Fuzzy dynamic reasoning method, device and equipment for industrial time series

Abstract

The embodiment of the invention provides a fuzzy dynamic reasoning method, device and equipment for an industrial time sequence. The method is applied to artificial intelligence and industrial control fields. The method comprises the following steps: preprocessing pre-collected industrial time sequence data of target industrial equipment to obtain industrial time sequence sample data, inputting the industrial time sequence sample data into a fuzzy dynamic network obtained by pre-training, and predicting to obtain the residual life of the target industrial equipment. By means of the method, the problem that uncertainty in industrial time series data is difficult to process is solved, the problem that industrial edge scene computing resources are insufficient is solved, and therefore prediction accuracy is improved.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and industrial control, and particularly to a fuzzy dynamic reasoning method, device, and equipment for industrial time series. Background Art

[0002] In complex industrial scenarios, during the operation of equipment, affected by the long-term internal degradation factors of the equipment itself and external fluctuating factors such as load, temperature, and sensor errors, the time series data collected inevitably has a large amount of uncertainty. The data-driven deep learning method is a completely deterministic representation and cannot handle the uncertainty in the data, seriously hindering its deployment in complex industrial scenarios.

[0003] Industrial time series data is usually long-term and high-dimensional. Using complex deep neural networks will result in a significant increase in computational complexity. And in edge industrial scenarios, the computing resources are limited, and complex neural networks are difficult to meet the real-time requirements. Existing training methods for time series dynamic networks ignore the contributions of different difficulty samples to different exit predictors. The early exit behavior in the inference stage is ignored, resulting in a gap between training and testing, affecting the performance of the dynamic model.

[0004] In summary, how to handle the uncertainty of time series data in industrial scenarios is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] Embodiments of this application provide a fuzzy dynamic reasoning method, device, and equipment for industrial time series to solve the problem of how to handle the uncertainty of time series data in industrial scenarios.

[0006] In a first aspect, embodiments of this application provide a fuzzy dynamic reasoning method for industrial time series, including:

[0007] Preprocess the industrial time series data of the target industrial equipment collected in advance to obtain industrial time series sample data;

[0008] Input the industrial time series sample data into a pre-trained fuzzy dynamic network to predict the remaining life of the target industrial equipment. The fuzzy dynamic network includes a fuzzy logic representation layer, a sample information fusion layer, and a time series multi-exit dynamic network layer, and the multi-exit dynamic network layer includes multiple sub-networks.

[0009] In a possible implementation manner, the inputting the industrial time series sample data into a pre-trained fuzzy dynamic network to predict the remaining life of the target industrial equipment includes:

[0010] Performing fuzzy logic processing on the industrial time-series sample data through the fuzzy logic representation layer to obtain fuzzy membership information;

[0011] Concatenating the fuzzy membership information and the industrial time-series sample data to obtain concatenated information;

[0012] Processing the concatenated information through the sample information fusion layer to obtain hybrid features;

[0013] Performing inference on the hybrid features through the time-series multi-exit dynamic network layer to obtain the remaining life of the target industrial device.

[0014] In a possible implementation manner, the performing inference on the hybrid features through the time-series multi-exit dynamic network layer to obtain the remaining life of the target industrial device includes:

[0015] Step a: Inputting the hybrid features into the first-level sub-network, extracting first-level deep features through feature extraction, and generating a first-level prediction value and a first-level confidence level;

[0016] Step b: If the first-level confidence level is less than a preset threshold, inputting the first-level deep features into the second-level sub-network to obtain second-level deep features, superimposing the first-level deep features weighted by a preset attenuation coefficient and the second-level deep features to generate second-level fusion features, and generating a second-level prediction value and a second-level confidence level based on the second-level fusion features;

[0017] Step c: Repeating step b until a prediction value corresponding to a sub-network with a confidence level greater than the preset threshold is obtained, and determining the prediction value as the remaining life of the target industrial device.

[0018] In a possible implementation manner, the processing the concatenated information through the sample information fusion layer to obtain hybrid features includes:

[0019] Successively performing mean trend extraction, peak trend extraction, and cross-channel dependence extraction on the concatenated information in the time dimension, and fusing the extracted features to obtain the hybrid features.

[0020] In a possible implementation manner, each sub-network includes a feature extraction layer, a confidence level predictor, and a regression predictor. The feature extraction layer of the first-level sub-network is a one-dimensional convolutional neural network, and the feature extraction layers of other sub-networks except the first-level sub-network are Transformer encoders.

[0021] In a possible implementation manner, the method further includes:

[0022] The preset initial fuzzy dynamic network is trained based on pre-acquired training data to obtain the fuzzy dynamic network, and the training data includes multiple industrial time series data.

[0023] In a possible implementation manner, the training of the preset initial fuzzy dynamic network based on pre-acquired training data to obtain the fuzzy dynamic network includes:

[0024] According to the training data, the outputs of the shallow sub-network and the deep sub-network are weighted and superimposed, and the network parameters of the initial fuzzy dynamic network are iteratively updated based on a preset gradient redistribution rule until a preset number of iterations is reached, to obtain the fuzzy dynamic network;

[0025] Wherein, the gradient redistribution rule includes: blocking the gradient transmission of the shallow sub-network during backpropagation, for the shared parameters of the nth-level sub-network, the gradient of the shared parameters is jointly calculated by the prediction losses of the nth-level to the last-level sub-networks, and the gradient value is weighted and averaged according to the number of sub-networks participating in the calculation, and n is an integer greater than or equal to 1.

