CN116541680B

CN116541680B - Industrial equipment state prediction method and device based on light-weight time sequence reduction

Info

Publication number: CN116541680B
Application number: CN202310819842.9A
Authority: CN
Inventors: 任磊; 王海腾
Original assignee: Beihang University
Current assignee: Beihang University
Priority date: 2023-07-06
Filing date: 2023-07-06
Publication date: 2023-09-12
Anticipated expiration: 2043-07-06
Also published as: CN116541680A

Abstract

The embodiment of the application provides an industrial equipment state prediction method and device based on light-weight time sequence reduction, wherein the method comprises the following steps: acquiring operation data of an engine; inputting the operation data to a first neuron group in a group linear transformation embedded layer of a prediction model to perform linear transformation processing on the operation data to obtain first data; inputting the first data to an attention layer to obtain second data; inputting the second data to a time sequence reduction layer for importance scoring, and removing the data with the importance scoring lower than the first preset value from the second data to obtain third data; inputting the third data into a group feedforward network layer for linear transformation processing to obtain fourth data; and inputting the fourth data into a learning layer of the prediction model, and learning the fourth data according to a multi-level learning mechanism strategy in the learning layer to obtain the residual service life of the engine. In this way, the speed at which the predictive model processes data can be increased.

Description

Industrial equipment state prediction method and device based on light-weight time sequence reduction

Technical Field

The application relates to the technical field of equipment state prediction, in particular to an industrial equipment state prediction method and device based on light-weight time sequence reduction.

Background

In an industrial scenario, an engine of an industrial device may wear during use, etc., and may have an impact on the remaining life of the engine. Therefore, during use of the engine, a prediction of the remaining life of the engine is required.

Currently, the remaining life of an engine is generally predicted using a predictive model that includes a transducer layer, an attention layer, and a feed forward propagation layer. Using the model to predict the remaining life of the engine may include: and acquiring real-time data of the engine, and sequentially performing cyclic processing on the acquired data through a transducer layer, a attention layer and a feedforward propagation layer to obtain final data, wherein the number of the cyclic processing cycles is preset. And predicting the final data to obtain the residual life of the engine.

However, in order to ensure the accuracy of the prediction result, a large amount of data needs to be processed, however, the correlation between some data and the prediction result is low, so that redundant calculation is generated in the processing process, and the part of data may derive a large amount of data in the processing process, and the data occupies a large amount of memory resources. Thus, the processing speed of the predictive model currently used is slow.

Disclosure of Invention

The embodiment of the application provides an industrial equipment state prediction method and device based on light-weight time sequence reduction, which can reduce the data volume in the process of processing a prediction model and can improve the running speed of the prediction model under the condition of less influence on the precision of a prediction result.

In a first aspect, an embodiment of the present application provides a method for predicting a state of an industrial device based on lightweight time series reduction, where the method for predicting a state of an industrial device based on lightweight time series reduction includes:

acquiring operation data of an engine;

inputting the operation data to a first neuron group in a group linear transformation embedded layer of a prediction model to perform linear transformation processing on the operation data to obtain first data; the linear transformation embedded layers comprise a plurality of linear transformation layers, wherein a target linear transformation layer in the linear transformation embedded layers comprises L groups of first neurons, input data of the target linear transformation layer is divided into L groups of sub-data, the L groups of sub-data are in one-to-one correspondence with the L groups of first neurons, and L is a natural number greater than or equal to 1;

inputting the first data into an attention layer of the prediction model, and processing the first data in the attention layer by adopting an attention mechanism to obtain second data;

Inputting the second data to a time sequence reduction layer of the prediction model for importance scoring, and removing the data with the importance score lower than a first preset value in the second data to obtain third data;

inputting the third data into a group feedforward network layer of the prediction model to perform linear transformation processing to obtain fourth data;

and inputting the fourth data into a learning layer of the prediction model, and learning the fourth data according to a multi-level learning mechanism strategy in the learning layer to obtain the residual service life of the engine.

In a possible implementation manner, the input data of the target group linear transformation layer is first time series data;

in the target group linear transformation layer, the first time series data is divided into L groups of first sub-time series data in advance according to the dimension of the time series, and the L groups of first sub-time series data are in one-to-one correspondence with the L groups of first neurons.

In a possible implementation manner, the inputting the operation data into the neuron group in the group linear transformation embedded layer of the prediction model performs linear transformation processing on the operation data to obtain first data, including:

Inputting the first time series data to a set of linear transformation embedding layers of a predictive model;

aiming at the target group linear transformation layer, the L groups of first neurons respectively process corresponding first sub-time sequence data in the L groups of first sub-time sequence data to obtain a first processing result;

inputting the first processing result into a linear transformation layer of a next layer group of the target group of linear transformation layers for processing;

and processing the last linear transformation layer in the linear transformation layers to obtain the first data.

In a possible implementation manner, the second data includes a plurality of time steps;

inputting the second data to a time sequence reduction layer of the prediction model for importance scoring, and removing the data with the importance score lower than a first preset value from the second data to obtain third data, wherein the third data comprises the following steps:

inputting the second data into a time-series reduction layer of the predictive model, by an importance scoring function in the time-series reduction layerScoring each time step to obtain importance scores of each time step;

wherein ,representing time step- >Importance score of->Outputting the value of the ith row and jth column of the matrix for the attention layer, representing the time step +.>For step of time->N represents the number of time steps;

and removing the time steps with importance scores lower than the first preset value in the plurality of time steps to obtain third data.

In a possible implementation manner, the group feedforward network layer includes a plurality of groups of feedforward network layers, the target group feedforward network layer of the group feedforward network layers includes Q groups of second neurons, the input data of the target group feedforward network layer is second time series data, and Q is a natural number greater than or equal to 1;

in the target group feedforward network layer, the second time series data is divided into Q groups of second sub time series data in advance according to the dimension of the time series, and the Q groups of second sub time series data are in one-to-one correspondence with the Q groups of second neurons.

In a possible implementation manner, the inputting the third data to the group feedforward network layer of the prediction model to perform linear transformation processing to obtain fourth data includes:

inputting the second time series data to a group feed forward network layer of the predictive model;

Aiming at the target group feedforward network layer, the Q group second neurons respectively process corresponding second sub time sequence data in the Q group second sub time sequence data to obtain a second processing result;

inputting the second processing result into a feedforward network layer of a next layer of the feedforward network layer of the target group for processing;

and obtaining the fourth data through processing of a last layer of the feedforward network layers.

