Operation and maintenance data prediction method and system based on position information adaptive adjustment

Abstract

The invention discloses an operation and maintenance data prediction method and system based on position information adaptive adjustment, and the method comprises the following steps: obtaining historical operation and maintenance data, and constructing a time series data set of the operation and maintenance data; performing position code embedding on all the time sequence data in the time sequence data set, and decomposing all the time sequence data after embedding position codes into season items and trend items; adjusting the trend item according to periodic information in the season item, and constructing an adaptive item for each piece of time sequence data by using the season item and the adjusted trend item; respectively fusing the season item, the adjusted trend item and the adaptive item of each time sequence data to obtain a fused time sequence feature representation; and inputting the fused time sequence feature representation into a Transform network to obtain a prediction result of the operation and maintenance data.

Description

Technical Field

[0001] The present invention relates to the field of operation and maintenance data prediction, and in particular to an operation and maintenance data prediction method and system based on adaptive adjustment of location information. Background Art

[0002] In the wave of digital transformation, enterprises have put forward higher requirements for the efficient operation and maintenance of production equipment, facilities and systems. In order to ensure the continuity and stability of the production process, it is crucial to detect and prevent potential failures in a timely manner. With the development of sensor technology and the Internet of Things (IoT), various equipment in the factory can collect a large amount of operation and maintenance data in real time, such as temperature, pressure, vibration, power consumption and other parameters, which form a complex operation and maintenance time series data set. Effective use of this data for analysis and prediction can help enterprises predict equipment failures in advance, optimize maintenance plans, improve production efficiency and reduce operating costs.

[0003] At present, deep learning technology, especially positional embedding, has been widely used in time series forecasting tasks. Positional embedding can capture the position information in the time series, enhance the model's understanding of the characteristics of different time points, and thus improve the prediction accuracy. This method embeds the position information into the operation and maintenance data, so that the model can better understand the change pattern of the data over time, making it possible to make accurate predictions.

[0004] Although location embedding technology has shown potential in operation and maintenance data prediction, existing methods still face challenges in dealing with complex time dependencies. In particular, when faced with operation and maintenance data with non-stationary or multi-scale time patterns, traditional position encoding methods often find it difficult to effectively capture the deep-level characteristics of the data, making existing time series decomposition methods insufficient in terms of precision and dynamic adjustment of the relationship between seasonal terms, trend terms, and location information. These problems limit the prediction performance of the model under complex operation and maintenance conditions, especially in scenarios that require high-precision and real-time predictions. Therefore, the development of more sophisticated and flexible time series analysis and prediction methods is an urgent problem to be solved in the current field of operation and maintenance data prediction. Summary of the invention

[0005] The present invention proposes an operation and maintenance data prediction method and system based on adaptive adjustment of location information, which solves the problem that the existing operation and maintenance data prediction method cannot dynamically adjust the relationship between the seasonal items, trend items and location information of the operation and maintenance time series data.

[0006] In order to solve the above technical problems, the present invention provides an operation and maintenance data prediction method based on adaptive adjustment of location information, comprising the following steps:

[0007] Step S1: Obtain historical operation and maintenance data and construct a time series data set of operation and maintenance data;

[0008] Step S2: position encoding and embedding all time series data in the time series data set, and decomposing all embedded position encoded time series data into seasonal items and trend items;

[0009] Step S3: adjusting the trend item according to the periodic information in the season item, and constructing an adaptive item for each time series data using the season item and the adjusted trend item;

[0010] Step S4: fusing the seasonal item, the adjusted trend item and the adaptive item of each time series data respectively to obtain a fused time series feature representation;

[0011] Step S5: Input the fused time series feature representation into the Transformer network to obtain the prediction result of the operation and maintenance data.

[0012] Preferably, in step S2, the expression for position encoding and embedding all the time series data in the time series data set is:

[0013] X PE =X+PE(X);

[0014]

[0015] In the above formula, X PE is the time series data after embedding position encoding; X is the time series data; PE(X) means position embedding of X; pos is the position index; i is the dimension index; d is the embedding dimension.

