, and improving production efficiency.

[0003] Among them, actual industrial data often contains rich non-linear structures. Traditional feature extraction methods are based on linear models, and linear models cannot fully capture these complex non-linear dependence relationships, resulting in limited feature representation capabilities. When dealing with high-dimensional and strongly coupled data generated by complex industrial equipment, it is easy to fall into local minima. This leads to the fact that the feature extraction process cannot fully mine the potential non-linear dependence relationships inside the data, thereby affecting the accuracy and robustness of feature representation. Moreover, in the face of dynamically changing industrial data, existing processing methods may lose important information, or the retained features have insufficient discrimination for equipment status under different working conditions, reducing the effect of subsequent classification and recognition. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a method and device for predictive maintenance of equipment based on knowledge-driven, which can process complex linear data of equipment working parameters, accurately identify target features, and accurately capture and identify equipment failures according to current working parameters.

[0005] In a first aspect, an embodiment of the present invention provides a method for predictive maintenance of equipment based on knowledge-driven. The method includes: obtaining the working parameters of a preset industrial equipment, performing data preprocessing on the working parameters to generate a vector to be measured; inputting the vector to be measured into a pre-constructed feature extraction model to extract the target feature corresponding to the vector to be measured; wherein, the feature extraction model updates network parameters based on the feature non-linear dependence relationship corresponding to a preset training sample set; performing a data dynamic reshaping operation on the target feature, and using a preset autoencoder to perform dimensionality reduction processing on the dynamically reshaped target feature to generate a dimensionality reduction feature representation; wherein, the encoding layer of the autoencoder is sparsely optimized based on a non-linear coupling optimization strategy; using a pre-constructed classifier model to perform classification and recognition on the dimensionality reduction feature representation to determine the equipment failure phenomenon indicated by the current working parameters of the industrial equipment.

[0006] In combination with the first aspect, the embodiment of the present invention provides a first implementation manner of the first aspect. Among them, the steps of classifying and identifying the dimensionality-reduced feature representation by using a pre-constructed classifier model to determine the device fault phenomenon indicated by the current working parameters of the industrial device include: inputting the dimensionality-reduced feature representation into the pre-constructed classifier model to output the classification result of the device fault phenomenon; using the preset predictive maintenance knowledge graph system to diagnose the classification result of the device fault phenomenon to determine the potential fault cause corresponding to the classification result of the device fault phenomenon.

[0007] In combination with the first aspect, the embodiment of the present invention provides a second implementation manner of the first aspect. Among them, the steps of dynamically reshaping the data of the target feature include: analyzing the statistical characteristics corresponding to the target feature and calculating the data reshaping factor corresponding to the target feature; using the preset autoencoder to perform preliminary dimensionality reduction processing on the target feature, determining the loss function corresponding to the autoencoder, and determining the weight matrix reshaping factor corresponding to the loss function; calculating the reshaping matrix based on the data reshaping factor and the weight matrix reshaping factor; using the reshaping matrix to dynamically reshape the data of the target feature.

[0008] In combination with the first aspect, the embodiment of the present invention provides a third implementation manner of the first aspect. Among them, the steps of sparsely optimizing the encoding layer of the autoencoder based on the non-linear coupling optimization strategy include: using the preset coupling degree constraint function to determine the non-linear association between the features of the preset training sample set; based on the non-linear association and the preset feature sparsity constraint, determining the loss function of the preset autoencoder; based on the loss function and the preset momentum factor, performing backpropagation update on the weight matrix of the autoencoder to optimize the encoding layer of the autoencoder.

[0009] In combination with the first aspect, the embodiment of the present invention provides a fourth implementation manner of the first aspect. Among them, the construction method of the classifier model includes: inputting the preset training sample set into the preset support vector machine, and using the chaos theory algorithm to adjust the support vector machine parameters of the support vector machine; determining the Lagrange multiplier corresponding to the training sample set, and based on the Lagrange multiplier, adjusting the parameters of the kernel function of the support vector machine; until the support vector machine meets the preset training conditions, constructing a classifier model based on the support vector machine.

[0010] In combination with the first aspect, the embodiment of the present invention provides a fifth implementation manner of the first aspect. Among them, the steps of using the chaos theory algorithm to adjust the support vector machine parameters of the support vector machine include: generating random perturbations using the chaos theory; using the preset non-convex optimization algorithm to adjust the support vector machine parameters of the support vector machine based on the random perturbations.

[0011] In combination with the first aspect, an embodiment of the present invention provides a sixth implementation manner of the first aspect. The method for constructing a feature extraction model includes: obtaining a preset training sample set, performing non-linear decomposition on the training sample set by using a non-linear feature dual decomposition algorithm to obtain a plurality of dual components; inputting the plurality of dual components into a preset neural network for forward propagation, and determining the loss function corresponding to the neural network; updating the network parameters of the neural network based on the loss function and the feature non-linear dependence relationship indicated by the dual components; until the neural network meets the preset iteration condition, constructing a feature extraction model based on the neural network.

[0012] In combination with the first aspect, an embodiment of the present invention provides a seventh implementation manner of the first aspect. The method for constructing a preset training sample set includes: obtaining the working parameters pre-stored in a preset industrial production device, annotating the working parameters to construct an initial sample set; generating random noise corresponding to the initial sample set, and obtaining the current operation data of the industrial production device; determining the coupled input corresponding to the industrial production device based on the random noise and the current operation data; and constructing a training sample set by performing data augmentation on the initial sample set based on the coupled input.

[0013] In combination with the first aspect, an embodiment of the present invention provides an eighth implementation manner of the first aspect. The step of constructing a training sample set by performing data augmentation on the initial sample set based on the coupled input includes: inputting the coupled input into a pre-trained generative adversarial network, and using the generative adversarial network to generate augmented samples corresponding to the coupled input; merging the augmented samples with the initial sample set to construct a training sample set; wherein, the generative adversarial network updates the generator by performing an exponential calculation on the output of the discriminator.

[0014] In the second aspect, an embodiment of the present invention provides a knowledge-driven device predictive maintenance device. The device includes: a data acquisition module, configured to acquire the working parameters of a preset industrial device, perform data preprocessing on the working parameters to generate a vector to be measured; a feature extraction module, configured to input the vector to be measured into a pre-constructed feature extraction model to extract the target feature corresponding to the vector to be measured; wherein, the feature extraction model updates the network parameters based on the feature non-linear dependence relationship corresponding to the preset training sample set; a data processing module, configured to perform a data dynamic reshaping operation on the target feature, and perform dimensionality reduction processing on the dynamically reshaped target feature by using a preset autoencoder to generate a dimensionality-reduced feature representation; wherein, the encoding layer of the autoencoder is sparsely optimized based on a non-linear coupling optimization strategy; an execution module, configured to perform classification and recognition on the dimensionality-reduced feature representation by using a pre-constructed classifier model to determine the device fault phenomenon indicated by the current working parameters of the industrial device.

[0015] The embodiments of the present invention bring the following beneficial effects: A method and device for knowledge-driven predictive maintenance of equipment provided by the present invention can mine more complex feature relationships in a high-dimensional space, improve the accuracy of feature extraction. And on the premise of strengthening the internal connection between features, it can adaptively adjust its dimensionality reduction process according to the changes and patterns of data, effectively avoid information loss, maximize the discriminability of feature representation, and accurately capture potential fault phenomena of the equipment under current working parameters.