In a possible implementation manner, the inputting the fourth data into a learning layer of the prediction model includes:

performing cyclic processing on the fourth data through the attention layer, the time sequence reduction layer and the group feedforward network layer to obtain target data, wherein the output of the attention layer is the input of the time sequence reduction layer, the output of the time sequence reduction layer is the input of the group feedforward network layer, the output of the group feedforward network layer is the input of the attention layer, and the cyclic processing cycle number is a second preset value;

and inputting the target data into a learning layer of the prediction model.

In a possible implementation manner, in the learning layer, the learning the fourth data according to a multi-level learning mechanism policy to obtain a remaining life of the engine includes:

Flattening the second data in the learning layer to obtain a first vector;

flattening the target data to obtain a second vector;

propagating the first vector and the second vector forward in parallel using a first multi-layer perceptron;

processing the first vector propagated by the first multi-layer perceptron through a ReLU activation function and Batchnorm normalization to obtain a third vector;

processing the second vector propagated by the first multi-layer perceptron through a ReLU activation function and Batchnorm normalization to obtain a fourth vector;

concatenating the third vector and the fourth vector into a fifth vector;

and forward transmitting the fifth vector through a second multi-layer perceptron to obtain the residual life of the engine.

In a second aspect, an embodiment of the present application provides an industrial equipment state prediction apparatus based on lightweight time series reduction, including:

the acquisition module is used for acquiring the operation data of the engine;

the processing module is used for inputting the operation data to a first neuron group in a group linear transformation embedded layer of the prediction model to perform linear transformation processing on the operation data so as to obtain first data; the linear transformation embedded layers comprise a plurality of linear transformation layers, wherein a target linear transformation layer in the linear transformation embedded layers comprises L groups of first neurons, input data of the target linear transformation layer is divided into L groups of sub-data, the L groups of sub-data are in one-to-one correspondence with the L groups of first neurons, and L is a natural number greater than or equal to 1;

The processing module is further configured to input the first data to an attention layer of the prediction model, and in the attention layer, process the first data by adopting an attention mechanism to obtain second data;

the processing module is further configured to input the second data to a time sequence reduction layer of the prediction model to perform importance scoring, and remove data in the second data with importance scoring lower than a first preset value to obtain third data;

the processing module is further configured to input the third data to a group feedforward network layer of the prediction model to perform linear transformation processing, so as to obtain fourth data;

the processing module is further configured to input the fourth data to a learning layer of the prediction model, and learn the fourth data in the learning layer according to a multi-level learning mechanism policy to obtain a remaining life of the engine.

in the target group linear transformation layer, the first time series data is divided into L groups of first sub time series data in advance in accordance with the dimension of the time series.

In a possible implementation manner, the processing module is specifically configured to input the first time-series data into a set of linear transformation embedding layers of a prediction model; aiming at the target group linear transformation layer, the L groups of first neurons respectively process corresponding first sub-time sequence data in the L groups of first sub-time sequence data to obtain a first processing result; inputting the first processing result into a linear transformation layer of a next layer group of the target group of linear transformation layers for processing; and processing the last linear transformation layer in the linear transformation layers to obtain the first data.

In a possible implementation manner, the second numberThe method comprises a plurality of time steps; the processing module is specifically configured to input the second data into a time-series reduction layer of the prediction model, where the importance scoring function is used in the time-series reduction layerScoring each time step to obtain importance scores of each time step; wherein (1)>Representing time step->Importance score of->Outputting the value of the ith row and jth column of the matrix for the attention layer, representing the time step +.>For step of time- >N represents the number of time steps; and removing the time steps with importance scores lower than the first preset value in the plurality of time steps to obtain third data.

In a possible implementation manner, the group feedforward network layer includes a plurality of groups of feedforward network layers, the target group feedforward network layer of the group feedforward network layers includes Q groups of second neurons, the input data of the target group feedforward network layer is second time series data, and Q is a natural number greater than or equal to 1; in the target group feedforward network layer, the second time series data is divided into Q groups of second sub time series data in advance according to the dimension of the time series, and the Q groups of second sub time series data are in one-to-one correspondence with the Q groups of second neurons.

In a possible implementation manner, the processing module is specifically configured to input the second time-series data to a group feedforward network layer of the prediction model; aiming at the target group feedforward network layer, the Q group second neurons respectively process corresponding second sub time sequence data in the Q group second sub time sequence data to obtain a second processing result; inputting the second processing result into a feedforward network layer of a next layer of the feedforward network layer of the target group for processing; and obtaining the fourth data through processing of a last layer of the feedforward network layers.

In a possible implementation manner, the processing module is specifically configured to perform a loop processing on the fourth data through the attention layer, the time sequence reduction layer, and the group feedforward network layer to obtain target data, where an output of the attention layer is an input of the time sequence reduction layer, an output of the time sequence reduction layer is an input of the group feedforward network layer, an output of the group feedforward network layer is an input of the attention layer, and a number of loops of the loop processing is a second preset value; and inputting the target data into a learning layer of the prediction model.

In a possible implementation manner, the processing module is specifically configured to perform flattening processing on the second data in the learning layer to obtain a first vector; flattening the target data to obtain a second vector; propagating the first vector and the second vector forward in parallel using a first multi-layer perceptron; processing the first vector propagated by the first multi-layer perceptron through a ReLU activation function and Batchnorm normalization to obtain a third vector; processing the second vector propagated by the first multi-layer perceptron through a ReLU activation function and Batchnorm normalization to obtain a fourth vector; concatenating the third vector and the fourth vector into a fifth vector; and forward transmitting the fifth vector through a second multi-layer perceptron to obtain the residual life of the engine.

In a third aspect, an embodiment of the present application further provides an electronic device, including: a processor, and a memory communicatively coupled to the processor;

the memory stores computer-executable instructions;

the processor executes computer-executable instructions stored in the memory to implement the method as described in any one of the possible implementations of the first aspect.

In a fourth aspect, an embodiment of the present application further provides a computer readable storage medium, where computer executable instructions are stored, and when executed by a processor, implement the method described in any one of the possible implementation manners of the first aspect.

In a fifth aspect, embodiments of the present application further provide a computer program product comprising a computer program which, when executed by a processor, implements a method as described in any one of the possible implementations of the first aspect.

