Method and medium for industrial equipment residual life prediction based on classification and error correction knowledge distillation

Abstract

The application relates to an industrial equipment residual life prediction method and medium based on classification and error correction knowledge distillation, and relates to the technical field of industrial equipment life prediction. The similarity between the teacher output prediction value and the corresponding true value is used to divide the difficult samples and simple samples of each time sequence, so that the distillation loss function can be constructed based on the obtained difficult samples and simple samples, thereby obtaining the total loss function, so that the student network model can perform knowledge self-reflection in the training process combined with the loss function, instead of directly replacing the true value, emphasizing self-correction, and the adaptability of the student model to the real data is enhanced, so that the prediction accuracy can be improved while reducing the reasoning memory, thereby achieving a better balance between accuracy and reasoning memory, so as to facilitate easier deployment on mobile devices.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment life prediction technology, specifically to a method and medium for predicting the remaining life of industrial equipment based on classification and error correction knowledge distillation. Background Technology

[0002] Health management (PHM) is a crucial research area in industry, emphasizing real-time monitoring and analysis of equipment condition to predict health status and potential failures. Remaining useful life (RUL) prediction is a core aspect of PHM, playing a vital role in improving equipment reliability, optimizing maintenance plans, and reducing operating costs. Currently, methods for predicting RUL mainly include physical model-based methods, statistical model-based methods, and data-driven methods. Physical model-based methods predict RUL based on physical equations and models, according to the equipment's operating principles and wear mechanisms. These methods require a deep understanding of the equipment's physical characteristics and operating environment. Statistical model-based methods require historical failure data and lifetime distribution models to estimate the equipment's RUL. These models typically require large amounts of historical data to produce reliable results. Data-driven models, on the other hand, use machine learning and deep learning algorithms to predict RUL by analyzing sensor data such as vibration, temperature, and sound, and have become a hot research area in recent years.

[0003] However, traditional deep learning networks typically require significant computational resources and memory, limiting their deployment on edge devices. Therefore, achieving a better balance between accuracy and inference memory for easier deployment on mobile devices has become a pressing issue. Summary of the Invention

[0004] The technical problem to be solved by this application is to provide a method and medium for predicting the remaining life of industrial equipment based on classification and error correction knowledge distillation, which has the characteristics of achieving a better balance between accuracy and inference memory, so as to facilitate easier deployment on mobile devices.

[0005] In one aspect, an embodiment provides a method for predicting the remaining useful life of industrial equipment based on classification and error correction knowledge distillation, comprising: training an industrial equipment remaining useful life prediction network based on classification and error correction knowledge distillation by combining a loss function, and predicting the remaining useful life of industrial equipment based on the industrial equipment remaining useful life prediction network; wherein the industrial equipment remaining useful life prediction network is a neural network model trained based on classification and error correction knowledge distillation, and the training method includes:

[0006] The time series of parameters for predicting the remaining life of industrial equipment is obtained and labeled with the true value of the remaining life to obtain a labeled time series of parameters. The labeled time series of parameters is then preprocessed to obtain training samples.

[0007] The training samples are input into multiple teacher network models to obtain their respective output prediction values; the multiple teacher network models are neural network models with complementary advantages.

[0008] The teacher's output prediction value is obtained by linearly fusing the corresponding output prediction values.

[0009] Based on a preset similarity threshold, the predicted teacher output values ​​at each time point are classified into simple samples and difficult samples. The simple samples indicate that the similarity between the predicted teacher output value and the corresponding real value reaches the preset similarity threshold, while the difficult samples indicate that the similarity between the predicted teacher output value and the corresponding real value is less than the preset similarity threshold.

[0010] Based on the time points, the simple and difficult samples are mapped to the time series of predicted values ​​output by teachers, the time series of actual values ​​output by students, and the time series of predicted values ​​output by the student network model, thereby obtaining the time series of each corresponding labeled with difficult and simple samples.

[0011] A first distillation loss function is constructed based on the difficult samples of the time series of predicted values ​​output by teachers, the time series of actual values, and the time series of predicted values ​​output by the student network model.

[0012] Based on simple samples of the time series of predicted values ​​output by teachers and the time series of predicted values ​​output by student network models, a second distillation loss function is constructed.

[0013] The third distillation loss function is obtained based on the first and second distillation loss functions;

[0014] The total loss function is constructed based on the third distillation loss function and the loss function between the predicted and actual values ​​of the student network model.