[0016] Other features and advantages of the present invention will be described in the following specification, and, in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention are realized and obtained by the structures specifically pointed out in the specification and the drawings.

[0017] To make the above objectives, features, and advantages of the present invention more obvious and understandable, the following specific preferred embodiments are given, and in conjunction with the accompanying drawings, the detailed description is as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for use in the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0019] Figure 1 It is a flowchart of a method for knowledge-driven predictive maintenance of equipment provided by an embodiment of the present invention;

[0020] Figure 2 It is a flowchart of another method for knowledge-driven predictive maintenance of equipment provided by an embodiment of the present invention;

[0021] Figure 3 It is a flowchart of a method for constructing a training sample set provided by an embodiment of the present invention;

[0022] Figure 4 It is a schematic structural diagram of a device for knowledge-driven predictive maintenance of equipment provided by an embodiment of the present invention;

[0023] Figure 5 It is a schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following describes the implementation manners of the present disclosure through specific specific examples. Those skilled in the art can easily understand other advantages and effects of the present disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all the embodiments. The present disclosure can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present disclosure without creative efforts belong to the scope of protection of the present disclosure.

[0025] It should be noted that the following describes various aspects of the embodiments within the scope of the appended claims. It should be apparent that the aspects described in the present invention can be embodied in a wide variety of forms, and any specific structure and / or function described in the present invention is illustrative only. Based on the present disclosure, those skilled in the art should understand that one aspect described in the present invention can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described in the present invention can be used to implement the device and / or practice the method. In addition, this device can be implemented and this method can be practiced using other structures and / or functions in addition to one or more of the aspects described in the present invention.

[0026] It should also be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present disclosure in a schematic manner. The diagrams only show the components related to the present disclosure, rather than being drawn according to the number, shape, and size of the components in actual implementation. The type, quantity, and ratio of each component in actual implementation can be arbitrarily changed, and the component layout type may also be more complex. In addition, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0027] The embodiments of the present invention provide a knowledge-driven device predictive maintenance method and device, which can process complex linear data of device working parameters, accurately identify target features, and accurately capture and identify device faults based on the current working parameters.

[0028] For the convenience of understanding, first, a knowledge-driven device predictive maintenance method provided by the embodiments of the present invention will be described in detail. Figure 1 The flowchart of a knowledge-driven device predictive maintenance method provided by the embodiments of the present invention is shown. Refer to Figure 1, the method includes the following steps:

[0029] Step S102, obtain the working parameters of a preset industrial device, perform data preprocessing on the working parameters, and generate a vector to be measured.

[0030] In the embodiments of the present invention, data is collected from industrial production devices, including but not limited to mechanical devices, electrical devices, sensor devices, etc., for predictive maintenance of the devices. Among them, data is collected through interfaces such as monitoring systems, sensor networks, and SCADA systems (supervisory control and data acquisition) related to the devices. Among them, by obtaining various working parameters monitored in real time by sensors on the preset industrial device, the current state of the device is determined. Among them, the device status information monitored by the sensors can be uploaded to the central repository through the data acquisition system for data analysis. Historical data is extracted from the device logs, maintenance records, or historical monitoring data, and these data are transmitted through the network and stored in a distributed database or cloud storage; the data storage adopts a structured format, and all device operation data is stored in a unified format. Each piece of data consists of multiple attributes (such as temperature, pressure, vibration, etc.), and each attribute corresponds to a vector component. The data storage format is CSV format.

[0031] It should be noted that this embodiment is only to illustrate a data format and type of the present invention. In actual applications, the number of data attributes is usually more than 10, and the number of data attributes may reach dozens or even hundreds. Specifically, it may include extracting time-domain features and frequency-domain features from the collected time-series data. For example, the average amplitude, amplitude variance, root mean square amplitude, peak ratio, kurtosis, skewness, maximum amplitude, minimum amplitude, main frequency, etc. of vibration data.

[0032] Step S104, input the vector to be measured into a pre-constructed feature extraction model, and extract the target features corresponding to the vector to be measured.

[0033] Extract data features using a pre-trained feature extraction model, and mine important features from the original data that can reflect changes in the device state. The model will output a new set of feature vectors, which can more accurately describe the current state of the device. Among them, the model in the embodiment of the present invention is established based on a large number of historical samples. In specific implementation, the model updates network parameters based on the non-linear dependence relationship of features corresponding to a preset training sample set. In the training process of traditional neural networks, a symmetric topological structure is adopted, that is, the connection method between neurons is symmetric. Although this structure is simple and easy to implement, when dealing with complex non-linear data, it is easy to fall into local minima, resulting in poor training effects. In the embodiment of the present invention, network parameters are updated based on the non-linear dependence relationship of features. By non-linearly decomposing the data input into the neural network, the data is decomposed into multiple dual components, so that the connection method of each layer of the neural network is no longer simply symmetric, and then the network can mine more complex feature relationships in the high-dimensional space, thereby improving the accuracy of feature extraction.

[0034] Step S106, perform a data dynamic reshaping operation on the target features, and use a preset autoencoder to perform dimensionality reduction processing on the dynamically reshaped target features to generate a dimensionality-reduced feature representation.

[0035] Traditional data dimensionality reduction processing maps input data to a low-dimensional space, but when dealing with complex and non-linear data, it often has the defect of being unable to effectively extract the most representative features. In the embodiment of the present invention, the input features are first dynamically adjusted according to the feature distribution of the input data, so as to adaptively adjust its dimensionality reduction process according to the changes and patterns of the data, effectively avoiding information loss and maximizing the discriminability of feature representation.

[0036] The present invention uses an autoencoder for data dimensionality reduction processing. An autoencoder is a neural network architecture that can reduce the dimension of the feature space by learning the compressed representation of the input data. Among them, the encoding layer of the autoencoder in the embodiment of the present invention is sparsely optimized based on a non-linear coupling optimization strategy. The non-linear coupling optimization strategy is to handle the interaction between multiple variables in a complex system. In one embodiment, by adopting the constraint conditions of the coupling system, there is a certain non-linear correlation between each encoding neuron, strengthening the internal connection between features. By optimizing the coupling degree, the neural network can better preserve the original information of the data during the dimensionality reduction process and eliminate noise interference at the same time.

[0037] Step S108, use a pre-constructed classifier model to classify and identify the dimensionality-reduced feature representation, and determine the device fault phenomenon indicated by the current working parameters of the industrial device.

[0038] After the collected working parameters are processed as described above, they can be classified and recognized by a pre-constructed classifier model (such as a support vector machine, random forest, etc.), and it is determined whether the current working parameters of the device indicate a certain specific fault phenomenon according to the classification result. In one implementation, the dimensionality-reduced feature representation can be input into the pre-constructed classifier model to output the classification result of the device fault phenomenon, and the classification categories are different device fault phenomena, such as "hydraulic pressure too high", "knife chain running out of position", etc.

[0039] Furthermore, a preset predictive maintenance knowledge graph system can be used to diagnose the classification result of the device fault phenomenon to determine the potential fault causes corresponding to the classification result of the device fault phenomenon. The diagnosis methods include: 1 - Fault cause reasoning: infer potential fault causes from the input fault phenomenon. For example, if the fault phenomenon is "hydraulic pressure too high", the system may infer that it is "the output pressure of the hydraulic pump is set too high"; 2 - Concurrent fault reasoning: analyze the association between multiple fault phenomena. For example, "the knife chain running out of position" may cause "the machine tool to alarm automatically"; 3 - Operation-induced fault reasoning: certain operations may directly or indirectly cause fault phenomena. For example, "starting the hydraulic system" may cause "hydraulic pressure too high"; 4 - Alarm information association: the system can associate the fault phenomenon with the alarm information to help users identify the meaning of the fault alarm code. If the current knowledge base cannot provide an accurate diagnosis or solution, the system will automatically perform web crawling to obtain the latest solution and add it to the knowledge base, so as to achieve self-learning and continuous optimization of the knowledge base.

