Feature prediction method and device for non-stationary time series data

Abstract

The invention provides a non-stationary time series data-oriented feature prediction method and device, and relates to the technical field of time series prediction, and the method comprises the steps: obtaining a multi-source monitoring sequence; performing calculation according to a time sequence and a wavelet base sequence in the multi-source monitoring sequence to obtain a waveform coefficient, and performing optimal wavelet decomposition on the multi-source monitoring sequence under different decomposition layer numbers through the waveform coefficient to obtain a plurality of low-frequency components and a plurality of high-frequency components; inputting the plurality of low-frequency components into a non-stationary attention mechanism for sequence stationary processing and prediction to obtain a low-frequency prediction result; predicting the plurality of high-frequency components based on an attention mechanism of a preset sequence model to obtain a high-frequency prediction result, the attention mechanism including a feature attention mechanism and a time sequence attention mechanism; and performing reconstruction based on the low-frequency prediction result and the high-frequency prediction result to obtain a time sequence feature prediction result. The problem that the prediction accuracy is reduced is solved.

Description

Technical Field

[0001] This invention relates to the field of time series forecasting technology, and more specifically, to a method and apparatus for feature prediction of non-stationary time series data. Background Technology

[0002] In the field of time series forecasting technology, the structural characteristics of engineering structures, the diversity of external loads, and the continuous dynamic changes of environmental factors such as climate and traffic all work together to affect the monitoring data of vibration and displacement of structures, often resulting in a complex unstable state. When using existing time series data forecasting models to process unstable states, the models struggle to accurately capture the non-stationary characteristics in the data due to significant differences between the basic assumptions of the models and the properties of the actual monitoring data, leading to a decrease in forecast accuracy. Therefore, there is an urgent need for feature forecasting methods and devices for non-stationary time series data to solve the problem of reduced forecast accuracy. Summary of the Invention

[0003] The purpose of this invention is to provide a feature prediction method for non-stationary time-series data to improve the aforementioned problems. To achieve this objective, the technical solution adopted by this invention is as follows:

[0004] Firstly, this application provides a feature prediction method for non-stationary time series data, including:

[0005] Acquire multi-source monitoring sequences, which include structural response monitoring sequences and environmental load sequences;

[0006] Waveform coefficients are calculated based on the time series and preset wavelet basis sequences in the multi-source monitoring sequence. The waveform coefficients are then used to perform optimal wavelet decomposition on the multi-source monitoring sequence at different decomposition levels to obtain multiple low-frequency components and multiple high-frequency components.

[0007] Multiple low-frequency components are input into a preset non-stationary attention mechanism for sequence stationarity processing and prediction to obtain low-frequency prediction results.

[0008] The high-frequency components are predicted based on the attention mechanism of the preset sequence model to obtain the high-frequency prediction result. The attention mechanism includes a feature attention mechanism and a temporal attention mechanism.

[0009] Based on the low-frequency prediction results and the high-frequency prediction results, the temporal feature prediction results are obtained by reconstruction.

[0010] Secondly, this application also provides a feature prediction device for non-stationary time series data, including:

[0011] The acquisition module is used to acquire multi-source monitoring sequences, which include structural response monitoring sequences and environmental load sequences;

[0012] The decomposition module is used to calculate waveform coefficients based on the time series and preset wavelet basis sequences in the multi-source monitoring sequence. The waveform coefficients are then used to perform optimal wavelet decomposition on the multi-source monitoring sequence at different decomposition levels to obtain multiple low-frequency components and multiple high-frequency components.

[0013] The first prediction module is used to input multiple low-frequency components into a preset non-stationary attention mechanism for sequence stationarity processing and prediction to obtain low-frequency prediction results.

[0014] The second prediction module is used to predict multiple high-frequency components based on the attention mechanism of a preset sequence model to obtain high-frequency prediction results. The attention mechanism includes a feature attention mechanism and a temporal attention mechanism.

[0015] A construction module is used to reconstruct the time-series feature prediction results based on the low-frequency prediction results and the high-frequency prediction results.