[0015] The student network model is trained using the total loss function, and the trained student network model is used as a prediction network for the remaining life of industrial equipment.

[0016] In one embodiment, classifying the teacher output prediction values ​​at each time point into easy samples and hard samples based on a preset similarity threshold includes:

[0017] Construct a time series of predicted and actual values ​​of teacher outputs into a single... A two-dimensional time series, where n is the length of the time series;

[0018] The above Transpose the two-dimensional time series to generate a The feature matrix;

[0019] Set a local sliding window and slide the local sliding window according to the time sequence and the preset step size;

[0020] For each sliding window, the correlation between the two column vectors within the sliding window is calculated, including: generating a Gaussian kernel matrix based on the two column vectors, and decentering the Gaussian kernel matrix to obtain a Gaussian kernel correlation matrix; obtaining the feature entropy corresponding to each local sliding window based on the Gaussian kernel correlation matrix; and defining the data corresponding to the sliding window with feature entropy less than the entropy threshold as simple samples based on a preset entropy threshold, and defining the data corresponding to the sliding window with feature entropy greater than or equal to the entropy threshold as hard samples.

[0021] In one embodiment, generating a Gaussian kernel matrix based on two column vectors includes:

[0022]

[0023]

[0024] Where K represents the Gaussian kernel matrix, k represents the Gaussian kernel function, x1, x2, and x3 represent the time series values ​​of the first column vector, y1, y2, and y3 represent the time series values ​​of the second column vector, p represents the index of the time series value in the first column vector, and q represents the index of the time series value in the second column vector. This indicates the calculation of Euclidean distance. represents the kernel width of the Gaussian kernel function, and exp represents the natural exponential function;

[0025] The process of decentering the Gaussian kernel matrix to obtain the Gaussian kernel correlation matrix includes:

[0026]

[0027]

[0028]

[0029] in, This represents the decentralized Gaussian kernel function. Let m represent a square matrix with the same dimensions as the Gaussian kernel matrix K and all elements being identical, where m represents the dimension of the square matrix; C represents the Gaussian kernel correlation matrix; and T represents the transpose.

[0030] The method for obtaining the feature entropy corresponding to each local sliding window based on the Gaussian kernel correlation matrix includes:

[0031]

[0032] Where FE represents the feature entropy, and r represents the index of any one of the three eigenvalues ​​from the singular value decomposition of the Gaussian kernel correlation matrix C. denoted by , represents the eigenvalues ​​of the singular value decomposition of the Gaussian kernel correlation matrix C, and log represents the logarithmic function to the base 2.

[0033] In a second aspect, one embodiment provides a computer-readable storage medium storing a program that can be loaded by a processor and executed by the industrial equipment remaining life prediction method described in any of the above embodiments.

[0034] The beneficial effects of this invention are:

[0035] By linearly fusing the outputs of multiple complementary teacher network models to obtain the teacher output prediction value, the student network model can better learn the teacher's characteristics and learning ability, thereby improving the prediction performance of the student network model. Furthermore, by dividing each time series into hard and easy samples based on the similarity between the teacher output prediction value and the corresponding true value, a distillation loss function can be constructed based on these hard and easy samples to obtain the total loss function. This allows the student network model to perform self-reflection during training, rather than directly replacing the true value, emphasizing self-correction and enhancing the student model's adaptability to real data. This allows for improved prediction accuracy while reducing inference memory, achieving a better balance between accuracy and inference memory, and facilitating easier deployment on mobile devices. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of a training method for an industrial equipment remaining life prediction network according to an embodiment of this application.

[0037] Figure 2 This application Figure 1 A schematic diagram of the method flow of one embodiment of step S04;

[0038] Figure 3 This application Figure 2 A schematic diagram of the method flow for one embodiment of step S0404. Detailed Implementation

[0039] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of this application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to this application are not shown or described in the specification. This is to avoid obscuring the core parts of this application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0040] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0041] The serial numbers assigned to components in this article, such as "first" and "second", are used only to distinguish the objects being described and have no sequential or technical meaning.

[0042] To facilitate the explanation of the inventive concept of this application, the knowledge distillation technology will be briefly described below.