It can be seen that the embodiments of the present application provide a method and apparatus for predicting an industrial device state based on lightweight time series reduction, where the method includes: acquiring operation data of an engine; inputting the operation data to a first neuron group in a group linear transformation embedded layer of a prediction model to perform linear transformation processing on the operation data to obtain first data; the group linear transformation embedded layer comprises a plurality of groups of linear transformation layers, wherein a target group linear transformation layer in the group linear transformation embedded layer comprises L groups of first neurons, input data of the target group linear transformation layer is divided into L groups of sub-data, and the L groups of sub-data are in one-to-one correspondence with the L groups of first neurons; inputting the first data into an attention layer of the prediction model, and processing the first data in the attention layer by adopting an attention mechanism to obtain second data; inputting the second data to a time sequence reduction layer of the prediction model to carry out importance scoring, and removing data with importance scoring lower than a first preset value in the second data to obtain third data; inputting the third data into a group feedforward network layer of the prediction model to perform linear transformation processing to obtain fourth data; and inputting the fourth data into a learning layer of the prediction model, and learning the fourth data according to a multi-level learning mechanism strategy in the learning layer to obtain the residual service life of the engine. According to the method provided by the embodiment of the application, the data of the input group linear transformation layer and the neurons in the input group linear transformation layer are grouped, and each group of neurons process the corresponding data group, so that each neuron only needs to process one group of data, but not all the data, the data volume processed by each neuron can be reduced, the complexity of data processing is reduced, and the processing speed is improved. In addition, the data with low importance scores is removed through the time sequence reduction layer, so that the data amount in the prediction model is further reduced. Therefore, the method provided by the embodiment of the application can improve the speed of processing the data by the prediction model.

Drawings

FIG. 1 is a schematic structural diagram of a prediction model according to an embodiment of the present application;

FIG. 2 is a schematic flow chart of an industrial equipment state prediction method based on light-weight time sequence reduction according to an embodiment of the present application;

FIG. 3 is a schematic structural diagram of a set of linear transformation embedded layers according to an embodiment of the present application;

fig. 4 is a schematic structural diagram of a time sequence reduction layer according to an embodiment of the present application;

FIG. 5 is a flowchart of a method for processing target data by a learning layer according to an embodiment of the present application;

FIG. 6 is a schematic diagram of another prediction model according to an embodiment of the present application;

FIG. 7 is a flowchart of another method for predicting the status of an industrial device based on lightweight time series reduction according to an embodiment of the present application;

FIG. 8 is a schematic structural diagram of an industrial equipment state prediction device based on light-weight time-series reduction according to an embodiment of the present application;

fig. 9 is a schematic structural diagram of an electronic device according to the present application.

Specific embodiments of the present disclosure have been shown by way of the above drawings and will be described in more detail below. These drawings and the written description are not intended to limit the scope of the disclosed concepts in any way, but rather to illustrate the disclosed concepts to those skilled in the art by reference to specific embodiments.

Detailed Description

Reference will now be made in detail to exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, the same numbers in different drawings refer to the same or similar elements, unless otherwise indicated. The implementations described in the following exemplary examples are not representative of all implementations consistent with the present disclosure. Rather, they are merely examples of apparatus and methods consistent with some aspects of the present disclosure as detailed in the accompanying claims.

In embodiments of the present application, "at least one" means one or more, and "a plurality" means two or more. "and/or", describes an association relationship of an association object, and indicates that there may be three relationships, for example, a and/or B, and may indicate: there are three cases, a alone, a and B together, and B alone, wherein a, B may be singular or plural. In the text description of the present application, the character "/" generally indicates that the front-rear associated object is an or relationship.

Engines are installed in common industrial equipment, and the equipment is not operated by the engines. Thus, the health of the engine has a critical effect on whether the device can function properly. The health of the engine includes, but is not limited to, the remaining life of the engine, the degree of wear of the engine, etc.

In some implementations, to make the predicted outcome more accurate, the remaining life of the plant engine may be predicted by a predictive model. The prediction model comprises a transducer layer, an attention layer and a feedforward propagation layer. the transducer layer and the feedforward propagation layer each include a plurality of neurons therein. When the prediction model is used for predicting the residual life of the engine, the collected operation data of the engine can be input into a transducer layer, the transducer layer can perform linear transformation processing on the operation data of the engine, the processed data is input into an attention layer to be processed through an attention mechanism, the processed data is input into a feedforward propagation layer to continue the linear transformation processing, the steps are circularly executed until the cycle times reach the preset times, and finally output data is obtained. The remaining life of the engine is obtained by performing prediction processing on the finally output data.

The input data comprises a plurality of sub-data, and when the transducer layer processes the input data, each neuron in the transducer layer processes each sub-data in the input data, and when the feedforward propagation layer processes the input data, each neuron in the feedforward propagation layer processes each sub-data in the input data.

However, in the above implementation, each neuron in the transducer layer and the feedforward propagation layer needs to process a large amount of data, so that there is a large amount of linear conversion processing in the model processing. Furthermore, the data input into the predictive model may contain data that has less effect on predicting the remaining life of the engine, such that a significant amount of redundant computation is generated during the predictive processing of the data. Therefore, in the implementation, the prediction model may occupy a large amount of memory resources in the running process, and the operation speed is low.

Based on the above, the embodiment of the application provides an industrial equipment state prediction method based on light-weight time sequence reduction, for any one of the linear transformation layers in the linear transformation embedding layer, the neurons and the input data are divided into different groups, and the neuron groups are in one-to-one correspondence with the data groups, so that the neurons in each neuron group only need to process the data in the corresponding data group, and do not need to process other data, thereby reducing the data processing amount of the neurons. Since the neurons are in the same manner as the processing, the accuracy of the prediction result is not reduced by grouping. In addition, the data with low importance scores in the input data are removed through the time sequence reduction layer, so that the prediction model does not need to process the data with low importance scores in the subsequent process. Therefore, the method provided by the embodiment of the application reduces the data volume of the data processed by the prediction model under the condition of less influence on the prediction precision, can reduce the occupation of memory resources in the data processing process of the prediction model, and improves the data processing speed of the prediction model.

It should be noted that, the engine in the embodiment of the present application may be various engines in the industrial field, for example, an aircraft engine, and the embodiment of the present application is not limited to the engine specifically.

Hereinafter, the method for predicting the state of an industrial device based on lightweight time series reduction provided by the present application will be described in detail by way of specific examples. It is to be understood that the following embodiments may be combined with each other and that some embodiments may not be repeated for the same or similar concepts or processes.

In order to clearly describe the technical solution of the embodiment of the present application, the following describes the model structure used in the embodiment of the present application. Fig. 1 is a schematic structural diagram of a prediction model according to an embodiment of the present application.

As shown in fig. 1, a lightweight transducer layer and a learning layer may be included in the prediction model according to an embodiment of the present application. The lightweight transformer layer can comprise a group linear transformation embedded layer, an attention layer, a time sequence reduction layer and a group feedforward network layer.