[0026] In a second aspect, an embodiment of the present application provides a fuzzy dynamic inference device for industrial time series, including:

[0027] A preprocessing module preprocesses the industrial time series data of the target industrial equipment collected in advance to obtain industrial time series sample data;

[0028] A prediction module is configured to input the industrial time series sample data into a pre-trained fuzzy dynamic network to predict the remaining life of the target industrial equipment, and the fuzzy dynamic network includes a fuzzy logic representation layer, a sample information fusion layer, and a time series multi-output dynamic network layer, and the multi-output dynamic network layer includes multiple sub-networks.

[0029] In a possible implementation manner, the prediction module is configured to:

[0030] Perform fuzzy logic processing on the industrial time series sample data through the fuzzy logic representation layer to obtain fuzzy membership information;

[0031] Concatenate the fuzzy membership information and the industrial time series sample data to obtain concatenated information;

[0032] Process the concatenated information through the sample information fusion layer to obtain a mixed feature;

[0033] Perform inference on the mixed feature through the time series multi-output dynamic network layer to obtain the remaining life of the target industrial equipment.

[0034] In a possible implementation manner, the prediction module infers the remaining life of the target industrial device through the time-series multi-exit dynamic network layer, specifically including:

[0035] Step a: Input the mixed features into the first-level sub-network, extract the first-level deep features through feature extraction, and generate a first-level prediction value and a first-level confidence level.

[0036] Step b: If the first-level confidence level is less than a preset threshold, input the first-level deep features into the second-level sub-network to obtain second-level deep features, superimpose the first-level deep features weighted by a preset attenuation coefficient and the second-level deep features to generate second-level fusion features, and generate a second-level prediction value and a second-level confidence level based on the second-level fusion features.

[0037] Step c: Repeat Step b until a prediction value corresponding to a sub-network with a confidence level greater than the preset threshold is obtained, and determine the prediction value as the remaining life of the target industrial device.

[0038] In a possible implementation manner, the prediction module processes the splicing information through the sample information fusion layer to obtain mixed features, specifically including:

[0039] Extract the mean trend, peak trend, and cross-channel dependence in the time dimension of the splicing information in sequence, and fuse the extracted features to obtain the mixed features.

[0040] In a possible implementation manner, each sub-network includes a feature extraction layer, a confidence predictor, and a regression predictor. The feature extraction layer of the first-level sub-network is a one-dimensional convolutional neural network, and the feature extraction layers of other sub-networks except the first-level sub-network are Transformer encoders.

[0041] In a possible implementation manner, the device further includes:

[0042] A training module, configured to train a preset initial fuzzy dynamic network based on pre-acquired training data to obtain the fuzzy dynamic network, where the training data includes multiple industrial time series data.

[0043] In a possible implementation manner, the training module is specifically configured to:

[0044] According to the training data, weight and superimpose the outputs of the shallow sub-network and the deep sub-network, and iteratively update the network parameters of the initial fuzzy dynamic network based on a preset gradient reallocation rule until a preset number of iterations is reached to obtain the fuzzy dynamic network;

[0045] Among them, the gradient redistribution rule includes: blocking the gradient transmission of the shallow sub-network during backpropagation. For the shared parameters of the nth-level sub-network, the gradient of the shared parameters is jointly calculated by the prediction losses of the nth-level to the last-level sub-networks, and the gradient value is weighted and averaged according to the number of sub-networks participating in the calculation, where n is an integer greater than or equal to 1.

[0046] In a third aspect, an embodiment of the present application provides an electronic device, including: a memory, a processor;

[0047] The memory stores computer-executable instructions;

[0048] The processor executes the computer-executable instructions stored in the memory, so that the processor executes the above first aspect and / or various possible implementation manners of the first aspect.

[0049] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, in which computer-executable instructions are stored, and when the computer-executable instructions are executed by a processor, they are used to implement the above first aspect and / or various possible implementation manners of the first aspect.

[0050] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, and when the computer program is executed by a processor, it implements the above first aspect and / or various possible implementation manners of the first aspect.

[0051] The fuzzy dynamic inference method, device and equipment for industrial time series provided by the embodiments of the present application. Preprocess the industrial time series data of the target industrial equipment collected in advance to obtain industrial time series sample data, and input the industrial time series sample data into a pre-trained fuzzy dynamic network to predict the remaining life of the target industrial equipment. Through the above method, the problem of difficult processing of uncertainties in industrial time series data is solved, and the problem of insufficient computing resources in industrial edge scenarios is also solved, thereby improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The drawings here are incorporated into the description and constitute a part of this description, showing the embodiments consistent with the present application, and are used together with the description to explain the principles of the present application.

[0053] Figure 1 It is a schematic flowchart of the fuzzy dynamic inference method for industrial time series provided by the present application Figure 1 ;

[0054] Figure 2 It is a schematic diagram of the fuzzy dynamic network architecture;

[0055] Figure 3 It is a schematic flowchart of the fuzzy dynamic inference method for industrial time series provided by the present applicationFigure 2 ;

[0056] Figure 4 Flow schematic of the fuzzy dynamic inference method for industrial time series provided by this application Figure 3 ;

[0057] Figure 5 Schematic diagram for backpropagation gradient calculation;

[0058] Figure 6 Schematic diagram of the structure of the fuzzy dynamic inference device for industrial time series provided by this application;

[0059] Figure 7 Schematic diagram of the structure of the electronic device provided by this application.