[0043] Furthermore, the stringent requirements for data protection and operational efficiency in edge devices further highlight the need for lightweight models. Recent advances in lightweight models primarily focus on model compression techniques, including knowledge distillation (KD), quantization, pruning, and neural architecture search. Among these methods, KD has garnered significant attention due to its flexibility and efficiency. Knowledge distillation, as an effective model compression technique, extracts knowledge from a large, well-trained teacher network model and transfers it to a small, lightweight student network model, thereby significantly reducing the number of model parameters and computational complexity without significantly sacrificing predictive performance. While knowledge distillation was initially applied in fields such as computer vision and natural language processing, its application in RUL (Reliable Edge Utilization) industrial equipment is still in its early stages. Therefore, effectively and reliably deploying industrial equipment at the edge remains a major challenge.

[0044] In view of this, this application provides a method and medium for predicting the remaining useful life of industrial equipment based on classification and error correction knowledge distillation. In the training method of the remaining useful life prediction network for industrial equipment, the output prediction value of the teacher is obtained by linearly fusing the outputs of multiple teacher network models with complementary advantages. This allows the student network model to better learn the characteristics and learning ability of the teachers, thereby improving the prediction performance of the student network model. Since the difficult and easy samples of each time series are obtained by dividing the similarity between the teacher output prediction value and the corresponding real value, the distillation loss function can be constructed based on the obtained difficult and easy samples to obtain the total loss function. This allows the student network model to perform knowledge self-reflection during the training process by combining the loss function, rather than directly replacing the real value. This emphasizes self-correction and enhances the adaptability of the student model to real data. It can improve the prediction accuracy while reducing inference memory, thereby achieving a better balance between accuracy and inference memory, and making it easier to deploy on mobile devices.

[0045] One embodiment of this application provides a method for predicting the remaining useful life of industrial equipment based on classification and error correction knowledge distillation. The method includes: training an industrial equipment remaining useful life prediction network based on classification and error correction knowledge distillation using a loss function; and predicting the remaining useful life of industrial equipment based on this network. The industrial equipment remaining useful life prediction network is a neural network model trained using knowledge distillation based on classification and error correction. Please refer to... Figure 1 The training method for this industrial equipment remaining life prediction network includes:

[0046] Step S01: Obtain the time series of parameters for predicting the remaining life of industrial equipment and label the actual value of the remaining life to obtain a labeled time series of parameters. Then, preprocess the labeled time series of parameters to obtain training samples.

[0047] In one embodiment, the acquired parameter time series may include vibration, temperature, and sound time series of industrial equipment collected by sensors. The parameter times are labeled to obtain parameter time series training samples with labels containing the true values ​​of remaining lifetime.

[0048] Those skilled in the art will understand that, in order to obtain the required training samples, preprocessing methods can be adopted using existing techniques, which will not be elaborated here.

[0049] Step S02: Input the training samples into multiple teacher network models respectively to obtain their respective output prediction values.

[0050] Among them, several teacher network models are neural network models with complementary advantages.

[0051] In one embodiment, the multiple teacher network models include a first teacher network model and a second teacher network model. Training samples are input into the first and second teacher network models respectively to obtain corresponding first and second output prediction values. The first and second teacher network models are complementary neural network models. For example, one teacher network model may excel in fast inference speed, while the other may excel in high accuracy; thus, the two teacher network models complement each other in their strengths.

[0052] Step S03: Linearly fuse the corresponding output prediction values ​​to obtain the teacher output prediction value.

[0053] In one embodiment, step S03 includes:

[0054]

[0055] in, This indicates that the teacher outputs the predicted value. This represents the first output prediction value. This represents the second output prediction value. This represents the hyperparameters that pre-determine the contributions of the first and second teacher network models, where 0 < 0. <1.

[0056] By leveraging the complementary strengths of the teacher network model, the student network model can better learn the characteristics and learning abilities of the teachers, thereby improving the predictive performance of the student network model.

[0057] In one embodiment, to more comprehensively reflect the performance of the teacher network model, one can select... =0.5.

[0058] Step S04: Based on a preset similarity threshold, classify the predicted output values ​​of teachers at each time point into simple samples and difficult samples.

[0059] In the case of simple samples, the similarity between the predicted teacher output value and the corresponding real value reaches the preset similarity threshold, while in the case of hard samples, the similarity between the predicted teacher output value and the corresponding real value is less than the preset similarity threshold.

[0060] In one embodiment, please refer to Figure 2 Step S04 may include:

[0061] Step S0401: Construct a time series of predicted values ​​and actual values ​​output by the teacher into a single... Two-dimensional time series.