[0016] Preferably, in step S2, the expression for decomposing the time series data after embedding position encoding into seasonal terms and trend terms is:

[0017] D t =AvgPool(Padding(X PE ));

[0018] D s =X PE -D t ;

[0019] In the above formula, D t , D s They are the trend item and seasonal item of the time series data set respectively; AvgPool means moving average; Padding means filling the sequence.

[0020] Preferably, in step S3, the expression for adjusting the trend item according to the periodic information in the season item is:

[0021] D p =Dt -PE(D t );

[0022] Where D p is the adjusted trend item.

[0023] Preferably, the expression for constructing an adaptive term for each time series data using the seasonal term and the adjusted trend term in step S3 is:

[0024] H=D s +D p ⊙softmax(D s );

[0025] Where H is the adaptive term; ⊙ represents the corresponding multiplication of elements; and softmax is the activation function.

[0026] Preferably, the expression for fusing the seasonal term, the adjusted trend term and the adaptive term in step S4 is:

[0027] E=(D s W s +b s )+(D p W p +b p )+(HW h +b h );

[0028] Where E is the fused time series feature representation; D s , D p , H are the seasonal term, adjusted trend term and adaptive term of the time series data set respectively; W s , W p , W h D s , D p , the weight matrix of H; b s , b p , b h D s , D p , the bias term of H.

[0029] Preferably, in step S5, the fused time series feature representation is used as the input sequence of the Transformer network, the encoder of the Transformer network is used to perform deep feature extraction on the input sequence, and the feature representation output by the encoder is input into the fully connected layer of the Transformer network to be mapped into the prediction result of the operation and maintenance data.

[0030] Preferably, the encoder calculates the correlation between the query vector, key vector and value vector of the input sequence through a multi-head attention mechanism, models the global dependency relationship of each time position in the input sequence, performs residual connection and layer normalization processing on the output of the multi-head attention mechanism, inputs the feedforward neural network for nonlinear transformation, and again performs residual connection and layer normalization processing to obtain the final encoder output;

[0031] The expression of the multi-head attention mechanism is:

[0032] MultiHead(Q,K,V)=Concat(head1,...,head h )W O ;

[0033]

[0034] The expression of the residual connection and layer normalization processing is:

[0035] E'=LayerNorm(E+MultiHead(E,E,E));

[0036] The expression for the nonlinear transformation of the input feedforward neural network is:

[0037] FFN(E')=max(0,(E'W1+b1)W2+b2);

[0038] The output of the encoder is expressed as:

[0039] Z = LayerNorm(E'+FFN(E'));

[0040] In the above formula, MultiHead is a multi-head attention mechanism; Q, K, and V are the query vector, key vector, and value vector of the multi-head attention mechanism respectively; Concat is the connection function; head i is the i-th head of the multi-head attention mechanism; W O is a learnable parameter; Attention is the attention weight; Head i The projection matrix of the query vector, key vector and value vector; softmax is the activation function; d is the embedding dimension; E' is the time series feature representation after residual connection and layer normalization; E is the fused time series feature representation; Z is the output of the encoder; LayerNorm is the layer normalization function; FFN is the output of the feedforward neural network; W1 and W2 are the weight matrices of E'; b1 and b2 are the bias terms of E'.

[0041] Preferably, the expression of the prediction result of the operation and maintenance data in step S5 is:

[0042] Y=ZW y +b y ;

[0043] In the formula, Y is the prediction result of operation and maintenance data; is the weight matrix of Y, and f is the length of the prediction result; is the bias term of Y.

[0044] The present invention also provides an operation and maintenance data prediction system based on adaptive adjustment of position information, which is implemented based on the above-mentioned operation and maintenance data prediction method based on adaptive adjustment of position information, and includes: a data acquisition module, a position coding embedding module, a sequence decomposition module, a sequence fusion module, a feature coding module and a data prediction module;

[0045] The data collection module collects historical operation and maintenance data and constructs an operation and maintenance time series data set;

[0046] The position code embedding module is used to embed the position code of all the operation and maintenance time series data in the operation and maintenance time series data set;

[0047] The sequence decomposition module decomposes the operation and maintenance time series data after embedding position coding to obtain the seasonal items and trend items of the operation and maintenance time series data;

[0048] The sequence fusion module: constructs an adaptive term using the seasonal term and the adjusted trend term, fuses the seasonal term, the adjusted trend term and the adaptive term, and obtains a fused time series feature representation;

[0049] The feature encoding module: uses the encoder of the Transformer network to perform deep feature extraction on the fused time series feature representation;

[0050] The data prediction module: inputs the feature representation output by the feature encoding module into the fully connected layer of the Transformer network for prediction, and uses the mean square error loss function to optimize the parameters of the Transformer network.