[0060] Through the above-mentioned drawings, specific embodiments of this application have been shown, and there will be more detailed descriptions hereinafter. These drawings and textual descriptions are not intended to limit the scope of the concept of this application in any way, but to illustrate the concept of this application to those skilled in the art by referring to specific embodiments. Specific Embodiments

[0061] Here, exemplary embodiments will be described in detail, and their examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0062] In complex industrial scenarios, during the operation of equipment, affected by the long-term internal degradation factors of the equipment itself and external fluctuating factors such as load, temperature, and sensor errors, the time series data collected inevitably has a large amount of uncertainty. Deep learning methods based on data-driven are completely deterministic representations and cannot handle the uncertainty existing in the data, seriously hindering their deployment in complex industrial scenarios. Industrial time series data is usually long-term and high-dimensional. Using complex deep neural networks will lead to a significant increase in the amount of calculation. And in edge industrial scenarios, the computing resources are limited, and complex neural networks are difficult to meet the real-time requirements. Existing training methods for time series dynamic networks ignore the contributions of different difficulty samples to different exit predictors. The early exit behavior in the inference stage is ignored, resulting in a gap between training and testing and affecting the performance of the dynamic model. To sum up, how to handle the uncertainty of time series data in industrial scenarios is a technical problem that urgently needs to be solved in this field.

[0063] In view of the above problems, the present application provides a fuzzy dynamic reasoning method, device and equipment for industrial time series, which solves the problem that it is difficult to handle the uncertainty in industrial time series data. Specifically, there are inevitably a large amount of uncertainties in the currently collected time series, while the data-driven deep learning method is a completely deterministic representation and cannot handle the uncertainties existing in the data, which seriously hinders its deployment in complex industrial scenarios. Industrial time series are usually long-term and high-dimensional. Using complex deep neural networks such as Transformer will lead to a significant increase in the amount of calculation. And the computing resources in edge industrial scenarios are limited, and it is difficult for complex neural networks to meet the real-time requirements. In addition, the existing training methods for time series dynamic networks ignore the contributions of different difficulty samples to different exit predictors. The early exit behavior in the inference stage is ignored, resulting in a gap between training and testing and affecting the performance of the dynamic model. Considering these problems, it is studied whether the fuzzy learning method can be introduced into the deep learning model to solve the problems of insufficient computing resources and uncertainty in industrial edge scenarios, and to obtain the fuzzy feature information in the fuzzy membership degree to supplement the sample information, making the sample information representation more comprehensive and solving the problem that it is difficult to handle the uncertainty in industrial time series data. Then, by jointly training regression predictors, weighting the outputs of shallow subnetworks and deep subnetworks, and redistributing subnetwork gradients, etc., the gap between the training and inference of the time series multi-exit network is bridged. Based on this, the solution of the present application is proposed.

[0064] The following uses specific embodiments to elaborate in detail on the technical solution of the present application and how the technical solution of the present application solves the above technical problems. These several specific embodiments below can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0065] Figure 1 Flow schematic of the fuzzy dynamic reasoning method for industrial time series provided by the present application Figure 1 , as Figure 1 shown, the method includes:

[0066] S101: Preprocess the industrial time series data of the target industrial equipment collected in advance to obtain industrial time series sample data.

[0067] In this step, in order to accurately predict the remaining life of the target industrial equipment, the industrial time series data of the target industrial equipment can be collected in advance and the data can be preprocessed.

[0068] Exemplarily, clearly identify the types of industrial equipment for which the remaining life needs to be predicted, such as mechanical components, electronic components, etc. According to the equipment type and performance indicators, select appropriate sensors and data acquisition devices. Ensure the accuracy and stability of the data acquisition devices to obtain high-quality time series data. During the normal operation of the equipment, continuously collect time series data. For example, record the operating time, load conditions, environmental parameters, etc. of the equipment for subsequent analysis.

[0069] The preprocessing of the data can include:

[0070] 1. Data cleaning:

[0071] Identify and remove outliers: Check for outliers in the data that may significantly deviate from the normal range. These values may be caused by sensor failures, data transmission errors, etc. For outliers, options include deletion, replacement, or smoothing.

[0072] Fill in missing values: For missing data points, interpolation methods (such as linear interpolation, spline interpolation, etc.) can be used for filling, or estimated based on the operating state and trend of the equipment.

[0073] 2. Data transformation:

[0074] Normalization: Convert data with different dimensions to the same scale to improve the convergence speed and prediction accuracy of the model.

[0075] Feature extraction: Extract meaningful features from the original data, such as statistical quantities like mean, standard deviation, kurtosis, skewness, and frequency domain features (such as spectral analysis).

[0076] 3. Data partitioning:

[0077] Partition the dataset into a training set, a validation set, and a test set. The training set is used to construct and train the prediction model, the validation set is used to adjust the model parameters and optimize the model performance, and the test set is used to evaluate the final prediction accuracy of the model.

[0078] S102: Input the industrial time series sample data into the pre-trained fuzzy dynamic network to predict the remaining life of the target industrial equipment.

[0079] In this step, after obtaining the industrial time series sample data, input this data into the pre-trained fuzzy dynamic network, and then predict the remaining life of the target industrial equipment.

[0080] Specifically, during the operation of the device, affected by its own inherent long-term degradation factors and external factors such as load, temperature, and sensor errors, the collected time series inevitably has a large amount of uncertainty. Therefore, the fuzzy dynamic network has a fuzzy logic representation layer, a sample information fusion layer, and a time-series multi-output dynamic network layer. The multi-output dynamic network layer includes multiple sub-networks.