[0016] The beneficial effects of this invention are as follows:

[0017] This invention selects the optimal wavelet basis from multi-source monitoring sequences at different decomposition levels, decomposes the multi-source monitoring sequences using the optimal wavelet basis to obtain low-frequency and high-frequency components, and predicts the low-frequency components using a non-stationary attention mechanism. This dynamically focuses on key information in the sequence, performs stationarization processing on the low-frequency components, and predicts future trends based on historical data, thus obtaining low-frequency prediction results. The high-frequency components are predicted using the attention mechanism of the sequence model, which focuses on key features and temporal relationships in the data, thereby improving the accuracy and robustness of the prediction. The reconstructed prediction results comprehensively reflect the dynamic behavior of the engineering structure and the impact of environmental loads, providing a scientific basis for structural health monitoring, safety assessment, and maintenance decisions, and solving the problem of reduced prediction accuracy.

[0018] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the feature prediction method for non-stationary time series data described in this embodiment of the invention;

[0021] Figure 2 The cable force sequence in this embodiment of the invention is shown under different decomposition levels. Line chart;

[0022] Figure 3 The anchor span cable force sequence in this embodiment of the invention under different decomposition layers Line chart;

[0023] Figure 4 The stress sequence of the main beam in this embodiment of the invention under different decomposition layers Line chart;

[0024] Figure 5 This is a graph showing the prediction effect of the non-stationary feature-focusing network model on the cable force monitoring sequence in the first quarter of this invention.

[0025] Figure 6 This is a graph showing the prediction effect of the non-stationary feature-focusing network model on the cable force monitoring sequence in the second quarter, according to an embodiment of the present invention.

[0026] Figure 7 This is a graph showing the prediction effect of the non-stationary feature-focusing network model on the cable force monitoring sequence in the third quarter of this invention.

[0027] Figure 8 This is a graph showing the prediction effect of the non-stationary feature-focusing network model on the cable force monitoring sequence in the fourth quarter of this invention.

[0028] Figure 9 This is a schematic diagram of the feature prediction device for non-stationary time-series data as described in an embodiment of the present invention.

[0029] The diagram is labeled as follows: 800, Feature prediction device for non-stationary time series data; 801, Processor; 802, Memory; 803, Multimedia component; 804, I / O interface; 805, Communication component. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0031] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0032] Example 1:

[0033] This embodiment provides a feature prediction method for non-stationary time series data.

[0034] See Figure 1 The figure shows that the method includes steps S1 to S5, including:

[0035] S1: Obtain a multi-source monitoring sequence, which includes a structural response monitoring sequence and an environmental load sequence;

[0036] S2: Calculate waveform coefficients based on the time series and preset wavelet basis sequence in the multi-source monitoring sequence, and perform optimal wavelet decomposition on the multi-source monitoring sequence under different decomposition levels using the waveform coefficients to obtain multiple low-frequency components and multiple high-frequency components.

[0037] To clarify the specific acquisition methods for multiple low-frequency components and multiple high-frequency components, step S2 includes S21 to S25, specifically:

[0038] S21: Obtain multiple decomposition levels;

[0039] S22: Calculate the waveform coefficients based on the time series and the wavelet basis sequence;

[0040] To clarify the specific method for obtaining the waveform coefficients, step S22 includes S221 to S224, specifically:

[0041] S221: Calculate the approximate error signal based on the time series and the wavelet basis sequence;

[0042] In this step, the expression for the approximate error signal is:

[0043] (1)

[0044] In the above formula (1), This is an approximate error signal. It is a time series. It is a constant. It is a wavelet basis sequence;

[0045] The approximation error signal represents sequence sum Sequence in The similarity at time points; the wavelet basis sequence includes the Haar wavelet basis, Morlet wavelet basis, Mexihat wavelet basis, or db4 wavelet;

[0046] The larger the absolute value of the approximate error signal, the smaller the similarity between the time series and the wavelet basis sequence.

[0047] S222: Based on a preset adjustment ratio parameter, the approximate error signal is converted into approximate error energy, and the minimum signal energy is obtained by taking the minimum derivative of the approximate error energy;

[0048] In this step, the approximate error signal is... Convert to approximate error energy ,Will To obtain the constant The value of the constant Substituting the value of into the approximate error signal expression, we obtain the minimized signal energy;

[0049] The expression for minimizing signal energy is:

[0050] (2)

[0051] In the above formula (2), The coefficients are the approximate error signal energy. The integral symbol is used. The square of the time series, For time intervals, It is a time series. It is a wavelet basis sequence. The width of the time window. It is the square of the wavelet basis sequence.