[0062] Where n is the length of the time series.

[0063] Step S0402, the above Transpose the two-dimensional time series to generate a The characteristic matrix of .

[0064] Step S0403: Set a local sliding window and slide the local sliding window according to the time sequence and the preset step size.

[0065] In one embodiment, a can be defined A sliding window with a step size of 3 is used. The sliding window slides sequentially in chronological order. Each sliding window contains two columns of data: one column contains the predicted values ​​from three teachers, and the other column contains the actual values ​​from three teachers, for a total of six data points. The specific sliding window W can be represented as:

[0066]

[0067] Where X represents the column vector of predicted values ​​output by the teacher, Y represents the column vector of true values, x1, x2 and x3 represent the time series values ​​of the X column vector, and y1, y2 and y3 represent the time series values ​​of the Y column vector.

[0068] Since the data within the window consists of continuous time points, the causal relationship of the time series can be guaranteed.

[0069] Step S0404: For each sliding window, calculate the correlation between two column vectors within the sliding window.

[0070] Based on the sliding window given above, the correlation between two column vectors within the sliding window can be calculated. In one embodiment, please refer to... Figure 3 Step S0404 may include:

[0071] Step S100: Generate a Gaussian kernel matrix based on the two column vectors.

[0072] To analyze the correlation between two column vectors, in one implementation, we apply a Gaussian kernel function to calculate the correlation between each element of the two column vectors of the matrix. Combined with step S0403, step S100 may include:

[0073]

[0074]

[0075] Where K represents the Gaussian kernel matrix, k represents the Gaussian kernel function, x1, x2, and x3 represent the time series values ​​of the first column vector, y1, y2, and y3 represent the time series values ​​of the second column vector, p represents the index of the time series value in the first column vector, and q represents the index of the time series value in the second column vector. This indicates the calculation of Euclidean distance. represents the kernel width of the Gaussian kernel function, and exp represents the natural exponential function.

[0076] As a hyperparameter, in one embodiment, it can be preset to the average Euclidean distance of the data within the corresponding window.

[0077] Combining with step S0403, here x1, x2 and x3 represent the time series values ​​of the column vector of predicted values ​​output by the teacher, and y1, y2 and y3 represent the time series values ​​of the column vector of true values.

[0078] Based on the above process, the Gaussian kernel matrix can be mapped and calculated for each local window.

[0079] Step S200: Decenter the Gaussian kernel matrix to obtain the Gaussian kernel correlation matrix.

[0080] By decentering each Gaussian kernel correlation matrix, global biases can be removed, preventing individual points with large deviations from affecting the overall result. In one embodiment, step S200 may include:

[0081]

[0082]

[0083]

[0084] in, This represents the decentralized Gaussian kernel function. Let m represent a square matrix with the same dimension as the Gaussian kernel matrix K and all elements being identical, where m represents the dimension of the square matrix; C represents the Gaussian kernel correlation matrix; and T represents the transpose.

[0085] The Gaussian kernel correlation matrix C can be obtained through matrix operations, which can then be used to further extract correlations.

[0086] Step S300: Obtain the feature entropy corresponding to each local sliding window based on the Gaussian kernel correlation matrix.

[0087] In one embodiment, the singular values ​​of the generated Gaussian kernel correlation matrix C are decomposed into three eigenvalues, and then a eigenentropy value is generated using the eigenentropy formula, ultimately generating m / 3 eigenentropy values. Step S300 may include:

[0088]

[0089] Where FE represents the feature entropy, and r represents the index of any one of the three eigenvalues ​​from the singular value decomposition of the Gaussian kernel correlation matrix C. denoted by , represents the eigenvalues ​​of the singular value decomposition of the Gaussian kernel correlation matrix C, and log represents the logarithmic function to the base 2.

[0090] Step S400: Based on a preset entropy threshold, the data corresponding to the sliding window with feature entropy less than the entropy threshold is defined as simple samples, and the data corresponding to the sliding window with feature entropy greater than or equal to the entropy threshold is defined as difficult samples.

[0091] We arrange the generated feature entropy values ​​according to the order of the previous time series, and based on the preset entropy threshold, we can distinguish between simple and difficult samples.