[0051] The benefits of the present invention include at least:

[0052] 1. Position codes are embedded in the operation and maintenance time series data. The embedding of position codes can enhance the model's ability to recognize different position features in the time series, and help capture seasonal changes and trend changes in the operation and maintenance time series data;

[0053] 2. Decomposing the time series into seasonal terms and trend terms can enable the model to more accurately identify and understand the cyclical changes and long-term trends in the data. Adjusting the trend terms according to the cyclical information in the seasonal terms can make the trend terms more consistent with the actual cyclical changes, which helps to improve the accuracy of model predictions.

[0054] 3. By constructing adaptive items, the model can dynamically adjust the relationship between seasonal items, trend items and location information, so that the model can adapt to different operation and maintenance conditions and changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A schematic diagram of a method flow chart of an embodiment of the present invention;

[0056] Figure 2 The figure is a schematic diagram of the process of obtaining a fused operation and maintenance time series data set in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of the present invention.

[0058] Operation and maintenance time series data is a collection of data arranged in chronological order, which is common in many fields such as power analysis and weather forecasting. The core of the time series prediction task is to establish a time evolution model between historical data points to predict future numerical changes.

[0059] Suppose there is a set of operation and maintenance time series data sets containing multiple variables Where m represents the total number of time steps in the time series, and n represents the number of multivariate variables. Define a historical lookback window of length l Based on the given historical lookback window X l Predict the output Y for the next f time steps f = This prediction process can be achieved through Y f =δ(X l ), where δ is the constructed prediction model. The prediction task involves using X l To predict Y f , where Y f =δ(X l ), δ is the prediction model.

[0060] Compared with other types of data, time series data often lacks clear semantic information, which makes the position features in the data play an extremely critical role in the time series prediction task. Therefore, before training a multivariate time series prediction model, it is necessary to effectively extract the position features in the time series data to enhance the model's expressive power and prediction accuracy. This feature extraction process can be expressed as a mapping from the original time series data to the d-dimensional representation space:

[0061] To this end, the present invention proposes an operation and maintenance data prediction method based on adaptive adjustment of location information. This method can not only effectively capture the time location information in the data, but also reduce noise interference through adaptive adjustment mechanism and improve the generalization ability of the model. The overall flow chart of this method is as follows Figure 1 As shown in the figure, by integrating position information and adaptive adjustment strategies, efficient and accurate representation learning of time series data is achieved.

[0062] Specifically, the steps include:

[0063] Step S1: Acquire historical operation and maintenance data and construct an operation and maintenance time series data set. Divide the operation and maintenance time series data set and perform the initial operation and maintenance time series data X * Perform standardized preprocessing to obtain operation and maintenance time series data The expression of standardized preprocessing is:

[0064]

[0065] Where, X is the preprocessed operation and maintenance time series data; x * is the original operation and maintenance time series data; μ is the mean of the operation and maintenance time series data; σ is the standard deviation of the operation and maintenance time series data.

[0066] This normalization process can eliminate the dimensional differences between different features, allowing the model to better learn the intrinsic patterns of the data.

[0067] Step S2: Perform position encoding and embedding on all operation and maintenance time series data in the operation and maintenance time series data set, and decompose all the operation and maintenance time series data after embedding position encoding into seasonal items and trend items.

[0068] Specifically, for the input operation and maintenance time series data X, we first embed its position, and then add it to the original data to perform time series decomposition of the season-trend term. The expression for position encoding embedding of the operation and maintenance time series data set is:

[0069] X PE =X+PE(X);

[0070]

[0071] In the above formula, XPE is the operation and maintenance time series data after embedding position encoding; X is the operation and maintenance time series data; PE(X) means position embedding of X; pos is the position index; i is the dimension index; d is the embedding dimension.