[0052] S223: Perform normalized relative error processing on the minimized signal energy to obtain the signal energy relative error;

[0053] In this step, the expression for the relative error of the signal energy is:

[0054] (3)

[0055] In the above formula (3), The relative error of the energy sign, To minimize signal energy, It is a time series. It is a wavelet basis sequence. The width of the time window. The integral symbol is used. The square of the time series, The square of the wavelet basis sequence, For time intervals.

[0056] S224: Based on the relative error of the signal energy, perform integral calculation on the time series and the wavelet basis sequence to obtain the waveform coefficients.

[0057] In this step, the waveform coefficients are:

[0058] (4)

[0059] In the above formula (4), Indicates waveform coefficients. It is a time series. It is a wavelet basis sequence. The width of the time window. The integral symbol is used. The square of the time series, The square of the wavelet basis sequence, For time intervals.

[0060] The waveform coefficient The larger the time series With the wavelet basis sequence The higher the degree of relevance, the better.

[0061] S23: Discretize the waveform coefficients, perform waveform similarity analysis on the time series and the wavelet basis sequence using the discretized waveform coefficients, and select the optimal wavelet basis based on the analysis results to obtain the optimal wavelet basis;

[0062] In this step, the expression for the discrete waveform coefficients is:

[0063] (5)

[0064] In the above formula (5), These are the discrete waveform coefficients. The total number of sampling points. For signal In the The value of each sampling point, For signal In the The value of each sampling point;

[0065] Preferably, the waveform similarity between the multi-source monitoring sequence and the db4 wavelet basis is the highest, reaching 99.98%, and the db4 wavelet basis is selected as the optimal wavelet basis.

[0066] S24: Perform wavelet decomposition on the multi-source monitoring sequence according to the optimal wavelet basis to obtain multiple initial low-frequency components and multiple initial high-frequency components;

[0067] In this step, the multi-source monitoring sequence is decomposed using the db4 wavelet basis.

[0068] S25: Process the multiple initial high-frequency components according to the structural response monitoring sequence under different decomposition levels to obtain periodic components. Optimize the multiple initial low-frequency components and multiple initial high-frequency components using the periodic components to obtain multiple low-frequency components and multiple high-frequency components.

[0069] To clarify the specific processing steps for multiple low-frequency components and multiple high-frequency components, step S25 includes S251 to S254, specifically:

[0070] S251: Based on the structural response monitoring sequence under different decomposition levels, multiple initial high-frequency components are superimposed to obtain periodic components;

[0071] S252: Calculate the dynamic time bending distance based on the periodic components and the environmental load sequence;

[0072] To clarify the specific method for obtaining the dynamic time bending distance, step S252 includes steps S2521 to S2523, specifically:

[0073] S2521: Construct a matrix based on the periodic components and the environmental load sequence, and obtain the time series similarity degree by calculating the distance values ​​in the matrix;

[0074] In this step, the periodic component is The periodic component is set as The environmental load sequence is set as follows: The periodic component The environmental load sequence The periodic components and the environmental load sequence are used to construct a matrix to obtain a matrix expression;

[0075] The matrix expression is:

[0076] (6)

[0077] In the above formula (6), For data matrix, For elements and The distance between them , ,…, For set The elements in , ,…, For set The elements in

[0078] Preferably, the matrix Each element in , , , Representing time series point With time series point The Euclidean distance between them.

[0079] S2522: Evaluate the preset curved path according to the time series similarity criterion, and determine the curved path when the curved path meets the time series similarity criterion;

[0080] In this step, the bending path is , for The One element;

[0081] The criteria for time series similarity are as follows:

[0082] (1) Boundedness: ;

[0083] (2) Boundary conditions: , ;

[0084] (3) Monotonicity and continuity: For , Must meet , .

[0085] S2523: Select the optimal path for the curved path based on the preset cumulative cost matrix, and obtain the dynamic time-varying curved distance by calculating the total cost of the optimal path.

[0086] In this step, the cumulative cost matrix is ​​expressed as follows:

[0087] (7)

[0088] In the above formula (7), For periodic components The former Individual element and environmental load sequence The former The minimum distance between elements For periodic components The Middle Individual element and environmental load sequence The Middle The distance between elements;

[0089] The expression for the dynamic time bending distance is:

[0090] (8)

[0091] In the above formula (8), For periodic components and environmental load sequence The dynamic time-warped distance between them For path, For path The total number of elements in the middle. For curved paths The Middle The distance between elements.

[0092] Preferably, when the When the .