The group linear transformation embedding layer may be used to perform linear transformation processing on input data. The attention layer may be used to process the input data using an attention mechanism. The time sequence reduction layer may be used to filter the input data and remove the data that does not meet the condition in the input data. The group feed forward network layer may be used to perform linear transformation processing on the input data.

As shown in fig. 1, the operation data of the engine may be acquired, and the operation data of the engine may be input to a group linear transformation embedding layer in the lightweight transformer layer of the prediction model, and the group linear transformation embedding layer may perform linear transformation processing on the input data to output data a. The data a may be input to an attention layer of the lightweight transformer layer, which may process the data a using an attention mechanism to output the data B. The data B may be input to a time-series reduction layer of the lightweight transformer layer, the time-series reduction layer performs screening processing on the data B, removes data having a low importance score, and outputs the screened data C. The data C may be input to a group feed forward network layer of the lightweight transformer layer, which processes the data C to output the data D. The data D can be input into the attention layer, the attention layer processes the data D, the data output by the attention layer is input into the time sequence reduction layer for processing, the data output by the time sequence reduction layer is input into the group feedforward network layer, the processing processes of the attention layer, the time sequence reduction layer and the group feedforward network layer on the data are circularly executed, when the number of times of circulation reaches the preset number of times, the group feedforward network layer outputs the data E, the data E can be input into the learning layer of the prediction model, and the learning layer processes the data E to obtain the residual service life of the engine.

Therefore, when the prediction model shown in fig. 1 is used for predicting the residual life of the engine, the time sequence reduction layer in the prediction model can remove the data with lower importance scores, so that the data amount of the data processed by the prediction model can be reduced, the memory resources occupied in the data processing process can be reduced, and the data processing speed of the prediction model can be improved under the condition that the accuracy of the prediction result is less affected.

Based on the structure of the prediction model shown in fig. 1, fig. 2 is a schematic flow chart of an industrial equipment state prediction method based on light-weight time sequence reduction according to an embodiment of the present application. The lightweight time series reduction based industrial equipment state prediction method may be performed by software and/or hardware means, for example, the hardware means may be lightweight time series reduction based industrial equipment state prediction means, and the lightweight time series reduction based industrial equipment state prediction means may be an electronic equipment or a processing chip in the electronic equipment. For example, referring to fig. 2, the lightweight time series reduction-based industrial equipment state prediction method may include:

s201, acquiring operation data of an engine.

The operational data of the engine may be data relating to the equipment in which the engine is located, and when the engine is an engine on an aircraft, the operational data of the engine may include data of the aircraft's speed of flight, pressure, altitude, etc. The embodiment of the application does not specifically limit the operation data of the engine.

The operation data of the engine can be collected through a plurality of sensors or through other devices, and the mode of collecting the engine is not limited in the embodiment of the application.

For example, after the electronic device obtains the operation data of the engine, the normalization processing may be performed on the operation data of the engine, and the data after the normalization processing may be executed in step S202 described below. The embodiment of the application does not limit the specific method of normalization processing.

S202, inputting the operation data into a first neuron group in a group linear transformation embedded layer of a prediction model to perform linear transformation processing on the operation data, so as to obtain first data.

The linear transformation embedding layers comprise a plurality of linear transformation layers, the target linear transformation layer in the linear transformation embedding layers comprises L groups of first neurons, the input data of the target linear transformation layer is divided into L groups of sub-data, the L groups of sub-data are in one-to-one correspondence with the L groups of first neurons, and L is a natural number greater than or equal to 1. The embodiment of the application does not limit the number of layers of the group linear transformation layer and the number of groups of the first neurons in the group linear transformation layer.

For example, the set of linear transformation embedded layers includes 3 sets of linear transformation layers, a first set of linear transformation layers in the set of linear transformation embedded layers includes 1 set of first neurons, a second set of linear transformation layers includes 2 sets of first neurons, and a third set of linear transformation layers includes 2 sets of first neurons.

Thus, in the first layer group linear transformation layer, the input data is 1 group, i.e., the input data is not divided. In the second layer group linear transformation layer, the input data is 2 groups, i.e., the input data is divided into 2 groups. In the third layer group linear transformation layer, the input data is 2 groups, i.e., the input data is divided into 2 groups.

When the first neuron group in the first group linear transformation embedded layer of the prediction model performs linear transformation processing on the operation data, the first neuron group in the first group linear transformation embedded layer of the group linear transformation embedded layer performs linear transformation processing on the operation data, the processed data is input into the second group linear transformation layer, the first neuron group in the second group linear transformation layer performs linear transformation processing on the input data, and the processing is sequentially performed until the first neuron group in the last group linear transformation layer of the group linear transformation embedded layer performs linear transformation processing on the input data to obtain the first data.

When the group linear transformation layer among the group linear transformation embedded layers is as in the above exemplary case, in the first group linear transformation layer, 1 group of first neurons perform linear transformation processing on all the operation data. In the second-layer group linear transformation layer, 2 groups of first neurons perform linear transformation processing on 2 groups of input data, for example, the first group of neurons perform linear transformation processing on the first group of input data, the second group of neurons perform linear transformation processing on the second group of input data, or the first group of neurons perform linear transformation processing on the second group of input data, and the second group of neurons perform linear transformation processing on the first group of input data. In the third layer group linear transformation layer, 2 groups of first neurons perform linear transformation processing on 2 groups of input data, for example, the first group of neurons perform linear transformation processing on the first group of input data, the second group of neurons perform linear transformation processing on the second group of input data, or the first group of neurons perform linear transformation processing on the second group of input data, and the second group of neurons perform linear transformation processing on the first group of input data.

S203, inputting the first data into an attention layer of the prediction model, and processing the first data by adopting an attention mechanism in the attention layer to obtain second data.

In the attention layer, the first data may be processed by a self-attention mechanism to obtain the second data, or the first data may be processed by a multi-head attention mechanism to obtain the second data.

The self-attention mechanism and the multi-head attention mechanism are determined according to the complexity of the use scene of the engine, the multi-head attention mechanism is adopted when the complexity of the use scene of the engine is high, and the self-attention mechanism is adopted when the complexity of the use scene of the engine is low. The embodiment of the application does not limit the attention mechanism in detail.

S204, inputting the second data into a time sequence reduction layer of the prediction model to score importance, and removing data with importance scores lower than a first preset value from the second data to obtain third data.

The importance score may represent the importance of each sub-data in the second data to the predicted outcome, the higher the importance score, the higher the importance of the sub-data to the predicted outcome, the lower the importance score, the lower the importance of the sub-data to the predicted outcome.

The first preset value is a value representing the importance score, and may be different or the same for different engines or different usage scenarios of the same engine.