[0081] Exemplarily, Figure 2 is a schematic diagram of the fuzzy dynamic network architecture, as Figure 2 shown. The fuzzy logic representation layer consists of different fuzzy membership functions. The time-series multi-output dynamic neural network layer consists of 4 sub-networks. Each sub-network includes a feature extraction layer, a confidence predictor, a regression predictor, etc. The feature extraction layer used by the first-level sub-network is a one-dimensional convolutional neural network, and the remaining sub-networks use Transformer encoders. First, the sample data enters the fuzzy logic representation layer to obtain fuzzy membership information. Then, the fuzzy membership information is concatenated with the original sample and enters the sample information fusion layer. Through information adaptive fusion, hybrid feature information containing fuzzy features is obtained, enhancing the ability of the deep neural network to process uncertainty. Finally, the hybrid feature information enters the time-series multi-output dynamic network for inference. According to the sample difficulty, the depth of the network to be output is determined, and then the predicted value is obtained.

[0082] The fuzzy dynamic inference method for industrial time series provided by the embodiments of this application preprocesses the industrial time series data of the target industrial device collected in advance to obtain industrial time series sample data, and inputs the industrial time series sample data into the pre-trained fuzzy dynamic network to predict the remaining life of the target industrial device. Through the above method, the problem of difficult processing of uncertainty in industrial time series data is solved, and the problem of insufficient computing resources in industrial edge scenarios is also solved, thereby improving the prediction accuracy.

[0083] Figure 3 is a flowchart of the fuzzy dynamic inference method for industrial time series provided by this application Figure 2 as Figure 3 shown. On the basis of the above embodiments, step S102 specifically includes:

[0084] S301: Perform fuzzy logic processing on the industrial time series sample data through the fuzzy logic representation layer to obtain fuzzy membership information.

[0085] In this step, through fuzzy feature extraction technology, the fuzzy feature information in the fuzzy membership is obtained, and the fuzzy feature representation is concatenated with the original sample data to make the information representation more comprehensive, solving the problem of difficult processing of uncertainty in industrial time series data.

[0086] Exemplarily, as Figure 2As shown, for each sensor data in the industrial time-series sample data, the fuzzy logic representation layer processes the sensor data through the preset fuzzy membership function corresponding to the sensor to obtain the result corresponding to the sensor, and combines the results corresponding to each sensor to obtain the fuzzy membership information corresponding to the industrial time-series sample data.

[0087] First, the input industrial time-series sample data is decomposed dimension by dimension. Here, L represents the time length of the sequence, and K represents the dimension of K sensor data. Each sensor feature dimension is processed separately. For each sensor dimension , an independent fuzzy membership function is set for it. The fuzzy membership function for each dimension calculates the degree to which the input data belongs to the fuzzy set of the current dimension, and maps the input data of this dimension to the fuzzy membership interval of [0, 1].

[0088] The fuzzy logic representation layer uses a Gaussian membership function with fuzzy mean and fuzzy variance as parameters. The fuzzy mean represents the core trend of the sensor dimension data and reflects the typical value or expected value of the sensor. The fuzzy variance quantifies the degree of dispersion of the data around the fuzzy mean, thus reflecting the uncertainty of the data. Through the fuzzy mean and fuzzy variance, the uncertainty characteristics of the sensor measurement can be effectively captured, so as to more accurately reflect the operating state of the industrial equipment. The Gaussian membership function is as follows:

[0089]

[0090] The results of the membership functions for each sensor dimension are combined into fuzzy features , which can be described as:

[0091]

[0092] The fuzzy feature , that is, the final expression of the fuzzy membership information is:

[0093]

[0094] The fuzzy membership obtained through the Gaussian membership function can effectively capture the uncertainty in the input data and provide effective fuzzy feature information for subsequent sample information fusion.

[0095] S302: Concatenate the fuzzy membership information and the industrial time-series sample data to obtain concatenated information.

[0096] In this step, in order to obtain enhanced features, the fuzzy membership information and the industrial time-series sample data can be concatenated.

[0097] Specifically, it can be concatenated along the feature dimension (i.e., the column direction) so that the feature vector at each time point contains both the original sensor data and the fuzzy membership information.

[0098] S303: Process the concatenated information through the sample information fusion layer to obtain the mixed features.

[0099] In this step, the concatenated information obtained in the above step is input into the sample information fusion layer for processing, so as to obtain the mixed features containing fuzzy feature information.

[0100] Specifically, perform mean trend extraction, peak trend extraction, and cross-channel dependence extraction on the concatenated information in the time dimension in sequence, and fuse the extracted features to obtain the mixed features.

[0101] Exemplarily, combined with Figure 2 , the sample information fusion layer captures the mean trend and peak trend of the concatenated information in the time dimension, and combines the correlation information between channels to realize feature extraction and information fusion of the concatenated information . The sample information fusion layer uses average pooling to extract the average trend feature, which can be expressed as:

[0102]

[0103] where K represents the window size.

[0104] The sample information fusion layer uses max pooling to capture significant local changes and emphasizes the local extrema in the time series data. Its representation method is as follows:

[0105]

[0106] where K represents the window size.