[0093] S253: The preset minimum decomposition layer number is weighted according to the dynamic time bending distance and the preset entropy method to obtain the optimal decomposition layer number;

[0094] In this step, to clarify the specific method for obtaining the entropy value, the following are included:

[0095] The environmental load sequence under different decomposition levels is normalized to obtain the dynamic time-normalized value.

[0096] The expression for the dynamic time warp normalized value is:

[0097] (9)

[0098] In the above formula (9), For the first The environmental load sequence in the _ ... Normalized values ​​under layer decomposition For the first The environmental load sequence in the _ ... Specific values ​​under layer decomposition, The number of environmental load sequences. The number of decomposition levels;

[0099] The entropy value of the decomposition level is calculated based on the dynamic time warping normalized value, and the deviation of the decomposition level is calculated using the entropy value of the decomposition level.

[0100] The entropy expression for the number of decomposition layers is:

[0101] (10)

[0102] In the above formula (10), For the first The entropy value of the number of layers in the layer decomposition. It is the natural logarithm. The number of environmental load sequences. For the first The environmental load sequence in the _ ... Normalized values ​​under layer decomposition;

[0103] The expression for the deviation of the number of decomposition layers is:

[0104] (11)

[0105] In the above formula (11), For the first Deviation in the number of layers in the layer decomposition For the first The entropy value of the number of layers in the layer decomposition;

[0106] The deviation of the decomposition level is normalized to obtain the weight. The entropy weighted distance under the decomposition level is then calculated by calculating the weight.

[0107] The weight expression after the deviation is normalized is:

[0108] (12)

[0109] In the above formula (12), For the first The weights after normalizing the deviation of the layer decomposition layer number. For the first Deviation in the number of layers in the layer decomposition The number of decomposition levels;

[0110] The entropy-weighted distance expression for the decomposition level is:

[0111] (13)

[0112] In the above formula (13), For the bridge in the Entropy-weighted distance at each level of the layer decomposition The number of environmental load sequences. For the first The weights after normalizing the deviation of the layer decomposition layer number. For the first Data sequence after layer decomposition With the A sequence of environmental loads The dynamic time-normalized distance between them.

[0113] S254: Based on the optimal decomposition layer, optimize the selection of multiple initial low-frequency components and multiple initial high-frequency components to obtain multiple low-frequency components and multiple high-frequency components.

[0114] like Figures 2 to 4 As shown, in this step, the non-stationary structural response monitoring sequences, including the cable force SL-S1-05, anchor strand force SL-01-01, and main beam stress YB-05-01 sequences, under different decomposition levels... The line graph shows that, based on the selected optimal decomposition layer number k=6 for the SL-S1-05 monitoring sequence of cable force, 4 for the SL-01-01 monitoring sequence of anchor strand force, and 3 for the YB-05-01 monitoring sequence of main beam stress. The optimal decomposition layer number k for the environmental load sequence is the same as the optimal decomposition layer number for the structural response monitoring sequence related to its factors.

[0115] S3: Input the multiple low-frequency components into a preset non-stationary attention mechanism for sequence stationarity processing and prediction to obtain low-frequency prediction results;

[0116] In this step, the non-stationary attention mechanism dynamically focuses on key information in the sequence to predict low-frequency components, performs stationarization processing on low-frequency components, and predicts future trends based on historical data, thereby obtaining low-frequency prediction results.

[0117] To clarify the specific method for obtaining low-frequency prediction results, step S3 includes steps S31 to S36, specifically:

[0118] S31: Input the multiple low-frequency components into the non-stationary attention mechanism for normalization processing to obtain normalized low-frequency components;

[0119] In this step, the expression for the standardized low-frequency component is:

[0120] (14)

[0121] In the above formula (14), To standardize low-frequency components, Low-frequency components, This is the average value of the low-frequency components. This represents the standard deviation of the low-frequency components.

[0122] S32: In the case of sequence stabilization, each of the low-frequency components is input into a preset converter model for scaling transformation, and the normalized low-frequency components are received through the converter model to obtain a sequence stabilization result.

[0123] In this step, the converter model is: The input low-frequency component is ,exist Each query token in the system can be accessed via To perform the calculation, the The converter model is a feedforward network, and after normalization, it is... For the specified low-frequency components Receive.