[0072] The expression for decomposing the operation and maintenance time series data set after embedding position encoding is:

[0073] D t =AvgPool(Padding(X PE ));

[0074] D s =X PE -D t ;

[0075] In the above formula, D t , D s They are the trend item and seasonal item of the operation and maintenance time series data set respectively; AvgPool means moving average; Padding means padding the sequence to keep the length of the sequence unchanged.

[0076] This decomposition method can effectively separate the periodic patterns and long-term trends in time series data, laying the foundation for subsequent adaptive adjustments.

[0077] Step S3: Adjust the trend item according to the periodic information in the season item, and construct an adaptive item for each operation and maintenance time series data using the season item and the adjusted trend item.

[0078] Specifically, the expression for adjusting the trend term is:

[0079] D p =D t -PE(D t );

[0080] Where D p is the adjusted trend term; PE(X) represents position embedding.

[0081] The expression for constructing the adaptive term using the seasonal term and the adjusted trend term is:

[0082] H = Ds + Dp⊙softmax(Ds);

[0083] Where H is the adaptive term; ⊙ represents the corresponding multiplication of elements; and softmax is the activation function.

[0084] This step removes the noise position information in the trend term and uses the periodic information in the seasonal term to guide the adjustment of the trend term. At the same time, the introduction of the softmax activation function enables the model to adaptively focus on the important features in the seasonal term, thereby more effectively guiding the adjustment of the trend term.

[0085] Step S4: The seasonal item, the adjusted trend item and the adaptive item of each operation and maintenance time series data are fused respectively to obtain a fused time series feature representation.

[0086] Specifically, the features are fused to obtain an enhanced embedding representation

[0087] E=(D s W s +b s )+(D p W p +b p )+(HW h +b h );

[0088] Where, E is the fused operation and maintenance time series data set; D s , D p , H are the seasonal term, adjusted trend term and adaptive term of the operation and maintenance time series data set respectively; W s , W p , W h D s , D p , the weight matrix of H; b s , b p , b h D s , D p , the bias term of H.

[0089] Step S5: Input the fused time series feature representation into the Transformer network to obtain the prediction result of the operation and maintenance data.

[0090] Specifically, the fused time series feature representation is used as the input sequence of the Transformer network, and the encoder of the Transformer network is used to extract deep features of the input sequence. The encoder calculates the correlation between the query vector, key vector and value vector of the input sequence through the multi-head attention mechanism, models the global dependency relationship of each time position in the input sequence, performs residual connection and layer normalization on the output of the multi-head attention mechanism, inputs the feedforward neural network for nonlinear transformation, and again undergoes residual connection and layer normalization to obtain the final encoder output.

[0091] The expression of the multi-head attention mechanism is:

[0092] MultiHead(Q,K,V)=Concat(head1,...,head h )W O ;

[0093]

[0094] The expressions for residual connection and layer normalization are:

[0095] E'=LayerNorm(E+MultiHead(E,E,E)).

[0096] The expression for nonlinear transformation of the input feedforward neural network is:

[0097] FFN(E')=max(0,(E'W1+b1)W2+b2).

[0098] The final expression of the encoder output is:

[0099] Z = LayerNorm(E'+FFN(E')).

[0100] In the above formula, MultiHead is a multi-head attention mechanism; Q, K, and V are the query vector, key vector, and value vector of the multi-head attention mechanism respectively; Concat is the connection function; head i is the i-th head of the multi-head attention mechanism; W O is a learnable parameter; Attention is the attention weight; Head i The projection matrix of the query vector, key vector and value vector; softmax is the activation function; d is the dimension of the key vector, that is, the embedding dimension; E' is the time series feature representation after residual connection and layer normalization; E is the fused time series feature representation; Z is the output of the encoder; LayerNorm is the layer normalization function; FFN is the output of the feedforward neural network; W1 and W2 are the weight matrices of E'; b1 and b2 are the bias terms of E'.