[0040] Among them, in the embodiment of the present invention, a device predictive maintenance knowledge graph system is constructed to determine the device fault phenomenon and the possible causes. The user inputs parameters related to the device (such as fault phenomenon, fault code, abnormal parameter, fault description, etc.), and the system infers the fault cause and repair solution through the knowledge graph and provides real-time diagnosis.

[0041] Specifically, the device predictive maintenance knowledge graph system includes the following functions: 1 - Fault diagnosis function: after the user provides fault information, the system reasons through the knowledge graph and gives specific fault causes and repair solutions; 2 - Question and answer function: the system answers the user's questions about fault causes, operation errors, alarm information, etc. The four types of questions supported include: the phenomenon caused by a certain fault cause, the error encountered when performing a certain operation, the faults that often occur in a certain part, and the meaning of a certain fault alarm.

[0042] In one embodiment, the above knowledge graph can be constructed based on an ontology model. The key entities and relationships are stored in the form of triples, and each triple consists of "subject - predicate - object". Among them, the entity nodes are: 1 - Fault phenomenon: The specific manifestation when the device fails, which is the key entry for fault diagnosis; 2 - Fault cause: The root cause leading to the fault phenomenon; 3 - Operation steps: The steps executed by the user during the device operation process; 4 - Alarm information: The alarm code or information displayed when the device fails; 5 - Fault location: The component in the device where the fault occurs; 6 - Solution: The fault repair method feedback by the user. The relationship nodes are: 1 - Operation triggers phenomenon: A certain operation may trigger a specific fault phenomenon; 2 - Phenomenon causes reason: The relationship between a certain fault phenomenon and its underlying reason; 3 - Phenomenon co - occurrence relationship: The co - occurrence relationship between different fault phenomena; 4 - Phenomenon and location relationship: The association between the fault phenomenon and the specific location or component; 5 - Phenomenon and alarm information relationship: The association between the fault phenomenon and the alarm information. 6 - Fault cause and solution: The relationship between the fault cause and the repair measure. Among them, the above - mentioned fault phenomenon is predicted by the classifier model of the embodiment of the present invention for the processed data.

[0043] Further, on the basis of the above - mentioned embodiment, the embodiment of the present invention also provides another knowledge - driven device predictive maintenance method, Figure 2 which shows the flowchart of the embodiment of the present invention. Referring to Figure 2 , it includes the following steps:

[0044] Step S202, obtain the working parameters of the preset industrial device, perform data pre - processing on the working parameters, and generate a vector to be measured.

[0045] Step S204, input the vector to be measured into the pre - constructed feature extraction model, and extract the target feature corresponding to the vector to be measured.

[0046] The embodiment of the present invention mainly explains the construction method of the feature extraction model. The present invention uses a 5 - layer fully - connected neural network for feature extraction. Through the following construction method, it solves the problem that in some neural network structures of the prior art, gradient vanishing, gradient explosion or getting stuck in local optimal solutions may occur, affecting the training stability and model performance. Among them, the steps of updating network parameters based on the feature non - linear dependence relationship corresponding to the preset training sample set to construct the feature extraction model are as follows:

[0047] 1) Obtain the preset training sample set, and perform non - linear decomposition on the training sample set using the non - linear feature dual decomposition algorithm to obtain multiple dual components.

[0048] Among them, the embodiment of the present invention preprocesses the data input into the neural network by using the non-linear feature dual decomposition algorithm. By non-linearly decomposing the data, the data is decomposed into multiple dual components, providing a richer feature input for the neural network. After the input data of each layer passes through the non-linear mapping, it is transmitted to the next layer for further processing. Specifically, the non-linear feature dual decomposition realizes the non-linear decomposition of the input data through local decomposition, which is expressed as:

[0049]

[0050] In the formula, Ke is the number of components of the dual decomposition, is the weight coefficient of the k-th component, is the transformation matrix of the corresponding component. Preferably, Ke is set to 3.

[0051] Furthermore, the non-linear feature dual decomposition is separately supervised and trained, which is constrained by the regularization loss function to avoid the feature mapping being too complex, ensure that the model can generalize, and prevent overfitting. The calculation method of the regularization loss function is expressed as:

[0052]

[0053] In the formula, represents the Frobenius norm of the matrix , reflecting the complexity of the model parameters; is the Frobenius norm.

[0054] 2) Input multiple dual components into a preset neural network for forward propagation, and determine the loss function corresponding to the neural network.

[0055] Suppose the data input into the neural network is , n c is the dimension of the input feature, represents the dimensional space, and the weight matrix of the neural network is , m c is the number of neurons in the network hidden layer. In order to process the non-linear mapping of each layer, the initialization method of the neural network is expressed as:

[0056]

[0057] In the formula, is the initial weight of the neural network; represents a normal distribution with a mean of 0 and a variance of ; is the normal distribution; is the variance of the neural network parameter initialization. Preferably, is set to 0.01.

[0058] The input data is linearly transformed through the weight matrix and bias term of each layer of the neural network, and then non-linearly mapped through the activation function. The forward propagation method is expressed as:

[0059]

[0060]

[0061] In the formula, is the output of the (l-1)-th layer of the neural network, is the weight matrix of the l-th layer of the neural network, is the bias term of the l-th layer of the neural network, is the Sigmoid activation function; is the linear output of the l-th layer of the neural network; is the non-linear output of the l-th layer of the neural network.

[0062] 3) Update the network parameters of the neural network based on the loss function and the feature non-linear dependence relationship indicated by the dual components.

[0063] Among them, in each embodiment of the present invention, after each gradient calculation of the loss function, the weights of the network are updated according to the optimization algorithm. The update method is expressed as:

[0064]

[0065]

[0066] In the formula, is the parameter update operation; is the bias gradient of the -th layer of the neural network; is the regularization coefficient of the neural network; is the L2 regularization. Among them, .

[0067] Among them, in each embodiment of the present invention, the gradient of the loss function of the neural network is calculated through the backpropagation algorithm, and the weights of the network are updated according to the gradient. The traditional neural network uses the gradient descent method to update the weights. When calculating the gradient of the loss function in the present invention, the error is propagated through each layer of the neural network by the backpropagation algorithm. The calculation method of the partial derivative of the loss function of the neural network with respect to the weight is expressed as:

[0068]

[0069] Similarly, the calculation method of the partial derivative of the loss function of the neural network with respect to the bias is expressed as:

[0070]

[0071] In the formula, is the output of the (l - 1)-th layer of the neural network; is the non-linear output of the l-th layer of the neural network; is the linear output of the l-th layer of the neural network.

[0072] Furthermore, when considering the error term propagation, in addition to considering the single gradient information, the non-linear dependence relationship between features is also utilized to calculate the partial derivative of the loss function of the neural network with respect to the non-linear output, which is expressed as:

[0073]

[0074] In the formula, is the partial derivative of the loss function of the neural network with respect to the final output layer (i.e., the ls-th layer); is the non-linear output of the k-th layer of the neural network; is the linear output of the k-th layer of the neural network.