S205, inputting the third data into a group feedforward network layer of the prediction model to perform linear transformation processing, and obtaining fourth data.

In the embodiment of the application, the structure of the group feedforward network layer is similar to that of the group linear transformation embedded layer, so that the group feedforward network layer can perform linear transformation processing on the third data to obtain fourth data.

The structure of the group feedforward network layer may be referred to the above description of the structure of the group linear transformation embedded layer, and the process of processing the third data by the group feedforward network layer may be similar to the process of performing the linear transformation processing on the operation data by the group linear transformation embedded layer in the step S202, so that the content of this part may be referred to the above steps and will not be repeated here.

S206, inputting the fourth data into a learning layer of the prediction model, and learning the fourth data according to a multi-level learning mechanism strategy in the learning layer to obtain the residual service life of the engine.

The learning layer may also be referred to as a multi-level learning layer, and the multi-level learning mechanism may be implemented by a multi-level perceptron.

Therefore, according to the industrial equipment state prediction method based on light-weight time sequence reduction, the first neurons in any one group of linear transformation layers in the group of linear transformation embedded layers and the data input to the group of linear transformation layers are grouped, so that the first neurons in each group of first neurons only need to process the data in the data group corresponding to the first neurons, and each data input is not required to process, the number of times of linear transformation of a prediction model is reduced, the complexity of data processing of the prediction model is reduced, and the occupation of data to memory resources can be reduced. And a time sequence reduction layer is arranged in the prediction model, and the data is screened through the time sequence reduction layer to remove unimportant data, so that the data quantity processed by the prediction model can be reduced. Therefore, the industrial equipment state prediction method based on the light-weight time series reduction can improve the speed of predicting the residual life of the engine by the prediction model.

In order to facilitate understanding of the industrial equipment state prediction method based on light-weight time sequence reduction provided by the embodiment of the application, a method for processing data by a group linear transformation embedding layer, an attention layer, a time sequence reduction layer, a group feedforward network layer and a learning layer in a prediction model is described below.

In the embodiment of the present application, the obtained operation data of the engine may be time-series data, and thus, the input data of the target group linear transformation layer is first time-series data.

In step S202, the L sets of sub-data may be L sets of first sub-time series data. Thus, in the target group linear transformation layer, the first time-series data may be divided into L groups of first sub-time-series data in advance in the dimension of the time series.

Thus, by dividing the time series data according to the dimension of the time series, different neuron groups can process the time series data of different dimensions.

In the embodiment of the present application, taking an example that the group linear transformation embedded layer includes 3 groups of linear transformation layers, the structure of the group linear transformation embedded layer and the method of performing data processing on the group linear transformation embedded layer are described.

Fig. 3 is a schematic structural diagram of a set of linear transformation embedded layers according to an embodiment of the present application.

As shown in fig. 3, the group linear transformation embedding layer includes 3 groups of linear transformation layers, the first group linear transformation layer in the group linear transformation embedding layer is a group=1, that is, the first group linear transformation layer includes 1 group of first neurons, the second group linear transformation layer is a group=2, that is, the second group linear transformation layer includes 2 groups of first neurons, and the third group linear transformation layer is a group=2, that is, the third group linear transformation layer includes 2 groups of first neurons.

Accordingly, in fig. 3, the data input to the first layer group linear transformation layer is 1 group, i.e., the input data is not divided. The data input to the second layer group linear transformation layer is 2 groups, i.e., the input data is divided into 2 groups. The data input to the third layer group linear transformation layer is 2 groups, i.e., the input data is divided into 2 groups.

Based on the structure of the target group linear transformation layer shown in fig. 3, the method for performing linear transformation processing on the operation data by using the neuron group in the group linear transformation embedded layer of the prediction model to obtain first data may include the following steps:

inputting the first time series data to a set of linear transformation embedding layers of the predictive model; aiming at a target group linear transformation layer, the L groups of first neurons respectively process corresponding first sub-time sequence data in the L groups of first sub-time sequence data to obtain a first processing result; inputting the first processing result into the linear transformation layer of the next layer group of the target group of linear transformation layers for processing; and obtaining first data through processing of a last linear transformation layer in the linear transformation layers.

The target set of linear transformation layers may be any one of the set of linear transformation embedding layers.

When the next group of linear transformation layers of the target group of linear transformation layers processes the first processing result, the corresponding sub-time series data can be processed through each group of first neurons in the next group of linear transformation layers. When the last group of linear transformation layers of the target group of linear transformation layers processes input data, the first neurons of each group in the last group of linear transformation layers can process the corresponding sub-time sequence data.

It should be noted that, when the structure of the set of linear transformation embedded layers is the structure shown in fig. 3, the process of processing the input operation data by the set of linear transformation embedded layers may be referred to the related description in step S202 in the above embodiment, which is not repeated here.

Exemplary, assume that the time series of inputs isT is the length of the time series and d is the dimension of the time series. Group linear transformation can divide X into g groups +.>. wherein ,/>Is the divided vector. The divided vector can pass g learning matrices +.>Conversion to group output->. After processing of a plurality of group linear transformation layers, the outputs of the last group linear transformation layer are combined to +. >And obtaining the final output of the group linear transformation.

In this way, in the process of processing the input operation data by the group linear transformation embedded layer, as any group of first neurons in any group of linear transformation layer can process the corresponding sub-time sequence data without processing other sub-time sequence data, the number of times of linear transformation of the group linear transformation embedded layer is reduced, the data quantity generated in the processing process is reduced, and the memory occupation condition of the device is reduced.

Further, the first data is input into an attention layer of the prediction model, and in the attention layer, an attention mechanism is adopted to process the first data, so that second data is obtained.

By way of example, the following two possible implementations may be included when the first data is processed using the attention mechanism.

In one possible implementation, the first data may be processed by a self-attention mechanism, which may include the following equations (1) and (2):

（1）

（2）

wherein ,is the output of the linear transformation layer of the ith layer group,/->The matrix is a learning matrix, and Q, K and V are matrices after the learning matrix transformation. softmax is the activation function, K ^T Is the transposed K matrix. d, d _k Is the scaling factor.

It will be appreciated that when the structure of the set of linear transformation embedded layers is as shown in FIG. 3, including 3-layer set of linear transformation layers, the above equation (1)Is->。

In another possible implementation, the first data may be processed through a multi-head attention mechanism, which may include the following equation (3) and equation (4):

（3）

（4）

wherein ,，/>is a parameter matrix (also called learning matrix), +.>Is self-attention. Let us take h=1, h being the number of heads in the multi-head attention. />，/>Is the dimension of the learning matrix Q, K, V.

It should be noted that, the self-attention mechanism and the multi-head attention mechanism are realized by the attention encoder.