[0101] The feature representation output by the encoder is then input into the fully connected layer of the Transformer network and mapped to the prediction result of the operation and maintenance data:

[0102] Y=ZW y +b y ;

[0103] In the formula, Y is the prediction result of operation and maintenance data; is the weight matrix of Y, and f is the length of the prediction result; is the bias term of Y.

[0104] To train the model and evaluate its performance, MSE is used as the loss function for the Transformer network:

[0105]

[0106] In the formula, y ji and They represent the true value and predicted value of the j-th variable at the ith step, respectively. The choice of MSE loss function is based on its wide application and mathematical properties in regression problems.

[0107] The present invention also provides an operation and maintenance data prediction system based on adaptive adjustment of position information, which is implemented based on the above-mentioned operation and maintenance data prediction method based on adaptive adjustment of position information, and includes: a data acquisition module, a position coding embedding module, a sequence decomposition module, a sequence fusion module, a feature coding module and a data prediction module.

[0108] The data collection module is used to collect historical operation and maintenance data and construct an operation and maintenance time series data set.

[0109] The position coding embedding module is used to embed the position coding of all operation and maintenance time series data in the operation and maintenance time series data set.

[0110] The sequence decomposition module is used to decompose the operation and maintenance time series data after embedding position encoding to obtain the seasonal items and trend items of the operation and maintenance time series data.

[0111] The sequence fusion module constructs an adaptive term using the seasonal term and the adjusted trend term, and fuses the seasonal term, the adjusted trend term and the adaptive term to obtain the fused time series feature representation.

[0112] The feature encoding module uses the encoder of the Transformer network to perform deep feature extraction on the fused time series feature representation.

[0113] The data prediction module inputs the feature representation output by the feature encoding module into the fully connected layer of the Transformer network for prediction, and uses the mean square error loss function to optimize the parameters of the Transformer network.

[0114] The advantages of the method of the present invention are verified by experiments as follows:

[0115] 1. Data Description

[0116] In order to verify the effectiveness of the time series prediction performed by the present invention, an embodiment of the present invention adopts the ETT power operation and maintenance data set, which contains a variety of operation and maintenance indicators, such as load and oil temperature, including a data set sampled every hour (ETTh2) and a data set sampled every 15 minutes (ETTm2).

[0117] 2. Description of Benchmark Method

[0118] In order to evaluate the effectiveness of the present invention, a comparative verification scheme is designed in the embodiment of the present invention. Specifically, the method of the present invention is compared with the existing mainstream time series models. These models can be divided into two categories: Transformer network-based models: iTransformer, Non-stationary, Autoformer and non-Transformer models: TimesNet, DLinear, SCINet.

[0119] 3. Experimental Setup

[0120] In order to ensure a fair comparison between the model of the present invention and the comparison model, the experiment strictly reproduced these comparison methods according to the methods described in the iTransformer paper, and used the officially provided or hosted code for fair comparison. All experiments were conducted in an environment equipped with an NVIDIA Tesla P100 GPU and implemented using the PyTorch framework. The optimizer selected Adam, the initial learning rate was set in the range of {0.0001, 0.0005, 0.001}, the number of training rounds was 10 rounds, and the batch size was 32. The hyperparameter adjustment range of all Transformer layers in the model is {2, 3, 4}, the hidden layer dimension adjustment range is set to {128, 256, 512}, and the Dropout value is set to 0.1. In addition, the evaluation indicators use mean square error (MSE) and mean absolute error (MAE), which are widely used in time series prediction tasks and can fully reflect the prediction accuracy of the model.

[0121] 4. Results of Algorithm Performance Comparison

[0122] In order to compare the performance of different models in the prediction of operation and maintenance time series data, the experiment uniformly sets the input length to 96 time steps, and the prediction lengths of the evaluation models are: 96, 192, 336, and 720 time steps. The experimental results are shown in Table 1 below, where MSE is the mean square error and MAE is the mean absolute error.