[0075] Among them, in order to improve the training stability of the model and avoid falling into local minima when tuning the neural network parameters. When the input data propagates forward through each layer in the neural network, the present invention adopts a gradient optimization strategy for the non-symmetric topological stable points, and uses the historical gradient information to accelerate convergence and improve stability. The calculation method of the update amount of the weight parameters of the neural network is expressed as:

[0076]

[0077]

[0078] In the formula, is the update amount of the weight of the l-th layer of the neural network; is the update amount of the weight of the (l - 1)-th layer of the neural network; is the learning rate of the neural network, is the weight gradient of the l-th layer of the neural network; is the weight gradient of the i-th layer of the neural network; is the loss function of the neural network; is a parameter dynamically adjusted by the neural network, controlling the influence of the weight update of the previous layer of the neural network; is the sum of the squares of the historical gradients, representing the magnitude of the gradient change. By performing weighted averaging on the gradients, the model can obtain a smoother update process at different levels, avoiding the problems of local minima and oscillations; ls is the total number of layers of the neural network, preset to 5; is a small constant to prevent division by zero. Preferably, is set to 0.001, Set to 0.2.

[0079] 4) Until the neural network meets the preset iteration conditions, construct a feature extraction model based on the neural network.

[0080] Repeat the above steps iteratively until the preset stop iteration condition is met, which indicates that the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0081] In summary, the present invention uses a neural network based on non - linear feature dual decomposition. By non - linear decomposition, the data is decomposed into multiple dual components, improving the ability of the neural network to mine complex feature relationships in high - dimensional space, solving the problem that traditional neural networks are prone to falling into local minima resulting in poor training effects, and being able to improve the accuracy of feature extraction.

[0082] Furthermore, data reshaping is performed through the following steps S206 - S212.

[0083] Step S206, analyze the statistical characteristics corresponding to the target feature, and calculate the data reshaping factor corresponding to the target feature.

[0084] Step S208, use a preset auto - encoder to perform preliminary dimensionality reduction on the target feature, determine the loss function corresponding to the auto - encoder, and determine the weight matrix reshaping factor corresponding to the loss function.

[0085] Step S210, calculate the reshaping matrix based on the data reshaping factor and the weight matrix reshaping factor.

[0086] Step S212, perform data dynamic reshaping on the target feature using the reshaping matrix.

[0087] In specific implementation, the embodiment of the present invention combines data reshaping with an auto - encoder. The data reshaping operation can also be used to train the auto - encoder to adjust the structure and feature dimensions of the encoding layer according to the current characteristics of the data, ensuring that the network can be optimized for a specific data set in each training cycle, and avoiding over - fitting or under - fitting phenomena that may exist in the static dimensionality reduction structure of traditional auto - encoders.

[0088] Specifically, let the new representation of the input data after dynamic reshaping be , and the reshaping process is expressed as:

[0089]

[0090] In the formula, is the dynamic reshaping matrix, whose function is to dynamically adjust the input features according to the feature distribution of the input data; is the reshaped input data; is a feedback signal that reflects the optimization effect of the current dimensionality reduction path.

[0091] Among them, in the embodiment of the present invention, a reshaping matrix is used for data reshaping. The reshaping matrix consists of two parts, which not only depends on the characteristics of the original data, but also is dynamically updated according to the current state of the network (autoencoder). That is, by inputting the data into the autoencoder to determine the data characteristics, and comprehensively calculating the state of the autoencoder and the data characteristics to determine the corresponding reshaped data. It is expressed as:

[0092]

[0093] In the formula, is the reshaping factor based on the data input to the autoencoder, which determines the main distribution of the data; is the reshaping factor based on the weight matrix of the encoder, which adjusts the feature representation after dimensionality reduction.

[0094] The reshaping factor based on the data input to the autoencoder determines the reshaping weight of each feature by analyzing the statistical characteristics (such as mean and variance) of the input data. The calculation method is expressed as:

[0095]

[0096] In the formula, is the input data is the mean of is the first dynamic adjustment factor of the reshaping matrix; is the second dynamic adjustment factor of the reshaping matrix; is the second derivative of the loss function of the autoencoder with respect to the input data, which measures the sensitivity of the change of the input data to the loss, so as to adjust the reshaping factor; is the loss function of the autoencoder. Preferably, α r is set to 0.1, and β r is set to 0.3.

[0097] Furthermore, the reshaping factor based on the weight matrix of the encoder is dynamically adjusted according to the change of the current weight during the training process of the network. The calculation method is expressed as:

[0098]

[0099] In the formula, is the gradient of the loss function of the autoencoder with respect to the network weight, which represents the influence of the current weight on the reshaping matrix; is the increment of the autoencoder, which represents the rate of weight change during the training process; is the third dynamic adjustment factor of the reshaping matrix; is the fourth dynamic adjustment factor of the reshaping matrix.

[0100] Furthermore, the above feedback signal reflects the optimization effect of the current dimensionality reduction path. When the model shows overfitting or underfitting on some samples, by introducing a non-linear coupling feedback mechanism, the network will automatically reconfigure its dimensionality reduction structure. The configuration process can be determined based on the training of the autoencoder. By adjusting the parameters to adapt to the actual distribution of the data, it not only makes the training process more flexible but also effectively avoids the problems of over-simplification or information loss existing in traditional dimensionality reduction algorithms. Specifically, the feedback signal dynamically adjusts the training path (dimensionality reduction path) according to the weight matrix of the next iteration of the autoencoder, and the calculation method is expressed as:

[0101]

[0102] In the formula, is the gradient of the loss function with respect to the reshaping matrix; is the feedback adjustment factor, which is used to balance the gradient update and the contribution of the feedback signal; is the feedback signal; is the weight of the feedback signal. Preferably, is set to 1, is set to 0.05.

[0103] In summary, based on the above dynamic reshaping operation, the autoencoder can perform data dimensionality reduction by analyzing the distribution characteristics and complexity of the input data, can solve the overfitting or underfitting phenomena that may exist in the traditional autoencoder with a static dimensionality reduction structure, effectively avoid information loss, and at the same time maximize the discriminability of feature representation.

[0104] Step S214: Use a preset autoencoder to perform dimensionality reduction on the dynamically reshaped target features to generate a dimensionality-reduced feature representation.

[0105] Furthermore, the encoding layer of the sparse optimization autoencoder based on the non-linear coupling optimization strategy is described in the steps of training the autoencoder:

[0106] 1) Use a preset coupling degree constraint function to determine the non-linear association between the features of a preset training sample set.

[0107] The non-linear coupling degree constraint is the optimization of the non-linear correlation between feature representations. By constructing complex coupling relationships through high-order interaction terms, the calculation method of the coupling degree constraint function is expressed as:

[0108]

[0109] In the formula, is the i-th element of the low-dimensional feature representation, is the j-th element of the low-dimensional feature representation; Kas is the dimension of the low-dimensional feature representation; as a feature and the coupling weight between them, which characterizes the non - linear dependence between the two; is the bandwidth parameter in the Gaussian kernel function, which controls the interaction range between features. Preferably, is set to 0.1.

[0110] 2) Determine the loss function of the preset auto - encoder based on non - linear correlation and the preset feature sparsity constraint.