In the embodiment of the present application, the second data output by the attention layer output in the prediction model may be input into the time sequence reduction layer for screening, and the time sequence reduction layer may include:

inputting the second data into a time series reduction layer of the prediction model through an importance scoring function in the time series reduction layerScoring each time step to obtain importance scores of each time step The method comprises the steps of carrying out a first treatment on the surface of the Wherein the second data includes a plurality of time steps.

wherein ,representing time step->Importance score of->Outputting the value of the ith row and jth column of the matrix for the attention layer, representing the time step +.>For step of time->N represents the number of time steps.

Therefore, the time sequence data with low importance scores can be removed through the time sequence reduction layer, the data volume of the data processed by the prediction model is reduced, the occupation condition of memory resources in the data processing process of the prediction model is further reduced, and the data processing speed of the prediction model is improved.

Based on the method for processing data by the time-series reduction layer, fig. 4 is a schematic structural diagram of the time-series reduction layer according to an embodiment of the present application.

As shown in fig. 4, the plurality of time steps input to the time series reduction layer may process each time step by an importance scoring function to obtain an importance score for each time step. For example, time steps having importance scores of 0.3, 0.2, 0.1, … …, 0.4, etc. can be obtained, and time steps having importance scores of 0.1 can be eliminated, resulting in eliminated time steps. And inputting the time series data obtained after the elimination into a feed-forward network layer of the group for processing.

It should be noted that the embodiment of the present application is only illustrated in fig. 4, but the time-series reduction layer is not limited in any way.

Further, the third data output by the time series reduction layer may be input to the group feed forward network layer for processing. The group feedforward network layer may include a multi-layer group feedforward network layer, the target group feedforward network layer in the group feedforward network layer includes Q groups of second neurons, and the input data of the target group feedforward network layer is second time-series data. Wherein Q is a natural number greater than or equal to 1. In the target group feedforward network layer, the second time-series data may be divided into Q groups of second sub-time-series data in advance according to the dimension of the time-series, the Q groups of second sub-time-series data corresponding to the Q groups of second neurons one by one.

The target group feedforward network layer may be any layer of the group feedforward network layers, and the target group feedforward network layer is not limited in the embodiment of the present application.

In the embodiment of the present application, the number of layers of the group feedforward network layer included in the group feedforward network layer and the number of groups of the second neurons included in the group feedforward network layer are not particularly limited.

Based on the above description, the structure of the group feedforward network layer is similar to that of the group linear transformation embedded layer described in the above embodiment, and the description of the group linear transformation embedded layer in the above embodiment may be referred to, and will not be repeated here.

In the embodiment of the present application, inputting the third data to the group feedforward network layer of the prediction model to perform linear transformation processing to obtain the fourth data may include: inputting the second time series data to a group feed forward network layer of the predictive model; aiming at a target group feedforward network layer, respectively processing corresponding second sub-time sequence data in the Q groups of second sub-time sequence data by the Q groups of second neurons to obtain second processing results; inputting the second processing result into a feedforward network layer of the next layer of the feedforward network layer of the target group for processing; and obtaining fourth data through processing of a last layer of the group of feedforward network layers.

For example, when the second processing result is processed by the feedforward network layer of the next layer of the target group feedforward network layer, the corresponding sub-time series data may be processed by each group of second neurons in the feedforward network layer of the next layer. When the last group of feedforward network layers of the target group of feedforward network layers processes the input data, each group of second neurons in the last group of feedforward network layers can process the corresponding sub-time sequence data.

In this way, in the process of processing the input third data by the group feedforward network layer, as any group of second neurons in the group feedforward network layer can process the corresponding sub-time sequence data without processing other sub-time sequence data in any group of feedforward network layer, the number of times of linear transformation of the group feedforward network layer is reduced, the data quantity generated in the processing process is reduced, and the memory occupation condition of the equipment is reduced.

In the embodiment of the present application, inputting the fourth data output by the group feedforward network layer into the learning layer of the prediction model may include: circularly processing fourth data through the attention layer, the time sequence reduction layer and the group feedforward network layer to obtain target data; the target data is input to a learning layer of the predictive model.

The output of the attention layer is the input of the time sequence reduction layer, the output of the time sequence reduction layer is the input of the group feedforward network layer, the output of the group feedforward network layer is the input of the attention layer, and the cycle number of the cycle processing is a second preset value.

The number of cyclic processes may be a number obtained by subtracting 1 from the number of the transducer layers, which is not limited in the embodiment of the present application.

The cyclic processing procedure may be referred to the description of the attention layer, the time sequence reduction layer and the group feedforward network layer, and will not be described herein.

After the cyclic processing of the attention layer, the time series reduction layer, and the group feedforward network layer, the target data can be obtained and input into the learning layer for processing, as described in connection with the above embodiments. Fig. 5 is a flowchart of a method for processing target data by a learning layer according to an embodiment of the present application.

As shown in fig. 5, the method for processing target data by the learning layer may include the steps of:

s501, flattening the second data in the learning layer to obtain a first vector.

Exemplary, the second data is feature data of the un-pruned branches, and flattening the feature data of the un-pruned branches to obtain a first vector a ₁ 。

S502, flattening the target data to obtain a second vector.

The target data is characteristic data of complete pruning, and flattening is carried out on the characteristic data of complete pruning to obtain a second vector a _N 。

S503, the first vector and the second vector are propagated forward in parallel by using the first multi-layer perceptron.

The number of layers of the first multi-layer perceptron may be 2, 3, or other values, which are not limited in the embodiment of the present application.

S504, processing the first vector propagated by the first multi-layer perceptron through a ReLU activation function and Batchnorm normalization to obtain a third vector.

Exemplary, for the first vector a ₁ And (5) processing to obtain a third vector.

S505, processing the second vector propagated by the first multi-layer perceptron through a ReLU activation function and Batchnorm normalization to obtain a fourth vector.

Exemplary, for the second vector a _N Processing to obtain fourth vector。

S506, the third vector and the fourth vector are connected in series to form a fifth vector.

Illustratively, concatenating the above-described third and fourth vectors results in a fifth vector。

S507, forward transmitting a fifth vector through the second multi-layer perceptron to obtain the residual life of the engine.

In the embodiment of the present application, the number of layers of the second multi-layer perceptron may be 2, or may be 3, or other values. The number of layers of the second multi-layer perceptron may be the same as or different from the number of layers of the first multi-layer perceptron. The embodiment of the application does not limit the layer number of the second multilayer perceptron.

Exemplary, propagating the fifth vector forward through the second multi-layer perceptronThe remaining life of the engine is obtained.