[0092] In one embodiment, the entropy threshold can be preset based on the distribution of feature entropy and the performance of the student network model on a specific dataset. One embodiment sets the preset entropy threshold such that easy samples are greater than hard samples. In one embodiment, the preset entropy threshold can be 0.3. Since a larger feature entropy indicates lower similarity, and a smaller feature entropy indicates higher similarity, based on the preset feature entropy, data corresponding to sliding windows with feature entropy less than the entropy threshold are defined as easy samples, and data corresponding to sliding windows with feature entropy greater than or equal to the entropy threshold are defined as hard samples. Thus, easy samples indicate that the similarity between the obtained teacher output prediction value and the corresponding true value reaches the preset similarity threshold, while hard samples indicate that the similarity between the obtained teacher output prediction value and the corresponding true value is less than the preset similarity threshold.

[0093] Since both predicted and true values ​​are continuous within a local window, the above method allows us to focus not only on the relationship between predicted and true values ​​at the same time point, but also on the relationship between predicted and true values ​​at adjacent time points. Mathematically, a Gaussian kernel function is used to measure the correlation, and a Gaussian kernel matrix is ​​constructed. The correlation between predicted and true values ​​is then transformed into an entropy-based expression, and information classification is obtained through dispersion and correlation matrices. This information classification helps prevent incorrect transfer of teacher knowledge to the student network model, thereby improving the performance of the student network model.

[0094] Step S05: According to the time points, map the simple samples and difficult samples to the time series of predicted values ​​output by the teacher, the time series of actual values, and the time series of predicted values ​​output by the student network model, so as to obtain the time series of each corresponding labeled difficult and simple samples.

[0095] Based on the time points of the classified simple and difficult samples, and mapping them sequentially to the time series of predicted values ​​output by teachers, the time series of actual values ​​output by students, and the time series of predicted values ​​output by the student network model, we can obtain the simple sample of the teacher's predicted value time series. and difficult samples Simple samples of real-value time series and difficult samples A simple sample of the predicted time series output by the student network model. and difficult samples Where i represents the index of the difficult sample and j represents the index of the easy sample.

[0096] The applicant found in the research that in the current method, the prediction of the teacher network model is not accurate enough to be conveyed to the students at the time point corresponding to the difficult sample. Therefore, the true value is directly used to replace the difficult sample. However, this method of improving the performance of the student network model will hinder the student's ability to reflect on the teacher network model's wrong predictions.

[0097] In view of this, in one embodiment of this application, a novel method is employed to construct the loss function to overcome the defects caused by directly replacing difficult samples with true values.

[0098] Since the samples are divided into hard samples and easy samples, the distillation loss we obtain is divided into two parts: the first distillation loss for hard samples and the second distillation loss for easy samples.

[0099] Step S06: Construct the first distillation loss function based on the difficult samples of the teacher's output predicted value time series, the actual value time series, and the student network model's output predicted value time series.

[0100] By defining three objects—Hub, Positive, and Negative—you can then... Set as a hub, Set as negative, Set as a positive. Through dynamic learning, the hub will gradually adjust its position along the horizontal or vertical direction to reduce the Manhattan distance with the positive while increasing the Manhattan distance with the negative. Step S06 can then be expressed as:

[0101]

[0102]

[0103] in, Let represent the first distillation loss function, i represent the index of the hard sample, and I represent the number of hard samples, 1≤i≤I. This represents the i-th difficult sample in the predicted values ​​output by the student network model. This represents the i-th difficult sample in the true values. This represents the i-th difficult sample in the teacher's output predictions. This represents the margin of enforcement between positive and negative training sample pairs. Indicates the distance to Manhattan.

[0104] In this way, by employing the Manhattan distance strategy of knowledge self-reflection, and by having knowledge in the student network model reflect on itself rather than directly replacing the true value, the emphasis on self-correction enhances the adaptability of the student network model to real data and improves prediction accuracy.

[0105] Among them, for The constraints ensure that Hub samples and Positive samples are closer together.

[0106] Step S07: Based on simple samples of the time series of predicted values ​​output by teachers and the time series of predicted values ​​output by student network models, construct the second distillation loss function.

[0107] In one embodiment, step S07 includes:

[0108]

[0109] in, Let represent the distillation loss function, j represent the index of the simple sample, g represent the number of simple samples, and 1 ≤ j ≤ g. This represents the j-th simple sample in the predicted values ​​output by the student network model. This represents the j-th simple sample in the teacher's output prediction.

[0110] Step S08: Obtain the third distillation loss function based on the first and second distillation loss functions.