[0123] Table 1 Comparison of prediction results

[0124]

[0125]

[0126] It can be seen from the experimental results in Table 1 that the time series prediction method proposed in the present invention achieves state-of-the-art performance in 16 experimental settings. Compared with the existing best Transformer-based model iTransformer and the best non-Transformer-based model TimesNet, the MSE is reduced by 1.40% and 5.91% on average, respectively. This result strongly proves that the embedding effect can be significantly improved by decomposing the time series and regulating the trend and position information using seasonal cycles. In addition, some other latest innovative models, such as DLinear through seasonal trend decomposition and channel independence, Non-stationary that constructs time series information by building a non-stationary attention mechanism, and SCINet that designs recursive downsampling convolution interactions, etc., although these methods model time series information to varying degrees, they still cannot surpass the performance of the present invention in most cases. In general, the experimental results of the present invention on real operation and maintenance data sets show excellent performance and outstanding applicability. This further verifies the robustness of its time series modeling capabilities, while emphasizing the importance of accurately identifying time series position information in time series tasks.

[0127] The present invention effectively solves the problem that standard position embedding is not effective in time series prediction tasks through an improved adaptive position embedding mechanism, and can better capture the periodicity and long-term trend information in operation and maintenance data. By removing the noise position information in the trend item, the interference of irrelevant information on model prediction is effectively reduced, and the prediction accuracy is significantly improved. The periodic information in the seasonal item is used to guide the adjustment of the trend item, so that the model can adaptively focus on the important features in the time series data, and enhance the flexibility and generalization ability of the model. In addition, by linearly transforming and fusing the seasonal item, the adjusted trend item and the adaptive item, a richer and more accurate time series representation can be obtained, providing high-quality feature input for subsequent prediction tasks. In summary, the present invention effectively improves the representation ability of time series data through innovative time series decomposition and adaptive adjustment mechanism. This method has good interpretability in theory and has also shown superior performance in practice, which is of great significance for improving the accuracy of operation and maintenance prediction.

[0128] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. Only the preferred embodiments of the present invention are expressed. The description is more specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. As long as there is no contradiction in the combination of these technical features, they should be considered as the scope recorded in this specification.

[0129] It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these modifications and improvements all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for predicting operation and maintenance data based on adaptive adjustment of location information, characterized in that: The following steps are involved: Step S1: Obtain historical operation and maintenance data and construct a time series data set of operation and maintenance data; Step S2: position encoding and embedding all time series data in the time series data set, and decomposing all embedded position encoded time series data into seasonal items and trend items; Step S3: adjusting the trend item according to the periodic information in the season item, and constructing an adaptive item for each time series data using the season item and the adjusted trend item; Step S4: fusing the seasonal item, the adjusted trend item and the adaptive item of each time series data respectively to obtain a fused time series feature representation; Step S5: Input the fused time series feature representation into the Transformer network to obtain the prediction result of the operation and maintenance data.

2. The method for predicting operation and maintenance data based on adaptive adjustment of location information according to claim 1, characterized in that: The expression for position encoding and embedding all the time series data in the time series data set in step S2 is: X PE =X+PE(X); In the above formula, X PE is the time series data after embedding position encoding; X is the time series data; PE(X) means position embedding of X; pos is the position index; i is the dimension index; d is the embedding dimension.

3. The method for predicting operation and maintenance data based on adaptive adjustment of location information according to claim 2, characterized in that: The expression for decomposing the time series data after embedding position encoding into seasonal terms and trend terms in step S2 is: D t =AvgPool(Padding(X PE )); D s =X PE -D t ; In the above formula, D t , D s They are the trend item and seasonal item of the time series data set respectively; AvgPool means moving average; Padding means filling the sequence.

4. The method for predicting operation and maintenance data based on adaptive adjustment of location information according to claim 3, characterized in that: The expression for adjusting the trend item according to the periodic information in the season item in step S3 is: D p =D t -PE(D t ); Where D p is the adjusted trend item.

5. The method for predicting operation and maintenance data based on adaptive adjustment of location information according to claim 4, characterized in that: The expression for constructing the adaptive term for each time series data using the seasonal term and the adjusted trend term in step S3 is: H = Ds + Dp⊙softmax(Ds); Where H is the adaptive term; ⊙ represents the corresponding multiplication of elements; and softmax is the activation function.