Since the time series reduction layer may remove a part of the data, the part of the data removed by the time series reduction layer may not be redundant data. Therefore, the learning layer can obtain the remaining life of the engine together with the data not removed by the time series reduction layer and the data removed by the time series reduction layer, and can reduce the degradation of the accuracy of the prediction result due to the screening of the data by the time series reduction layer.

In connection with the above embodiments, fig. 6 is a schematic structural diagram of another prediction model according to an embodiment of the present application.

Illustratively, the results of the predictive model shown in FIG. 6 are similar to those of FIG. 1 and are not described in detail herein.

Referring to fig. 6, a method for predicting an industrial device state based on light-weight time-series reduction may be shown in fig. 7, and fig. 7 is a flow chart of another method for predicting an industrial device state based on light-weight time-series reduction according to an embodiment of the present application.

As shown in fig. 7, the industrial equipment state prediction method based on lightweight time series reduction may include the steps of:

s701, collecting data by multiple sensors, and carrying out normalization processing on the collected data.

S702, the normalized data is subjected to linear transformation of the group embedded layer in the lightweight transformer layer.

S703, the data output from the group linear transformation embedding layer passes through the attention layer in the lightweight transformer layer.

And S704, the data output by the attention layer passes through the time sequence reduction layer to obtain importance scores of different time steps, and the time steps after ranking are eliminated.

S705, the data output by the time sequence reduction layer passes through the group feedforward network layer in the lightweight transformer layer.

S706, repeating the steps S703-S705 for N-1 times, wherein N is the number of layers of the transducer network, and inputting the circularly processed data into a multi-level learning layer.

And S707, in the multi-level learning layer, obtaining the residual service life prediction according to the time sequence data which is not reduced in the first transducer layer and the data which is circularly processed by the last transducer layer.

In summary, according to the industrial equipment state prediction method based on light-weight time sequence reduction provided by the embodiment of the application, a light-weight transform structure is established in a prediction model by adjusting dimensions and using grouping transformation, so that the parameter number is reduced. The computational cost of industrial time series is reduced by adaptively eliminating redundant time steps by a time series reduction layer. In addition, a multi-level learning mechanism is also proposed to stabilize the time series reduction performance under various operating conditions.

Fig. 8 is a schematic structural diagram of an industrial equipment status prediction apparatus 80 based on lightweight time series reduction according to an embodiment of the present application, and as shown in fig. 8, for example, the industrial equipment status prediction apparatus 80 based on lightweight time series reduction may include:

An acquisition module 801 for acquiring engine operation data.

A processing module 802, configured to input operation data to a first neuron group in a group linear transformation embedding layer of the prediction model, perform linear transformation processing on the operation data, and obtain first data; the group linear transformation embedding layer comprises a plurality of groups of linear transformation layers, the target group linear transformation layer in the group linear transformation embedding layer comprises L groups of first neurons, the input data of the target group linear transformation layer is divided into L groups of sub-data, the L groups of sub-data are in one-to-one correspondence with the L groups of first neurons, and L is a natural number greater than or equal to 1.

The processing module 802 is further configured to input the first data to an attention layer of the prediction model, where an attention mechanism is used to process the first data to obtain the second data.

The processing module 802 is further configured to input the second data to the time-series reduction layer of the prediction model to perform importance scoring, and remove data with importance scoring lower than the first preset value in the second data, so as to obtain third data.

The processing module 802 is further configured to input the third data to a group feedforward network layer of the prediction model to perform linear transformation processing, so as to obtain fourth data.

The processing module 802 is further configured to input fourth data to a learning layer of the prediction model, where the fourth data is learned according to a multi-level learning mechanism policy to obtain a remaining life of the engine.

In one possible implementation, the input data of the target group linear transformation layer is first time series data; in the target group linear transformation layer, the first time series data is divided into L groups of first sub time series data in advance in accordance with the dimension of the time series.

In a possible implementation manner, the processing module 802 is specifically configured to input the first time-series data into a set of linear transformation embedding layers of the prediction model; aiming at a target group linear transformation layer, the L groups of first neurons respectively process corresponding first sub-time sequence data in the L groups of first sub-time sequence data to obtain a first processing result; inputting the first processing result into the linear transformation layer of the next layer group of the target group of linear transformation layers for processing; and obtaining first data through processing of a last linear transformation layer in the linear transformation layers.

In a possible implementation, the second data includes a plurality of time steps; the processing module 802 is specifically configured to input the second data into a time-series reduction layer of the prediction model, where the second data is passed through a weight in the time-series reduction layer Significance scoring functionScoring each time step to obtain importance scores of each time step; wherein,representing time step->Importance score of->For the value of the ith row and jth column of the attention layer output matrix, time step +.>For step of time->N represents the number of time steps; and removing the time steps with importance scores lower than the first preset value in the plurality of time steps to obtain third data.

In a possible implementation manner, the group feedforward network layer includes a multi-layer group feedforward network layer, the target group feedforward network layer includes Q groups of second neurons, the input data of the target group feedforward network layer is second time series data, and Q is a natural number greater than or equal to 1; in the target group feedforward network layer, the second time series data is divided into Q groups of second sub time series data in advance according to the dimension of the time series, and the Q groups of second sub time series data are in one-to-one correspondence with the Q groups of second neurons.

In a possible implementation manner, the processing module 802 is specifically configured to input the second time series data to a group feedforward network layer of the prediction model; aiming at a target group feedforward network layer, the Q group second neurons respectively process corresponding second sub time sequence data in the Q group second sub time sequence data to obtain a second processing result; inputting the second processing result into a feedforward network layer of the next layer of the feedforward network layer of the target group for processing; and obtaining fourth data through processing of a last layer of the group of feedforward network layers.

In a possible implementation manner, the processing module 802 is specifically configured to perform a loop processing on the fourth data through an attention layer, a time sequence reduction layer and a group feedforward network layer to obtain target data, where an output of the attention layer is an input of the time sequence reduction layer, an output of the time sequence reduction layer is an input of the group feedforward network layer, an output of the group feedforward network layer is an input of the attention layer, and a number of loops of the loop processing is a second preset value; the target data is input to a learning layer of the predictive model.

In a possible implementation manner, the processing module 802 is specifically configured to perform flattening processing on the second data in the learning layer to obtain a first vector; flattening the target data to obtain a second vector; propagating the first vector and the second vector forward in parallel using a first multi-layer perceptron; processing the first vector propagated by the first multi-layer perceptron through a ReLU activation function and Batchnorm normalization to obtain a third vector; processing the second vector propagated by the first multi-layer perceptron through a ReLU activation function and Batchnorm normalization to obtain a fourth vector; concatenating the third vector and the fourth vector into a fifth vector; and (5) forward transmitting the fifth vector through the second multi-layer perceptron to obtain the residual service life of the engine.