[0107] On the basis of extracting time features, the sample information fusion layer uses one-dimensional convolution to further capture the correlation between sensor channels. The convolutional kernel propagates along the time dimension and fuses the data of all sensor channels at the same time to extract the cross-channel dependence. This can be expressed as:

[0108]

[0109] where, is the concatenated information, and represent the weights and biases of the convolutional kernel, and represent dimensions.

[0110] Finally, all the feature information is concatenated along the feature dimension and further fused through the fully connected layer to obtain the mixed features The final output can be described as:

[0111]

[0112] where represents the concatenation operation, represents the linear layer.

[0113] By further fusing the sample and fuzzy membership information, it can provide strong support for handling uncertain factors in subsequent industrial time series prediction.

[0114] S304: Infer the remaining life of the target industrial equipment through the time series multi - exit dynamic network layer for the mixed features.

[0115] In this step, in order to achieve accurate prediction of the remaining life of industrial equipment, while balancing the computational efficiency and model performance, and adapting to the resource constraints of the edge industrial scenario, the mixed features can be adaptively inferred through the time series multi - exit dynamic network layer.

[0116] Among them, each sub - network of the time series multi - exit dynamic network layer includes a feature extraction layer, a confidence predictor, and a regression predictor. The feature extraction layer of the first - level sub - network is a one - dimensional convolutional neural network, and the feature extraction layers of other sub - networks except the first - level sub - network are Transformer encoders.

[0117] Specifically, input the mixed features into the first - level sub - network, obtain the first - level deep features through feature extraction, and generate the first - level prediction value and the first - level confidence. If the first - level confidence is less than the preset threshold, input the first - level deep features into the second - level sub - network to obtain the second - level deep features. Weight the first - level deep features by the preset attenuation coefficient and superimpose them with the second - level deep features to generate the second - level fusion features, and generate the second - level prediction value and the second - level confidence based on the second - level fusion features. Repeat the above operations until the prediction value corresponding to the sub - network with a confidence greater than the preset threshold is obtained, and determine the prediction value as the remaining life of the target industrial equipment.

[0118] The fuzzy dynamic inference method for industrial time series provided by the embodiments of the present application performs fuzzy logic processing on industrial time series sample data through the fuzzy logic representation layer to obtain fuzzy membership information, splices the fuzzy membership information and industrial time series sample data to obtain splicing information, processes the splicing information through the sample information fusion layer to obtain mixed features, and infers the mixed features through the time series multi - exit dynamic network layer to obtain the remaining life of the target industrial equipment. Through the above method, the uncertainty data modeling ability is improved, lightweight dynamic inference is realized, and the multi - scale feature representation ability is enhanced.

[0119] Figure 4Flow schematic of the fuzzy dynamic inference method for industrial time series provided by this application Figure 3 , such as Figure 4 shown. Based on the above embodiment, step S304 specifically includes:

[0120] S401: Input the mixed features into the first-level sub-network, obtain the first-level deep features through feature extraction, and generate the first-level prediction value and the first-level confidence level.

[0121] S402: If the first-level confidence level is less than the preset threshold, input the first-level deep features into the second-level sub-network to obtain the second-level deep features, superimpose the first-level deep features weighted by the preset attenuation coefficient and the second-level deep features to generate the second-level fusion features, and generate the second-level prediction value and the second-level confidence level based on the second-level fusion features.

[0122] S403: Repeat the above step S402 until the prediction value corresponding to the sub-network with a confidence level greater than the preset threshold is obtained, and determine the prediction value as the remaining life of the target industrial equipment.

[0123] After the mixed features are input into the time series multi-output dynamic network layer, they are first processed by the first-level sub-network. The one-dimensional convolutional neural network of the first-level sub-network performs feature extraction on the mixed features and outputs the first-level deep features. Specifically, the design of the one-dimensional convolutional neural network is that the convolutional kernel slides along the time dimension to capture local time series patterns (such as short-term fluctuations of vibration signals).

[0124] The regression predictor 1 performs a fully connected mapping on the first-level deep features and outputs the first-level prediction value. When calculating the confidence level through the confidence level predictor, optionally, the confidence level can be estimated based on the prediction residual or variance.

[0125] Furthermore, compare the confidence level of the first-level sub-network with the preset threshold. If the confidence level of the first-level sub-network is less than the preset threshold, trigger to enter the second sub-network, input the first-level deep features into the second sub-network for feature extraction to obtain the second-level deep features, then weight the first-level deep features by the attenuation coefficient, superimpose the weighted first-level deep features and the second-level deep features, and through the prediction of the regression predictor 2, obtain the second-level prediction value, calculate the confidence level corresponding to the second-level sub-network, compare the confidence level of the second-level sub-network with the preset threshold. If the confidence level of the second-level sub-network is less than the preset threshold, trigger to enter the third-level sub-network, repeat the above steps until the obtained confidence level is greater than the preset threshold, and determine the prediction value of the sub-network corresponding to the confidence level as the remaining life of the target industrial equipment.