[0124] S33: Based on the sequence stationarity result, the query vector in the preset attention function is standardized to obtain multiple stationarized sequences;

[0125] S34: In the case of non-stationary sequence stabilization, multiple non-stationary sequences are obtained by performing sequence normalization calculation on multiple stationary sequences and eliminating non-stationary information;

[0126] S35: The non-stationary information to be eliminated is re-analyzed. Multiple destationary factors are obtained by calculating the normalized exponential function on the preset scalar and the preset displacement vector respectively.

[0127] S36: Based on the multiple destationary factors, multiple stationary sequences, and multiple non-stationary sequences, a low-frequency prediction result is obtained.

[0128] S4: Based on the attention mechanism of the preset sequence model, multiple high-frequency components are predicted to obtain high-frequency prediction results. The attention mechanism includes a feature attention mechanism and a temporal attention mechanism.

[0129] In this step, the attention mechanism of the sequence model predicts high-frequency components, focusing on key features and temporal relationships in the data, thereby improving the accuracy and robustness of the prediction.

[0130] To clarify the specific method for obtaining high-frequency prediction results, step S4 includes steps S41 to S44, specifically:

[0131] S41: Integrate the multiple high-frequency components to obtain a high-frequency component set;

[0132] S42: Calculate weighted high-frequency components for the high-frequency component set based on the convolutional long short-term memory unit in the feature attention mechanism, and input the weighted high-frequency components into the convolutional long short-term memory unit for time-series information iteration to obtain the encoder hidden state.

[0133] To clarify the specific method for obtaining the encoder's hidden state, step S42 includes steps S421 to S425, specifically:

[0134] S421: Calculate the hidden state of the current time step based on the high-frequency component set and the previous time step hidden state in the encoder according to the convolutional long short-term memory unit;

[0135] In this step, the hidden state of the current time step is:

[0136] (15)

[0137] In the above formula (15), For the current time step The hidden state, For ConvLSTM units of convolutional long short-term memory, For the previous time step The hidden state, For the current time step High-frequency components;

[0138] The high-frequency component is used as input data, and the previous time step... The hidden state is the hidden state of the previous time step in the encoder.

[0139] Preferably, For the encoder at the current time step The hidden state.

[0140] S422: Perform a convolution operation based on the hidden state of the previous time step in the encoder and the first component in the high-frequency component set to obtain the attention weight at the current time step;

[0141] In this step, the expression for the attention weight at the current moment is:

[0142] (16)

[0143] In the above formula (16), For the current time step and spatial location Attention weights , and All of these are learnable parameters. It is the hyperbolic tangent function. For the previous time step The hidden state, For the first One high-frequency component;

[0144] Preferably, Dimensions: , Dimensions: , Dimensions: , Dimensions: The first The first component is a high-frequency component.

[0145] S423: Calculate the importance of the structural response monitoring sequence by performing a normalized exponential function on the attention weights at the current moment;

[0146] In this step, the normalized exponential function is: The function, at the current time step For features The importance of the corresponding structural response monitoring sequence to be predicted is expressed as: This level of importance belongs to the real number space. ;

[0147] The importance of the structural response monitoring sequence is as follows:

[0148] (17)

[0149] In the above formula (17), For the current time step and spatial location The importance of For the current time step and spatial location Attention weights It is an exponential function. For all spatial locations The summation of the exponential attention scores.

[0150] S424: Based on the importance of the structural response monitoring sequence and the preset structural response characteristics at the current moment, a weighted high-frequency component is constructed;

[0151] In this step, the importance of the structural response monitoring sequence is determined. With the current time step feature Multiplying yields the weighted high-frequency components. The weighted high-frequency components are used as the input to the encoder;

[0152] The weighted high-frequency component is:

[0153] (18)

[0154] In the above formula (18), For the current time step The weighted high-frequency components, , …, For the current time step The importance coefficient of the feature , …, For the current time step The various high-frequency components;

[0155] Wherein, the current time step Each high-frequency component The structural response characteristics at the current moment;

[0156] S425: Input the weighted high-frequency components into the convolutional long short-term memory unit, iterate the time series information of the hidden state at the current time step, and obtain the encoder hidden state.

[0157] In this step, the hidden state of the current time step is updated;

[0158] The encoder is in a hidden state as follows:

[0159] (19)

[0160] In the above formula (19), The updated encoder is now in a hidden state. For ConvLSTM units of convolutional long short-term memory, For the previous time step The hidden state, For the current time step The weighted high-frequency components.