[0111] Among them, in the training process, the parameters of the auto - encoder are first initialized. In one embodiment, the parameter initialization method of the auto - encoder is random initialization and follows a normal distribution with a mean of 0 and a variance of the identity matrix. Further, let the data input to the auto - encoder be , where N is the number of augmented samples and D is the dimension of the data features input to the auto - encoder.

[0112] The auto - encoder network consists of two parts: an encoder and a decoder. The encoder part maps the input data to a low - dimensional feature space, and the decoder part remaps the low - dimensional features back to the original space. The mapping method of the encoder is expressed as:

[0113]

[0114] In the formula, is the mapping function of the encoder, is the weight matrix of the encoder, and Z r is the low - dimensional feature representation. Preferably, the ReLU activation function is selected for the encoder mapping function.

[0115] The calculation method of the loss function of the auto - encoder is expressed as:

[0116]

[0117] In the formula, is the loss function of the auto - encoder. is the L2 norm; characterizes the reconstruction error and measures the difference between the feature mapping after dimensionality reduction and the original data; is the mapping function of the decoder; is the L1 norm.

[0118] characterizes the feature sparsity constraint, which promotes the feature representation to be as sparse as possible and reduces redundancy; is the coupling degree constraint function, aiming to maximize the non - linear correlation between features. is the weight coefficient of the reconstruction error, is the weight coefficient of the sparsity constraint, is the weight coefficient for the coupling degree constraint. Preferably, is set to 0.3, is set to 0.3.

[0119] 3) Based on the loss function and the preset momentum factor, perform backpropagation update on the weight matrix of the autoencoder to optimize the encoding layer of the autoencoder.

[0120] The autoencoder calculates the reconstruction error through forward propagation and updates the parameters through backpropagation. Different from the traditional autoencoder, the present invention dynamically adjusts the optimization strategy in each round of iteration to ensure that the network can automatically select the optimal dimensionality reduction path according to the actual performance of the data. The weight update method of the autoencoder is expressed as:

[0121]

[0122] In the formula, is the weight matrix of the autoencoder at the t-th iteration; is the weight matrix of the autoencoder at the (t + 1)-th iteration; is the learning rate of the autoencoder; is the gradient of the loss function of the autoencoder with respect to the weight matrix; is the momentum factor of the autoencoder, which is used to control the smoothness of the gradient update; is the second-order gradient of the loss function with respect to the weight matrix. Through the second-order gradient term, it can converge more efficiently and avoid the oscillation or local minimum problems that may occur in traditional gradient descent. Preferably, is set to 0.01, is set to 0.3. In summary, the sparse optimization of the encoding layer combined with the non-linear coupling optimization strategy is realized.

[0123] The weight coefficient of the reconstruction error in the embodiment of the present invention adopts an adaptive adjustment mechanism to dynamically adjust the weight of the reconstruction error during the training process, so that the contributions in different training stages can be balanced, thereby effectively avoiding overfitting or underfitting. The adjustment method is expressed as:

[0124]

[0125] In the formula, is the variance of the loss function, indicating the volatility of the current loss function between different training cycles. Repeat the above steps iteratively until the preset stop iteration condition is satisfied, which means the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0126] Step S216: Use the pre-constructed classifier model to classify and identify the dimensionality-reduced feature representation, and determine the device fault phenomenon indicated by the current working parameters of the industrial device.

[0127] The present invention uses the support vector machine algorithm to classify the dimensionality-reduced data. Different from the traditional support vector machine algorithm, the embodiments of the present invention introduce non-convex optimization to solve complex optimization problems in order to construct a classifier model. By optimizing when the objective function is not a convex function, the problem of difficultly finding the global optimal solution when there are multiple local extreme points is solved. Among them, the classification algorithms in the prior art cannot effectively apply non-convex optimization. Even if the extreme learning machine is used to map the input data to a non-linear high-dimensional space, the algorithm may still be attracted by "shallow" extreme points during the optimization process, and the non-convexity of the objective function still exists. The complex shape of the non-convex objective function will not disappear due to the transformation of the feature space, and there may still be multiple local optimal solutions, making it impossible to find the true global optimal solution, resulting in an insufficiently optimal decision boundary and limited classification accuracy. To solve the above problems, the present invention adopts a search mechanism based on chaos theory to explore more possible solution spaces during non-convex optimization, so as to find a better decision boundary and thus improve the classification accuracy. The construction method of the classifier model in the embodiments of the present invention is as follows:

[0128] 1) Input the preset training sample set into a preset support vector machine, and use the chaos theory algorithm to adjust the support vector machine parameters of the support vector machine.

[0129] Among them, the embodiments of the present invention use chaos theory to generate random perturbations; use a preset non-convex optimization algorithm to adjust the support vector machine parameters of the support vector machine based on the random perturbations.

[0130] First, utilize the initial value sensitivity of chaos theory to start the search process from multiple different initial points to ensure covering different regions of the solution space and avoid the algorithm falling into local optimal solutions, expressed as:

[0131]

[0132]

[0133] In the formula, is the value of the chaotic variable at the t-th iteration; is the value of the chaotic variable at the (t + 1)-th iteration; is the parameter of the chaotic mapping. Preferably, select a value that can make show chaotic behavior, such as μ u = 4; is the weight vector of the support vector machine at the t-th iteration; is the initial weight vector of the support vector machine; is the chaotic learning rate. Preferably, it is set to 0.01.

[0134] Furthermore, the role of the random perturbation generated based on the chaos theory is to increase the diversity of the search process, and to increase the randomness and exploration in the optimization process through the chaos theory, so as to avoid falling into the local optimal solution. The calculation method is expressed as:

[0135]

[0136] In the formula, is the coefficient that controls the magnitude of the perturbation effect. Preferably, it is set to 0.001.

[0137] On each search path, a non-convex optimization algorithm is executed to adjust the decision boundary of the support vector machine, and the optimization direction and step size are adjusted according to the classification error and the non-linearity output by the kernel function. The way of adjusting the support vector machine parameters by the non-convex optimization through the gradient descent method is expressed as:

[0138]

[0139]

[0140] In the formula, is the weight vector of the support vector machine at the (t + 1)-th iteration; is the step size of the non-convex optimization, is the gradient of the objective function of the support vector machine with respect to the weight, is the label of the i-th data sample, is the feature vector of the i-th data sample; represents the random perturbation generated based on the chaos theory. Preferably, it is set to 0.01.

[0141] Among them, the embodiments of the present invention also initialize the parameters of the support vector machine for training, including selecting a suitable kernel function type and initializing the regularization parameter, which is expressed as:

[0142]

[0143]

[0144] In the formula, is the radial basis function kernel, which is used to calculate the kernel function value between two data samples and ; is the first data sample input to the support vector machine; is the second data sample input to the support vector machine; is the width parameter of the kernel function; L u is the objective function of the support vector machine, w u is the weight vector of the support vector machine; represents the L2 norm; is the slack variable, used to handle the case of incomplete separability; is the regularization parameter, controlling the penalty intensity of the error term; N is the number of samples in the training sample set.

[0145] 2) Determine the Lagrange multipliers corresponding to the training sample set, and based on the Lagrange multipliers, adjust the parameters of the kernel function of the support vector machine.

[0146] Among them, the parameters of the kernel function can be dynamically adjusted based on the data characteristics obtained from the above feature dimensionality reduction model to better adapt to the distribution characteristics of the data. The dynamic adjustment of the width parameter of the kernel function is achieved through the following error minimization, expressed as:

[0147]

[0148]

[0149] In the formula, is the width parameter of the kernel function in the (t + 1)-th iteration; is the width parameter of the kernel function in the t-th iteration; is the adjustment rate of the width parameter of the kernel function, is the Lagrange multiplier of the i-th data sample; is the Lagrange multiplier of the j-th data sample; E total is the objective error function of the support vector machine; is the label of the i-th data sample; is the label of the j-th data sample.