[0126] The fuzzy dynamic inference method for industrial time series provided by the embodiments of the present application inputs mixed features into the first-level sub-network, obtains first-level deep features through feature extraction, and generates a first-level prediction value and a first-level confidence level. If the first-level confidence level is less than a preset threshold, the first-level deep features are input into the second-level sub-network to obtain second-level deep features. The first-level deep features are weighted by a preset attenuation coefficient and then superimposed with the second-level deep features to generate second-level fusion features, and a second-level prediction value and a second-level confidence level are generated based on the second-level fusion features. The above operations are repeated until a prediction value corresponding to a sub-network with a confidence level greater than the preset threshold is obtained, and the prediction value is determined as the remaining life of the target industrial device. Through the dynamic inference mechanism of multiple-level sub-networks, the above method combines confidence threshold judgment and hierarchical feature fusion to achieve an adaptive accuracy-efficiency balance for predicting the remaining life of industrial devices: the lightweight shallow sub-network (1D-CNN) quickly processes most simple samples and generates initial predictions. When the confidence level is insufficient, the deep sub-network (Transformer) is triggered to perform weighted residual correction on the features, gradually fusing local and global time series patterns. Finally, on the premise of ensuring prediction accuracy, the computational complexity is significantly reduced. At the same time, the uncertainty of industrial data is effectively addressed through fuzzy features and confidence quantification, providing a highly robust and low-latency life prediction solution for edge devices.

[0127] Based on the above embodiments, the method further includes training a preset initial fuzzy dynamic network based on pre-obtained training data to obtain a fuzzy dynamic network. The specific process of training includes: according to the training data, weighting and superimposing the outputs of the shallow sub-network and the deep sub-network, and iteratively updating the network parameters of the initial fuzzy dynamic network based on a preset gradient reallocation rule until a preset number of iterations is reached to obtain the fuzzy dynamic network;

[0128] Among them, the gradient reallocation rule includes: blocking the gradient transmission of the shallow sub-network during backpropagation. For the shared parameters of the nth-level sub-network, the gradient of the shared parameters is jointly calculated by the prediction losses of the nth-level to the last-level sub-networks, and the gradient value is weighted and averaged according to the number of sub-networks participating in the calculation, where n is an integer greater than or equal to 1.

[0129] Specifically, in the training method of traditional time series dynamic networks, the model optimizes all exits simultaneously during training, that is, given a batch of training samples, the regression predictors at all levels calculate losses and update parameters on the same batch of data. This is likely to lead to the problem of training-inference mismatch: in the inference stage, if the input sample is "simple", the network often exits at the previous shallow layer, while the deep classifier will almost only face "difficult" samples; but in the training stage, all regression predictors have seen all samples, resulting in the parameter distribution of the deep predictor being inconsistent with that during actual inference.

[0130] Different from traditional gradient boosting, which trains the first predictor until complete convergence and then trains the next predictor while fixing the predictors in front, the fuzzy dynamic network provided by the embodiments of the present application does not adopt a completely sequential training method. Most of the feature extraction layers are shared among all predictors. If the shallow layers are fixed first and then the deep layers are trained, the parameter coordination will be poor; moreover, separate stages will bring a large number of repeated iterations, resulting in a large overhead. Therefore, the outputs of all predictors are calculated uniformly, and then the losses of each predictor are summed up, and then the parameters are updated. The loss function L is:

[0131]

[0132] where x is the input sample, RMSE is the Root-mean-square error, Fn is the output of the nth predictor, and y is the true value label.

[0133] During the training process, all predictors will process the complete training samples, which means that the deep predictors will be affected by a large number of "simple" training samples and will not receive effective gradients for updating, resulting in poor learning effects for the deep layers. To solve this problem, output weighting paths are added between different sub-networks. As Figure 2 shown, the output of the second-level sub-network Transformer encoder 2 in the network is weighted with the output of the first-level sub-network one-dimensional convolutional neural network to obtain . The predicted value is obtained through the regression predictor 2 for the weighted . This can be understood as the second-level sub-network compensating for the prediction of the first-level sub-network on the basis of the first-level sub-network to obtain a more accurate prediction output. Similarly, the output of the ith sub-network of this model is regarded as: the sum of the outputs of all previous sub-networks plus the output of the current sub-network i:

[0134]

[0135] where is the weighting coefficient, used to weaken the accumulated output in front to prevent the shallow sub-networks from causing too much interference to the deep sub-networks. is set to a constant less than 1 to weaken the effective output of the front predictors, increase the prediction difficulty of the deep predictors, and thus enhance the training of the parameters of the deep predictors. It should be noted that during the training process the backpropagation link of is disabled.

[0136] In addition, during training, if the losses of all classifiers are directly backpropagated to the shared sub-network, it may cause excessive accumulation of gradients in the shallow layers, which is not conducive to the stable update of the early sub-network. Figure 5 Schematic diagram of backpropagation gradient calculation, as Figure 5 shown. During backpropagation, the gradients at each exit are weighted. The specific formula is expressed as:

[0137]

[0138] where is the total gradient of the nth sub-network, represents the RMSE loss of the ith exit, represents the gradient contribution of the loss of the ith exit to the nth sub-network. N - n + 1 is the number of exits from the nth exit to the Nth exit (a total of N exits), which is used to weight the gradients of each exit.

[0139] For the nth sub-network, since it participates in all exits from n to N, its gradient accumulation is . If this gradient accumulation is directly used, when N is large, the sub-network in the shallow layer may experience gradient explosion. The weighting factor functions to average all exits that contribute to this sub-network, so that the magnitude of its gradient remains within a controllable range.

[0140] The fuzzy dynamic inference method for industrial time series provided by the embodiments of the present application makes up for the gap between the training and inference of the time series multi-exit network by methods such as jointly training regression predictors, weighting the outputs of the shallow sub-network and the deep sub-network, and redistributing the sub-network gradients.