[0111] In one embodiment, step S08 includes:

[0112]

[0113] in, This represents the loss function of the third distillation.

[0114] Step S09: Obtain the total loss function based on the third distillation loss function and the loss function between the predicted value and the true value of the student network model.

[0115] In one embodiment, step S09 includes:

[0116]

[0117] in, Represents the total loss function. Indicates the loss weight, 0 < <1, This represents the loss function between the student network model's predicted values ​​and the actual values.

[0118] In one embodiment, The value is 0.8.

[0119] Those skilled in the art will understand that the loss function between the predicted and actual values ​​of the student network model can be obtained using existing techniques, which will not be elaborated here.

[0120] Step S10: Train the student network model using the total loss function, and use the trained student network model as the industrial equipment remaining life prediction network.

[0121] Based on the trained student network model, the preprocessed data to be predicted can be input into the student network model, and the final output is the RUL prediction value.

[0122] The industrial equipment remaining life prediction method based on any of the above embodiments obtains difficult and easy samples for each time series based on the similarity between the teacher's output predicted value and the corresponding true value. This allows for the construction of a distillation loss function based on the obtained difficult and easy samples, thereby obtaining the total loss function. This enables the student network model to perform knowledge self-reflection during training, rather than directly replacing the true value, emphasizing self-correction and enhancing the adaptability of the student model to real data. This allows for the improvement of prediction accuracy while reducing inference memory, thus achieving a better balance between accuracy and inference memory, making it easier to deploy on mobile devices.

[0123] One embodiment of this application provides a computer-readable storage medium storing a program, the stored program including methods that can be loaded by a processor and processed in any of the above embodiments.

[0124] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0125] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. A method for predicting the remaining life of industrial equipment based on classification and error correction knowledge distillation, characterized in that, include: By combining the loss function, an industrial equipment remaining life prediction network based on classification and error correction knowledge distillation is trained, and the remaining life of industrial equipment is predicted based on this industrial equipment remaining life prediction network. The industrial equipment remaining life prediction network is a neural network model trained based on knowledge distillation for classification and error correction. The training method includes: The time series of parameters for predicting the remaining life of industrial equipment is obtained and labeled with the true value of the remaining life to obtain a labeled time series of parameters. The labeled time series of parameters is then preprocessed to obtain training samples. The training samples are input into multiple teacher network models to obtain their respective output prediction values; the multiple teacher network models are neural network models with complementary advantages. The teacher's output prediction value is obtained by linearly fusing the corresponding output prediction values. Based on a preset similarity threshold, the predicted teacher output values ​​at each time point are classified into simple samples and difficult samples. The simple samples indicate that the similarity between the predicted teacher output value and the corresponding real value reaches the preset similarity threshold, while the difficult samples indicate that the similarity between the predicted teacher output value and the corresponding real value is less than the preset similarity threshold. Based on the time points, the simple and difficult samples are mapped to the time series of predicted values ​​output by teachers, the time series of actual values ​​output by students, and the time series of predicted values ​​output by the student network model, thereby obtaining the time series of each corresponding labeled with difficult and simple samples. A first distillation loss function is constructed based on the difficult samples of the time series of predicted values ​​output by teachers, the time series of actual values, and the time series of predicted values ​​output by student network models. Based on simple samples of the time series of predicted values ​​output by teachers and the time series of predicted values ​​output by student network models, a second distillation loss function is constructed. The third distillation loss function is obtained based on the first and second distillation loss functions; The total loss function is constructed based on the third distillation loss function and the loss function between the predicted and actual values ​​of the student network model. The student network model is trained using the total loss function, and the trained student network model is used as a prediction network for the remaining life of industrial equipment.

2. The method for predicting the remaining life of industrial equipment as described in claim 1, characterized in that, The process of inputting training samples into multiple teacher network models to obtain their respective output prediction values ​​includes: The training samples are input into the first teacher network model and the second teacher network model respectively to obtain the corresponding first output prediction value and second output prediction value. The method of linearly fusing the corresponding output prediction values ​​to obtain the teacher output prediction value includes: linearly fusing the first output prediction value and the second output prediction value to obtain the teacher output prediction value.

3. The method for predicting the remaining life of industrial equipment as described in claim 2, characterized in that, The method of linearly fusing the first output prediction value and the second output prediction value to obtain the teacher output prediction value includes: in, This indicates that the teacher outputs the predicted value. This represents the first output prediction value. This represents the second output prediction value. This represents the hyperparameters that pre-set the adjustments to the contributions of the first and second teacher network models.