[0150] 3) Until the support vector machine meets the preset training conditions, construct a classifier model based on the support vector machine.

[0151] Repeat the above steps iteratively until the preset stop iteration condition is met, which means the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0152] The embodiment of the present invention uses a support vector machine algorithm based on a search mechanism of chaos theory for classification. By exploring more possible solution spaces in the non-convex optimization process, a better decision boundary is found, solving the problem that the traditional support vector machine algorithm may fall into a local optimal solution in the non-convex optimization process.

[0153] Furthermore, the feature extraction model, autoencoder, and classifier model of the embodiments of the present invention are respectively trained through corresponding training sample sets. Among them, the samples after feature extraction can be used as the training sample set of the autoencoder, and further the data after dimensionality reduction processing can be used as the training sample set of the classifier model. In an actual industrial environment, the acquisition of high-quality fault data often comes with high costs. For example, a large number of sensors need to be installed and long-term data accumulation is required. In addition, due to the low frequency of equipment failures under normal operating conditions, the fault samples for training the model are relatively scarce. This not only limits the generalization ability of the model, but also makes it difficult for the model accuracy to meet the actual needs, and also leads to the problem of unbalanced training data. The acquisition, annotation, and preprocessing of training data are time-consuming and laborious, and insufficient training samples easily result in poor generalization ability of the model, while affecting the accuracy of the model. To solve this problem, the present invention adopts a data augmentation algorithm based on an adaptive coupling field to construct a training sample set after data augmentation. In the generation process, it not only depends on random noise, but also combines the context of device status information, and can more accurately generate samples with different fault categories, improving the diversity and coverage of the generated samples. Among them, Figure 3 shows a flowchart of a method for constructing a training sample set. Refer to Figure 3 and the method includes the following steps:

[0154] Step S10: Obtain the working parameters pre-stored in the preset industrial production equipment, annotate the working parameters, and construct an initial sample set.

[0155] Among them, the working parameters can refer to the above embodiments and will not be elaborated here. Further, the collected data is annotated. In one implementation, manual annotation can be used, and the annotation categories include different device fault phenomena, such as "hydraulic pressure too high", "knife chain running out of position", etc.

[0156] Step S20: Generate random noise corresponding to the initial sample set, and obtain the current operation data of the industrial production equipment.

[0157] Step S30: Based on the random noise and the current operation data, determine the coupling input corresponding to the industrial production equipment.

[0158] The embodiments of the present invention construct a training sample set by augmenting the data of the initial sample set. Among them, the input of the augmentation algorithm of the embodiments of the present invention is not only a random noise vector, but also adjusted in combination with the status information of the device. Each time of training, the current operation data of the device (such as vibration, temperature, etc.) is used as the coupling field condition for adjusting the input noise. In this way, the data augmentation algorithm can not only generate the form of data, but also generate virtual samples with specific attributes (such as fault type, fault degree, etc.) according to the specific device fault status. The input of the data augmentation algorithm is expressed as:

[0159]

[0160] In the formula, is an adaptive coupling field driving function, which generates a coupling input based on the weighted combination of device state information and noise, and is used to generate new input features; is the input feature of the generator that combines the adaptive coupling field; Z c is a random noise vector; S c is the device state information (including device parameters such as vibration and temperature); α c is the first adaptive adjustment coefficient, and β c is the second adaptive adjustment coefficient. The first adaptive adjustment coefficient and the second adaptive adjustment coefficient are used to adjust the weights of the noise and the device state information.

[0161] The first adaptive adjustment coefficient and the second adaptive adjustment coefficient follow the adaptive optimization mechanism of error backpropagation, and the update rule is:

[0162]

[0163]

[0164] In the formula, is the loss function used to characterize the preset data augmentation algorithm; is the learning rate of the adjustment coefficient, indicating the step size of the adjustment coefficient update; is the first adaptive adjustment coefficient at the (t + 1)-th iteration; is the second adaptive adjustment coefficient at the (t + 1)-th iteration; is the first adaptive adjustment coefficient at the (t + 1)-th iteration; is the first adaptive adjustment coefficient at the (t + 1)-th iteration; is the partial derivative symbol.

[0165] Step S40: Based on the coupling input, perform data augmentation on the initial sample set to construct a training sample set.

[0166] Furthermore, performing data augmentation on the above-mentioned coupling input to construct the training sample set of the embodiment of the present invention can improve the sample diversity and coverage, solve the problem of unbalanced training data caused by the high cost of actual device fault data collection and the scarcity of fault states, and the problem of poor model generalization ability and accuracy impact caused by insufficient training samples.

[0167] In one implementation manner, the data augmentation algorithm can adopt a generative adversarial network. Correspondingly, the above-mentioned It can be used to characterize the loss function of the generator of a generative adversarial network. Corresponding to the generative adversarial network, after obtaining the above-mentioned coupled input, the coupled input is input into a pre-trained generative adversarial network, and the generative adversarial network is used to generate an augmented sample corresponding to the coupled input; the augmented sample is combined with the initial sample set to construct a training sample set. Among them, the generative adversarial network updates the generator based on the exponential calculation of the output of the discriminator.

[0168] During the training process of the generative adversarial network, first initialize the parameters of the generative adversarial network. Let the generator be G c , and the discriminator be D c . The weight of the generator is , and the weight of the discriminator is . The bias of the generator is , and the bias of the discriminator is . Initialize the parameters of the generator and the discriminator to values on a normal distribution, and this normal distribution is a distribution with a mean of 0 and a variance of the identity matrix.

[0169] Among them, the goal of the discriminator is to identify whether the input data is real data. Specifically, the discriminator accepts two types of data: one is the real fault data collected from the device, and the other is the fake data generated by the generator; the discriminator optimizes its parameters so that it can accurately distinguish these two types of data. The loss function of the discriminator is expressed as:

[0170]

[0171] In the formula, preal represents the distribution of real data, Xc is the real data, is the data generated by the generator, is the discriminator function; obeys a specific distribution; represents the expectation.

[0172] After the discriminator is updated, the generator is optimized. The goal of the generator is to deceive the discriminator so that the discriminator cannot accurately judge which data is generated. Moreover, the optimization goal of the generator is to maximize the probability that the discriminator judges the generated data as real data. The loss function of the generator is expressed as:

[0173]

[0174] In the formula, P noise represents the distribution of noise; P state represents the distribution of the device state; is the generator function.

[0175] Different from the prior art, in the backpropagation of the generator in the embodiments of the present invention, the update rule is based on the feedback information of the discriminator, making the weight update of the generator more flexible and avoiding the problem of gradient disappearance or oscillation that may occur in traditional methods. The update method of the generator weights is expressed as:

[0176]

[0177] In the formula, is the learning rate of the generative adversarial network, is the rate constant; is the weight of the generator during the t-th iteration training; is the weight of the generator during the (t + 1)-th iteration training. is the exponential function.