[0141] Figure 6 Schematic diagram of the structure of the fuzzy dynamic inference device for industrial time series provided by the present application, as Figure 6 shown. The fuzzy dynamic inference device 600 provided in this embodiment includes:

[0142] A preprocessing module 601, configured to preprocess the industrial time series data of the target industrial device collected in advance to obtain industrial time series sample data.

[0143] A prediction module 602, configured to input the industrial time series sample data into a pre-trained fuzzy dynamic network to predict the remaining life of the target industrial device. The fuzzy dynamic network includes a fuzzy logic representation layer, a sample information fusion layer, and a time series multi-exit dynamic network layer. The multi-exit dynamic network layer includes multiple sub-networks.

[0144] In a possible implementation manner, the prediction module 602 is specifically configured to;

[0145] Perform fuzzy logic processing on industrial time-series sample data through the fuzzy logic representation layer to obtain fuzzy membership information;

[0146] Concatenate the fuzzy membership information and the industrial time-series sample data to obtain concatenated information;

[0147] Process the concatenated information through the sample information fusion layer to obtain hybrid features;

[0148] Perform inference on the hybrid features through the time-series multi-exit dynamic network layer to obtain the remaining life of the target industrial equipment.

[0149] In a possible implementation, the prediction module 602 performs inference on the hybrid features through the time-series multi-exit dynamic network layer to obtain the remaining life of the target industrial equipment, specifically including:

[0150] Step a: Input the hybrid features into the first-level sub-network, extract the first-level deep features through feature extraction, and generate a first-level prediction value and a first-level confidence level;

[0151] Step b: If the first-level confidence level is less than the preset threshold, input the first-level deep features into the second-level sub-network to obtain the second-level deep features, weight the first-level deep features according to the preset attenuation coefficient and superimpose them with the second-level deep features to generate the second-level fusion features, and generate a second-level prediction value and a second-level confidence level based on the second-level fusion features;

[0152] Step c: Repeat step b until the prediction value corresponding to the sub-network with a confidence level greater than the preset threshold is obtained, and determine the prediction value as the remaining life of the target industrial equipment.

[0153] In a possible implementation, the prediction module 602 processes the concatenated information through the sample information fusion layer to obtain hybrid features, specifically including:

[0154] Successively perform mean trend extraction, peak trend extraction, and cross-channel dependence extraction on the concatenated information in the time dimension, and fuse the extracted features to obtain hybrid features.

[0155] In a possible implementation, each sub-network includes a feature extraction layer, a confidence predictor, and a regression predictor. The feature extraction layer of the first-level sub-network is a one-dimensional convolutional neural network, and the feature extraction layer of other sub-networks except the first-level sub-network is a Transformer encoder.

[0156] In a possible implementation, the fuzzy dynamic inference device 600 for industrial time series further includes:

[0157] A training module 603, configured to train a preset initial fuzzy dynamic network based on pre-acquired training data to obtain a fuzzy dynamic network, where the training data includes multiple industrial time series data.

[0158] In a possible implementation, the training module 603 is specifically configured to:

[0159] According to the training data, by weighted superposition of the outputs of the shallow sub-network and the deep sub-network, and iteratively updating the network parameters of the initial fuzzy dynamic network based on a preset gradient redistribution rule until a preset number of iterations is reached, to obtain a fuzzy dynamic network;

[0160] Among them, the gradient redistribution rule includes: blocking the gradient transmission of the shallow sub-network during backpropagation. For the shared parameters of the nth-level sub-network, the gradient of the shared parameters is jointly calculated by the prediction losses of the nth-level to the last-level sub-networks, and the gradient value is weighted and averaged according to the number of sub-networks participating in the calculation, where n is an integer greater than or equal to 1.

[0161] The fuzzy dynamic inference device for industrial time series provided in this embodiment can execute the fuzzy dynamic inference method for industrial time series provided in the above method embodiment. The implementation principle and technical effect are similar, and will not be elaborated here in this embodiment.

[0162] Figure 7 It is a schematic structural diagram of an electronic device provided in this application. As Figure 7 shown, the electronic device 700 provided in this embodiment includes: at least one processor 701 and a memory 702. Optionally, the electronic device 700 further includes a communication component 703. Among them, the processor 701, the memory 702, and the communication component 703 are connected through a bus 704.

[0163] In a specific implementation process, at least one processor 701 executes computer execution instructions stored in the memory 702, so that at least one processor 701 executes the above method.

[0164] The specific implementation process of the processor 701 can be referred to in the above method embodiment. The implementation principle and technical effect are similar, and will not be elaborated here in this embodiment.

[0165] In the above embodiments, it should be understood that the processor may be a central processing unit (CPU for short), or other general-purpose processors, digital signal processors (DSP for short), application specific integrated circuits (ASIC for short), etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the method disclosed in combination with the invention can be directly implemented by the execution of the hardware processor, or by the combination of hardware and software modules in the processor.

[0166] The memory may include a random access memory (RAM), and may also include a non-volatile memory (NVM), such as at least one disk memory.

[0167] The bus may be an industry standard architecture (ISA) bus, a peripheral component interconnect (PCI) bus, an extended industry standard architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, the bus in the drawings of this application is not limited to only one bus or one type of bus.

[0168] This application also provides a computer program product, including a computer program, which implements the above method when executed by a processor.

[0169] This application also provides a computer-readable storage medium, in which computer-executable instructions are stored, and when the processor executes the computer-executable instructions, the above method is implemented.

[0170] The above-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a disk or an optical disc. The readable storage medium can be any available medium accessible by a general-purpose or special-purpose computer.