The industrial equipment state prediction device based on light-weight time sequence reduction provided by the embodiment of the application can execute the technical scheme of the industrial equipment state prediction method based on light-weight time sequence reduction in any embodiment, and the implementation principle and beneficial effects of the industrial equipment state prediction device based on light-weight time sequence reduction are similar to those of the industrial equipment state prediction method based on light-weight time sequence reduction, and can be seen from the implementation principle and beneficial effects of the industrial equipment state prediction method based on light-weight time sequence reduction, and are not repeated here.

Fig. 9 is a schematic structural diagram of an electronic device according to the present application. As shown in fig. 9, the electronic device 900 may include: at least one processor 901 and a memory 902.

A memory 902 for storing programs. In particular, the program may include program code including computer-operating instructions.

The memory 902 may include high-speed RAM memory or may further include non-volatile memory (non-volatile memory), such as at least one disk memory.

The processor 901 is configured to execute computer-executable instructions stored in the memory 902 to implement the lightweight time series reduction-based industrial equipment state prediction method described in the foregoing method embodiments. The processor 901 may be a central processing unit (Central Processing Unit, abbreviated as CPU), or an application specific integrated circuit (Application Specific Integrated Circuit, abbreviated as ASIC), or one or more integrated circuits configured to implement embodiments of the present application. Specifically, when the industrial equipment state prediction method based on the lightweight time series reduction described in the foregoing method embodiment is implemented, the electronic equipment may be, for example, an electronic equipment having a processing function, such as a terminal, a server, or the like.

Optionally, the electronic device 900 may also include a communication interface 903. In a specific implementation, if the communication interface 903, the memory 902, and the processor 901 are implemented independently, the communication interface 903, the memory 902, and the processor 901 may be connected to each other through buses and perform communication with each other. The bus may be an industry standard architecture (Industry Standard Architecture, abbreviated ISA) bus, an external device interconnect (Peripheral Component, abbreviated PCI) bus, or an extended industry standard architecture (Extended Industry Standard Architecture, abbreviated EISA) bus, among others. Buses may be divided into address buses, data buses, control buses, etc., but do not represent only one bus or one type of bus.

Alternatively, in a specific implementation, if the communication interface 903, the memory 902, and the processor 901 are integrated on a chip, the communication interface 903, the memory 902, and the processor 901 may complete communication through internal interfaces.

The present application also provides a computer-readable storage medium, which may include: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk or an optical disk, etc., in which program codes may be stored, and in particular, the computer-readable storage medium stores program instructions for the methods in the above embodiments.

The present application also provides a program product comprising execution instructions stored in a readable storage medium. The at least one processor of the electronic device may read the execution instructions from the readable storage medium, and execution of the execution instructions by the at least one processor causes the electronic device to implement the lightweight time series reduction-based industrial device state prediction method provided by the various embodiments described above.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present application, and not for limiting the same; although the application has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some or all of the technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit of the application.

Claims

1. An industrial equipment state prediction method based on lightweight time series reduction, which is characterized by comprising the following steps:

acquiring operation data of an engine;

inputting the fourth data into a learning layer of the prediction model, and learning the fourth data according to a multi-level learning mechanism strategy in the learning layer to obtain the residual service life of the engine;

the second data comprises a plurality of time steps;

inputting the second data into a time-series reduction layer of the predictive model, by an importance scoring function in the time-series reduction layer Scoring each time step to obtain importance scores of each time step;

wherein ,representing time step->Importance score of->Outputting the ith row and the jth column of the matrix for the attention layerValue, representing time step->For step of time->N represents the number of time steps;

2. The method of claim 1, wherein the input data of the target set of linear transformation layers is first time series data;

3. The method of claim 2, wherein the inputting the operational data into the set of neurons in the linear transformation embedding layer of the predictive model performs a linear transformation process on the operational data to obtain first data, comprising:

inputting the first time series data to a set of linear transformation embedding layers of the predictive model;

4. A method according to claim 3, wherein the group feedforward network layer includes a multi-layer group feedforward network layer, and the target group feedforward network layer includes Q groups of second neurons, and the input data of the target group feedforward network layer is second time series data, and Q is a natural number greater than or equal to 1;

5. The method of claim 4, wherein the inputting the third data into the group feedforward network layer of the predictive model for linear transformation to obtain fourth data comprises:

6. A method according to any one of claims 1-3, wherein said inputting said fourth data into a learning layer of said predictive model comprises:

and inputting the target data into a learning layer of the prediction model.

7. The method of claim 6, wherein learning the fourth data in the learning layer according to a multi-level learning mechanism strategy results in a remaining life of the engine, comprising:

flattening the second data in the learning layer to obtain a first vector;

flattening the target data to obtain a second vector;

concatenating the third vector and the fourth vector into a fifth vector;

8. An industrial equipment state prediction device based on lightweight time series reduction, comprising:

the acquisition module is used for acquiring the operation data of the engine;

The processing module is used for inputting the operation data to a first neuron group in a group linear transformation embedded layer of the prediction model to perform linear transformation processing on the operation data so as to obtain first data; the target group linear transformation layer in the group linear transformation embedded layer comprises L groups of first neurons, input data of the target group linear transformation layer is divided into L groups of sub-data, and the L groups of sub-data are in one-to-one correspondence with the L groups of first neurons;

the processing module is further configured to input the fourth data to a learning layer of the prediction model, where the learning layer learns the fourth data according to a multi-level learning mechanism policy to obtain a remaining life of the engine;

The second data comprises a plurality of time steps; the processing module is specifically configured to input the second data into a time-series reduction layer of the prediction model, where the importance scoring function is used in the time-series reduction layerScoring each time step to obtain importance scores of each time step; wherein (1)>Representing time step->Importance score of->Outputting the value of the ith row and jth column of the matrix for the attention layer, representing the time step +.>For step of time->N represents the number of time steps; and removing the time steps with importance scores lower than the first preset value in the plurality of time steps to obtain third data.

9. An electronic device, comprising: a processor, and a memory communicatively coupled to the processor;

the memory stores computer-executable instructions;

the processor executes the computer-executable instructions stored in the memory to implement a lightweight time series reduction based industrial equipment state prediction method as claimed in any one of claims 1-7.

10. A computer readable storage medium having stored therein computer executable instructions which, when executed by a processor, implement a lightweight time series reduction based industrial equipment state prediction method as claimed in any one of the preceding claims 1-7.