[0178] Among them, the weight update of the conventional generator is usually linear. When the quality of the samples generated by the generator is very high, the output of the discriminator is close to 1, resulting in the gradient approaching 0 and the problem of gradient disappearance. However, in the present invention, by using the exponential function, the weight update has non-linear characteristics and is more flexible. The generator dynamically adjusts the weights according to the feedback of the discriminator. When the quality of the samples generated by the generator is relatively high (close to "real"), the weight update amplitude is small; when the quality of the samples generated by the generator is relatively low (far from "real"), the weight update amplitude is large. Among them, through the adjustment of the exponential function, even when the discrimination probability of the discriminator is close to 1, a certain gradient can be maintained, ensuring that the weight update is smooth and bounded, avoiding drastic weight changes, avoiding gradient disappearance, and helping to improve the stability of training. Therefore, the weight update rule of the present invention is smoother and more flexible, can better cope with the fluctuations in the training process, and improve the training stability of the entire generative adversarial network.

[0179] Further, repeat the above steps iteratively until the preset stop iteration condition is satisfied, which means the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times. Among them, after the data augmentation model training is completed, the trained data augmentation model is used to increase the number of samples. In one embodiment, assuming the original collected samples are 800, and the data augmentation model expands and generates 200 samples, then the expanded dataset contains 1000 samples.

[0180] Further, on the basis of the above method embodiments, the embodiments of the present invention also provide a knowledge-driven device predictive maintenance device, Figure 4 shows the corresponding structural schematic diagram of the embodiments of the present invention. Refer to Figure 4, the device includes: a data acquisition module 100, configured to acquire the working parameters of a preset industrial device, perform data preprocessing on the working parameters, and generate a vector to be measured; a feature extraction module 200, configured to input the vector to be measured into a pre-constructed feature extraction model to extract the target feature corresponding to the vector to be measured; wherein, the feature extraction model updates network parameters based on the feature non-linear dependence relationship corresponding to a preset training sample set; a data processing module 300, configured to perform data dynamic reshaping operation on the target feature, and use a preset autoencoder to perform dimensionality reduction processing on the dynamically reshaped target feature to generate a dimensionality reduction feature representation; wherein, the encoding layer of the autoencoder is sparsely optimized based on a non-linear coupling optimization strategy; an execution module 400, configured to use a pre-constructed classifier model to classify and identify the dimensionality reduction feature representation, and determine the device fault phenomenon indicated by the current working parameters of the industrial device. The device for predictive maintenance of equipment based on knowledge drive provided by the embodiments of the present invention has the same technical features as the method for predictive maintenance of equipment based on knowledge drive provided by the above embodiments, so it can also solve the same technical problems and achieve the same technical effects.

[0181] The above-mentioned execution module 400 is further configured to input the dimensionality reduction feature representation into a pre-constructed classifier model and output a classification result of the device fault phenomenon; use a preset predictive maintenance knowledge graph system to diagnose the classification result of the device fault phenomenon, and determine the potential fault cause corresponding to the classification result of the device fault phenomenon. The above-mentioned data processing module 300 is further configured to analyze the statistical characteristics corresponding to the target feature, calculate the data reshaping factor corresponding to the target feature; use a preset autoencoder to perform preliminary dimensionality reduction processing on the target feature, determine the loss function corresponding to the autoencoder, and determine the weight matrix reshaping factor corresponding to the loss function; calculate a reshaping matrix based on the data reshaping factor and the weight matrix reshaping factor; use the reshaping matrix to perform data dynamic reshaping on the target feature. The above-mentioned data processing module 300 is further configured to use a preset coupling degree constraint function to determine the non-linear association between the features of a preset training sample set; based on the non-linear association and a preset feature sparsity constraint, determine the loss function of a preset autoencoder; based on the loss function and a preset momentum factor, perform backpropagation update on the weight matrix of the autoencoder to optimize the encoding layer of the autoencoder.

[0182] The above-mentioned execution module 400 is further configured to input a preset training sample set into a preset support vector machine, and use the chaos theory algorithm to adjust the parameters of the support vector machine; determine the Lagrange multipliers corresponding to the training sample set, and based on the Lagrange multipliers, adjust the parameters of the kernel function of the support vector machine; until the support vector machine meets the preset training conditions, construct a classifier model based on the support vector machine. The above-mentioned execution module 400 is further configured to generate random perturbations using the chaos theory; use a preset non-convex optimization algorithm to adjust the parameters of the support vector machine based on the random perturbations. Further, the above-mentioned feature extraction module 200 is further configured to obtain a preset training sample set, perform non-linear decomposition on the training sample set using the non-linear feature dual decomposition algorithm to obtain a plurality of dual components; input the plurality of dual components into a preset neural network for forward propagation, and determine the loss function corresponding to the neural network; update the network parameters of the neural network based on the loss function and the feature non-linear dependence relationship indicated by the dual components; until the neural network meets the preset iteration conditions, construct a feature extraction model based on the neural network.

[0183] The device further includes a construction module, configured to obtain the working parameters pre-stored in a preset industrial production device, label the working parameters, and construct an initial sample set; generate random noise corresponding to the initial sample set, and obtain the current operation data of the industrial production device; based on the random noise and the current operation data, determine the coupled input corresponding to the industrial production device; perform data augmentation on the initial sample set based on the coupled input to construct a training sample set. The above-mentioned construction module is further configured to input the coupling into a pre-trained generative adversarial network, and use the generative adversarial network to generate augmented samples corresponding to the coupled input; merge the augmented samples with the initial sample set to construct a training sample set; wherein, the generative adversarial network updates the generator based on the exponential calculation of the output of the discriminator.

[0184] An embodiment of the present invention further provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the above Figures 1 to 3 steps of any of the shown methods. An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, it executes the above Figures 1 to 3 steps of the shown method. An embodiment of the present invention further provides a schematic structural diagram of an electronic device, as shown in Figure 5 shown, which is a schematic structural diagram of the electronic device. Among them, the electronic device includes a processor 51 and a memory 50. The memory 50 stores computer-executable instructions that can be executed by the processor 51. The processor 51 executes the computer-executable instructions to implement the above Figures 1 to 3 steps of the shown method. InFigure 5 In the illustrated embodiment, the electronic device further includes a bus 52 and a communication interface 53. Among them, the processor 51, the communication interface 53, and the memory 50 are connected through the bus 52. Among them, the memory 50 may include a high-speed random access memory (RAM, Random Access Memory), and may also include a non-volatile memory, such as at least one disk memory. Through at least one communication interface 53 (which can be wired or wireless), a communication connection is realized between this system network element and at least one other network element, and the Internet, wide area network, local area network, metropolitan area network, etc. can be used. The bus 52 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc., and can also be an AMBA (Advanced Microcontroller Bus Architecture) bus. Among them, AMBA defines three buses, including the APB (Advanced Peripheral Bus) bus, the AHB (Advanced High-performance Bus) bus, and the AXI (Advanced eXtensible Interface) bus. The bus 52 can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience of representation, Figure 5 only a single bidirectional arrow is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0185] The processor 51 may be an integrated circuit chip with the ability to process signals. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor 51 or the instructions in the form of software. The above-mentioned processor 51 may be a general-purpose processor, including a central processing unit (CPU for short), a network processor (NP for short), etc.; it may also be a digital signal processor (DSP for short), an application specific integrated circuit (ASIC for short), a field-programmable gate array (FPGA for short), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in the embodiments of the present application can be directly embodied as being executed and completed by a hardware decoding processor, or executed and completed by a combination of the hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. This storage medium is located in the memory, and the processor 51 reads the information in the memory and combines its hardware to complete the foregoing Figures 1 to 3 any of the shown methods.