4. The method for predicting the remaining life of industrial equipment as described in claim 1, characterized in that, The method described above, based on a preset similarity threshold, classifies the predicted teacher output values ​​at each time point into easy and difficult samples, including: Construct a time series of predicted and actual values ​​of teacher outputs into a single... A two-dimensional time series, where n is the length of the time series; The above Transpose the two-dimensional time series to generate a The feature matrix; Set a local sliding window and slide the local sliding window according to the time sequence and the preset step size; For each sliding window, the correlation between the two column vectors within the sliding window is calculated, including: generating a Gaussian kernel matrix based on the two column vectors, and decentering the Gaussian kernel matrix to obtain a Gaussian kernel correlation matrix; obtaining the feature entropy corresponding to each local sliding window based on the Gaussian kernel correlation matrix; and defining the data corresponding to the sliding window with feature entropy less than the entropy threshold as simple samples based on a preset entropy threshold, and defining the data corresponding to the sliding window with feature entropy greater than or equal to the entropy threshold as hard samples.

5. The method for predicting the remaining life of industrial equipment as described in claim 4, characterized in that, The method for generating a Gaussian kernel matrix based on two column vectors includes: Where K represents the Gaussian kernel matrix, k represents the Gaussian kernel function, x1, x2, and x3 represent the time series values ​​of the first column vector, y1, y2, and y3 represent the time series values ​​of the second column vector, p represents the index of the time series value in the first column vector, and q represents the index of the time series value in the second column vector. This indicates the calculation of Euclidean distance. represents the kernel width of the Gaussian kernel function, and exp represents the natural exponential function; The process of decentering the Gaussian kernel matrix to obtain the Gaussian kernel correlation matrix includes: in, This represents the decentralized Gaussian kernel function. Let m represent a square matrix with the same dimensions as the Gaussian kernel matrix K and all elements being identical, where m represents the dimension of the square matrix; C represents the Gaussian kernel correlation matrix; and T represents the transpose. The method for obtaining the feature entropy corresponding to each local sliding window based on the Gaussian kernel correlation matrix includes: Where FE represents the feature entropy, and r represents the index of any one of the three eigenvalues ​​from the singular value decomposition of the Gaussian kernel correlation matrix C. denoted by , represents the eigenvalues ​​of the singular value decomposition of the Gaussian kernel correlation matrix C, and log represents the logarithmic function to the base 2.

6. The method for predicting the remaining life of industrial equipment as described in claim 1, characterized in that, The first distillation loss function is constructed based on the difficult samples of the teacher's output predicted value time series, the actual value time series, and the student network model's output predicted value time series, including: in, Let represent the first distillation loss function, i represent the index of the hard sample, and I represent the number of hard samples, 1≤i≤I. This represents the i-th difficult sample in the predicted values ​​output by the student network model. This represents the i-th difficult sample in the true values. This represents the i-th difficult sample in the teacher's output predictions. This represents the margin of enforcement between positive and negative training sample pairs. Indicates the distance to Manhattan.

7. The method for predicting the remaining life of industrial equipment as described in claim 6, characterized in that, The second distillation loss function is constructed based on simple samples of the time series of predicted values ​​output by teachers and the time series of predicted values ​​output by the student network model, including: in, Let represent the second distillation loss function, j represent the index of the simple sample, g represent the number of simple samples, and 1 ≤ j ≤ g. This represents the j-th simple sample in the predicted values ​​output by the student network model. This represents the j-th simple sample in the teacher's output prediction.

8. The method for predicting the remaining life of industrial equipment as described in claim 7, characterized in that, The method for obtaining the third distillation loss function based on the first and second distillation loss functions includes: in, This represents the loss function of the third distillation.

9. The method for predicting the remaining life of industrial equipment as described in claim 8, characterized in that, The total loss function, derived from the third distillation loss function and the loss function between the student network model's predicted and actual values, includes: in, Represents the total loss function. Indicates the loss weight, 0 < <1, This represents the loss function between the student network model's predicted values ​​and the actual values.

10. A computer-readable storage medium, characterized in that, The medium stores a program that can be loaded by a processor and executed as described in any one of claims 1 to 9 for predicting the remaining life of industrial equipment.

Images