[0171] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be part of the processor. The processor and the readable storage medium can be located in an Application Specific Integrated Circuit (ASIC). Of course, the processor and the readable storage medium can also exist as discrete components in a device.

[0172] The division of units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Additionally, the couplings or direct couplings or communication connections shown or discussed among each other can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0173] The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0174] Furthermore, in each embodiment of the present invention, the functional units can be integrated in one processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit.

[0175] 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 this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or 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 in each embodiment of the present invention. The aforementioned storage medium includes: 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 and other various media that can store program codes.

[0176] Those of ordinary skill in the art will understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments; and the foregoing storage medium includes: various media such as ROM, RAM, magnetic disk or optical disc that can store program codes.

[0177] Finally, it should be noted that: After considering the specification and practicing the invention disclosed herein, those skilled in the art will readily conceive of other embodiments of the present invention. The present invention is intended to cover any variations, uses, or adaptations of the present invention, which follow the general principles of the present invention and include known common knowledge or conventional technical means in the technical field not disclosed by the present invention. It is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present invention is only limited by the appended claims.

Claims

1. A fuzzy dynamic reasoning method for industrial time series, characterized in that Including: Preprocessing the industrial time series data of the pre-collected target industrial equipment to obtain industrial time series sample data; Inputting the industrial time series sample data into a pre-trained fuzzy dynamic network to predict the remaining life of the target industrial equipment. The fuzzy dynamic network includes a fuzzy logic representation layer, a sample information fusion layer, and a time series multi-output dynamic network layer. The multi-output dynamic network layer includes multiple sub-networks.

2. The method according to claim 1, wherein The step of inputting the industrial time series sample data into a pre-trained fuzzy dynamic network to predict the remaining life of the target industrial equipment includes: Performing fuzzy logic processing on the industrial time series sample data through the fuzzy logic representation layer to obtain fuzzy membership information; Concatenating the fuzzy membership information and the industrial time series sample data to obtain concatenated information; Processing the concatenated information through the sample information fusion layer to obtain a mixed feature; Inferring the remaining life of the target industrial equipment through the time series multi-output dynamic network layer for the mixed feature.

3. The method according to claim 2, wherein The step of inferring the remaining life of the target industrial equipment through the time series multi-output dynamic network layer for the mixed feature includes: Step a: Inputting the mixed feature into the first-level sub-network, extracting first-level deep features through feature extraction, and generating a first-level prediction value and a first-level confidence level; Step b: If the first-level confidence level is less than a preset threshold, inputting the first-level deep feature into the second-level sub-network to obtain a second-level deep feature, weighting the first-level deep feature by a preset attenuation coefficient and superimposing it with the second-level deep feature to generate a second-level fusion feature, and generating a second-level prediction value and a second-level confidence level based on the second-level fusion feature; Step c: Repeating step b until obtaining the prediction value corresponding to the sub-network with a confidence level greater than the preset threshold, and determining the prediction value as the remaining life of the target industrial equipment.

4. The method according to claim 2, wherein The step of processing the concatenated information through the sample information fusion layer to obtain a mixed feature includes: Successively extracting the mean trend, peak trend, and cross-channel dependence in the time dimension for the concatenated information, and fusing the extracted features to obtain the mixed feature.

5. The method according to claim 1, characterized in that Each sub-network includes a feature extraction layer, a confidence predictor, and a regression predictor. The feature extraction layer of the first-level sub-network is a one-dimensional convolutional neural network, and the feature extraction layers of other sub-networks except the first-level sub-network are Transformer encoders.

6. The method according to claim 1, wherein The method further includes: Training a preset initial fuzzy dynamic network based on pre-obtained training data to obtain the fuzzy dynamic network. The training data includes multiple industrial time series data.

7. The method according to claim 6, characterized in that, The step of training a preset initial fuzzy dynamic network based on pre-obtained training data to obtain the fuzzy dynamic network includes: According to the training data, iteratively updating the network parameters of the initial fuzzy dynamic network by weighted superposition of the outputs of the shallow sub-network and the deep sub-network and based on a preset gradient redistribution rule until reaching a preset number of iterations to obtain the fuzzy dynamic network; Among them, the gradient reallocation rule includes: blocking the gradient transfer of the shallow sub-network during backpropagation. For the shared parameters of the nth-level sub-network, the gradient of the shared parameters is jointly calculated by the prediction losses of the nth-level to the last-level sub-networks, and the gradient value is weighted and averaged according to the number of sub-networks participating in the calculation, where n is an integer greater than or equal to 1.

8. A fuzzy dynamic inference device for industrial time series, characterized in that including: A preprocessing module that preprocesses the industrial time series data of the target industrial device collected in advance to obtain industrial time series sample data; A prediction module for inputting the industrial time series sample data into a pre-trained fuzzy dynamic network to predict the remaining life of the target industrial device. The fuzzy dynamic network includes a fuzzy logic representation layer, a sample information fusion layer, and a time series multi-output dynamic network layer, and the multi-output dynamic network layer includes multiple sub-networks.

9. An electronic device, characterized in that, including: A memory and a processor; The memory stores computer execution instructions; The processor executes the computer execution instructions stored in the memory, so that the processor executes the fuzzy dynamic inference method for industrial time series according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Computer execution instructions are stored in the computer-readable storage medium, and when the computer execution instructions are executed by a processor, they are used to implement the fuzzy dynamic inference method for industrial time series according to any one of claims 1 to 7.

Images