[0186] A computer program product of a knowledge-driven device predictive maintenance method and device provided by an embodiment of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the method described in the foregoing method embodiments. For the specific implementation, reference can be made to the method embodiments and will not be elaborated here. Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the above-described system can refer to the corresponding process in the foregoing method embodiments and will not be elaborated here. Additionally, in the description of the embodiments of the present invention, unless otherwise clearly specified and limited, the terms "install", "connect", and "couple" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two elements. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations. If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program code. In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be understood as a limitation to the present invention. In addition, the terms "first", "second", and "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.Finally, it should be noted that the above embodiments are only specific embodiments of the present invention, used to illustrate the technical solutions of the present invention, rather than limiting it. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art within the technical scope disclosed by the present invention can still modify the technical solutions recorded in the foregoing embodiments or can easily think of changes, or perform equivalent replacements on some of the technical features; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A knowledge-driven device predictive maintenance method, characterized in that The method includes: Obtain the working parameters of a preset industrial device, perform data preprocessing on the working parameters to generate a vector to be measured; Input the vector to be measured into a pre-constructed feature extraction model to extract the target features corresponding to the vector to be measured; wherein, the feature extraction model updates network parameters based on the feature non-linear dependence relationship corresponding to a preset training sample set; Perform a data dynamic reshaping operation on the target features, and use a preset autoencoder to perform dimensionality reduction processing on the dynamically reshaped target features to generate a dimensionality-reduced feature representation; wherein, the encoding layer of the autoencoder is sparsely optimized based on a non-linear coupling optimization strategy; Use a pre-constructed classifier model to classify and identify the dimensionality-reduced feature representation, and determine the device fault phenomenon indicated by the current working parameters of the industrial device; The method for constructing the feature extraction model includes: Obtain a preset training sample set, perform non-linear decomposition on the training sample set using a non-linear feature dual decomposition algorithm to obtain multiple dual components; input the multiple dual components into a preset neural network for forward propagation, and determine the loss function corresponding to the neural network; based on the loss function and the feature non-linear dependence relationship indicated by the dual components, update the network parameters of the neural network; until the neural network meets the preset iteration conditions, construct a feature extraction model based on the neural network; The steps of sparsely optimizing the encoding layer of the autoencoder based on a non-linear coupling optimization strategy include: Use a preset coupling degree constraint function to determine the non-linear association between the features of a preset training sample set; based on the non-linear association and a preset feature sparsity constraint, determine the loss function of a preset autoencoder; based on the loss function and a preset momentum factor, perform backpropagation update on the weight matrix of the autoencoder to optimize the encoding layer of the autoencoder; The steps of performing data dynamic reshaping on the target features include: Analyze the statistical characteristics corresponding to the target features, and calculate the data reshaping factor corresponding to the target features; use a preset autoencoder to perform preliminary dimensionality reduction processing on the target features, determine the loss function corresponding to the autoencoder, and determine the weight matrix reshaping factor corresponding to the loss function; calculate a reshaping matrix based on the data reshaping factor and the weight matrix reshaping factor; use the reshaping matrix to perform data dynamic reshaping on the target features.

2. The method according to claim 1, characterized in that, The steps of using a pre-constructed classifier model to classify and identify the dimensionality-reduced feature representation and determine the device fault phenomenon indicated by the current working parameters of the industrial device include: Input the dimensionality-reduced feature representation into a pre-constructed classifier model to output a classification result of the device fault phenomenon; Use a preset predictive maintenance knowledge graph system to diagnose the classification result of the device fault phenomenon and determine the potential fault causes corresponding to the classification result of the device fault phenomenon.

3. The method according to claim 1, wherein The method for constructing the classifier model includes: Input a preset training sample set into a preset support vector machine, and use a chaos theory algorithm to adjust the support vector machine parameters of the support vector machine; Determine the Lagrange multipliers corresponding to the training sample set, and based on the Lagrange multipliers, adjust the parameters of the kernel function of the support vector machine; Until the support vector machine meets the preset training conditions, construct a classifier model based on the support vector machine.

4. The method according to claim 3, characterized in that The step of adjusting the parameters of the support vector machine of the support vector machine by using the chaos theory algorithm includes: Generate random perturbations using the chaos theory; Based on the random perturbations, adjust the parameters of the support vector machine of the support vector machine by using a preset non-convex optimization algorithm.

5. The method according to any one of claims 1 to 4, characterized in that, The construction method of the preset training sample set includes: Obtain the working parameters pre-stored in the preset industrial production equipment, label the working parameters, and construct an initial sample set; Generate random noise corresponding to the initial sample set, and obtain the current operation data of the industrial production equipment; Based on the random noise and the current operation data, determine the coupled input corresponding to the industrial production equipment; Based on the coupled input, perform data augmentation on the initial sample set to construct a training sample set.

6. The method according to claim 5, characterized in that, The step of performing data augmentation on the initial sample set based on the coupled input to construct a training sample set includes: Input the coupled input into a pre-trained generative adversarial network, and use the generative adversarial network to generate augmented samples corresponding to the coupled input; Combine the augmented samples with the initial sample set to construct a training sample set; Wherein, the generative adversarial network updates the generator based on the exponential calculation of the output of the discriminator.

7. A knowledge-driven device predictive maintenance device, characterized in that, The device includes: A data acquisition module, configured to acquire the working parameters of the preset industrial equipment, perform data preprocessing on the working parameters, and generate a vector to be measured; A feature extraction module, configured to input the vector to be measured into a pre-constructed feature extraction model to extract the target feature corresponding to the vector to be measured; wherein, the feature extraction model updates the network parameters based on the feature non-linear dependence relationship corresponding to the preset training sample set; A data processing module, configured to perform data dynamic reshaping operations on the target feature, and use a preset autoencoder to perform dimensionality reduction processing on the dynamically reshaped target feature to generate a dimensionality reduction feature representation; wherein, the encoding layer of the autoencoder is sparsely optimized based on a non-linear coupling optimization strategy; An execution module, configured to perform classification and recognition on the dimensionality reduction feature representation by using a pre-constructed classifier model, and determine the device fault phenomenon indicated by the current working parameters of the industrial equipment; The feature extraction module is further configured to: obtain a preset training sample set, perform non-linear decomposition on the training sample set by using a non-linear feature dual decomposition algorithm to obtain a plurality of dual components; input the plurality of dual components into a preset neural network for forward propagation, and determine the loss function corresponding to the neural network; based on the loss function and the feature non-linear dependence relationship indicated by the dual components, update the network parameters of the neural network; until the neural network meets the preset iteration conditions, construct a feature extraction model based on the neural network; The data processing module is further configured to: determine the non-linear association between features of a preset training sample set by using a preset coupling degree constraint function; determine a loss function of a preset autoencoder based on the non-linear association and a preset feature sparsity constraint; update the weight matrix of the autoencoder in a backpropagation manner based on the loss function and a preset momentum factor, so as to optimize the encoding layer of the autoencoder; The data processing module is further configured to: analyze the statistical characteristics corresponding to the target feature, and calculate a data reshaping factor corresponding to the target feature; perform preliminary dimensionality reduction processing on the target feature by using a preset autoencoder, determine a loss function corresponding to the autoencoder, and determine a weight matrix reshaping factor corresponding to the loss function; calculate a reshaping matrix based on the data reshaping factor and the weight matrix reshaping factor; perform data dynamic reshaping on the target feature by using the reshaping matrix.

Images