6. The method for predicting operation and maintenance data based on adaptive adjustment of location information according to claim 1, characterized in that: The expression for fusing the seasonal term, the adjusted trend term and the adaptive term in step S4 is: E=(D s W s +b s )+(D p W p +b p )+(HW h +b h ); Where E is the fused time series feature representation; D s , D p , H are the seasonal term, adjusted trend term and adaptive term of the time series data set respectively; W s , W p , W h D s , D p , the weight matrix of H; b s 、b p 、b h D s , D p , the bias term of H.

7. The method for predicting operation and maintenance data based on adaptive adjustment of location information according to claim 1, characterized in that: In step S5, the fused time series feature representation is used as the input sequence of the Transformer network, the encoder of the Transformer network is used to perform deep feature extraction on the input sequence, and the feature representation output by the encoder is input into the fully connected layer of the Transformer network to be mapped into the prediction result of the operation and maintenance data.

8. The method for predicting operation and maintenance data based on adaptive adjustment of location information according to claim 7, characterized in that: The encoder calculates the correlation between the query vector, key vector and value vector of the input sequence through a multi-head attention mechanism, models the global dependency relationship of each time position in the input sequence, performs residual connection and layer normalization on the output of the multi-head attention mechanism, inputs the feedforward neural network for nonlinear transformation, and again performs residual connection and layer normalization to obtain the final encoder output; The expression of the multi-head attention mechanism is: MultiHead(Q,K,V)=Concat(head1,...,head h )W O ; head i =Attention(QW i Q ,KW i K ,VW i v ); The expression of the residual connection and layer normalization processing is: E'=LayerNorm(E+MultiHead(E,E,E)); The expression for the nonlinear transformation of the input feedforward neural network is: FFN(E')=max(0,(E'W1+b1)W2+b2); The output of the encoder is expressed as: Z = LayerNorm(E'+FFN(E')); In the above formula, MultiHead is a multi-head attention mechanism; Q, K, and V are the query vector, key vector, and value vector of the multi-head attention mechanism respectively; Concat is the connection function; head i is the i-th head of the multi-head attention mechanism; W O is a learnable parameter; Attention is the attention weight; W i Q , W i K , W i v Head i The projection matrix of the query vector, key vector and value vector; softmax is the activation function; d is the embedding dimension; E' is the time series feature representation after residual connection and layer normalization; E is the fused time series feature representation; Z is the output of the encoder; LayerNorm is the layer normalization function; FFN is the output of the feedforward neural network; W1 and W2 are the weight matrices of E'; b1 and b2 are the bias terms of E'.

9. The method for predicting operation and maintenance data based on adaptive adjustment of location information according to claim 8, characterized in that: The expression of the prediction result of the operation and maintenance data in step S5 is: Y=ZW y +b y ; In the formula, Y is the prediction result of operation and maintenance data; is the weight matrix of Y, and f is the length of the prediction result; is the bias term of Y.

10. An operation and maintenance data prediction system based on adaptive adjustment of location information, implemented based on an operation and maintenance data prediction method based on adaptive adjustment of location information as claimed in any one of claims 1 to 9, characterized in that: include: Data acquisition module, position coding embedding module, sequence decomposition module, sequence fusion module, feature coding module and data prediction module; The data collection module collects historical operation and maintenance data and constructs an operation and maintenance time series data set; The position code embedding module is used to embed the position code of all the operation and maintenance time series data in the operation and maintenance time series data set; The sequence decomposition module decomposes the operation and maintenance time series data after embedding position coding to obtain the seasonal items and trend items of the operation and maintenance time series data; The sequence fusion module: constructs an adaptive term using the seasonal term and the adjusted trend term, fuses the seasonal term, the adjusted trend term and the adaptive term, and obtains a fused time series feature representation; The feature encoding module: uses the encoder of the Transformer network to perform deep feature extraction on the fused time series feature representation; The data prediction module inputs the feature representation output by the feature encoding module into the fully connected layer of the Transformer network for prediction, and optimizes the parameters of the Transformer network using the mean square error loss function.

Images