[0161] During the encoding phase, a Feature Attention (FPA) mechanism is employed to explore the correlation between the high-frequency components of each feature and the high-frequency components of the structural response monitoring sequence (represented as...). Based on correlation metrics, this method enables adaptive emphasis and weakening of high-frequency components of the original features, thereby generating weighted features. Subsequently, a Convolutional Long Short-Term Memory (ConvLSTM) encoder is used to iteratively process these weighted features for time-series information. This process ensures that each time step... Temporal coding hidden state They can all comprehensively and accurately encompass the complex relationships between various features.

[0162] S43: Calculate based on the hidden state of the previous time step in the decoder and the unit state in the decoder, and use the temporal attention mechanism to perform a weighted summation of the hidden states in the encoder to obtain the intermediate semantic vector;

[0163] In this processing step, existing technologies for processing non-stationary monitoring data generally result in prediction outputs exhibiting excessively high stationarity characteristics, accompanied by large prediction errors. To address this issue, this invention introduces a destationary attention mechanism. This mechanism effectively alleviates the over-stationarization phenomenon by reweighting key features in non-stationary data and reintegrating non-stationary information.

[0164] To clarify the specific method for obtaining the intermediate semantic vector, step S43 includes steps S431 to S433, specifically:

[0165] S431: Calculate the attention weight of the hidden state based on the hidden state of the previous time step in the decoder and the unit state in the decoder;

[0166] In this step, the attention weights for the hidden state are:

[0167] (20)

[0168] In the above formula (20), For the current time step Attention weights for each encoder hidden state. , and All of these are learnable parameters. It is the hyperbolic tangent function. The hidden state of the previous time step in the decoder. This refers to the unit state in the decoder. For the first The encoder at the current time step The hidden state of the output. The total number of encoders;

[0169] Preferably, 3D real vector: , A 3D real matrix: , A 3D real matrix: .

[0170] S432: Calculate the influence of the hidden state by performing a normalized exponential function on the attention weights of the hidden state;

[0171] In this step, The function is for the current time step. Attention weights for each encoder hidden state Normalization is performed to obtain the degree of influence of the encoder hidden layer state on the predicted value;

[0172] The degree of influence of the hidden state is:

[0173] (twenty one)

[0174] In the above formula (21), For time step Time The degree of influence of the encoder's hidden state For the current time step Attention weights for each encoder hidden state. It is an exponential function. This involves exponentially calculating and summing the attention weights for all encoder hidden states.

[0175] S433: Based on the influence of the temporal attention mechanism on the hidden state and the hidden state in the encoder, a weighted sum is performed to obtain an intermediate semantic vector.

[0176] In this step, the Temporal Attention (TPA) mechanism hides all encoder states. and time step Time The degree of influence of the encoder's hidden state A weighted calculation is performed to obtain the intermediate semantic vector. .

[0177] The intermediate semantic vector is:

[0178] (twenty two)

[0179] In the above formula (22), For time step The intermediate semantic vector, For time step Time The degree of influence of the encoder's hidden state For the first The encoder at the current time step The hidden state of the output. This represents the total number of encoders.

[0180] S44: Update the hidden state in the decoder based on the intermediate semantic vector from the previous time step and the output of the sequence model. Connect the updated hidden state and the intermediate semantic vector and input them into the fully connected layer of the temporal attention mechanism for prediction to obtain high-frequency prediction results.

[0181] In this step, the updated hidden state expression is:

[0182] (twenty three)

[0183] In the above formula (23), For time step The updated hidden state For spatiotemporal convolutional network TCN units, The hidden state of the previous time step in the decoder. For time step The output of the sequence model, For time step The intermediate semantic vector;

[0184] The high-frequency prediction result is as follows:

[0185] (twenty four)

[0186] In the above formula (24), For time step High-frequency prediction results and All are learnable parameter matrices for prediction. For time step The updated hidden state Time step The intermediate semantic vector, , All are bias terms;

[0187] in, A 1-dimensional real vector: , A 3D real matrix: , 3D real vector: , 3D real vector: , Scalar: The time step The intermediate semantic vector is the intermediate semantic vector of the previous time step.

[0188] S5: Reconstruct the time series feature prediction results based on the low-frequency prediction results and the high-frequency prediction results.

[0189] In this processing stage, the reconstructed time-series characteristic prediction results can comprehensively reflect the dynamic response characteristics of the engineering structure and the combined impact of environmental loads, providing a solid scientific basis for structural health monitoring, safety assessment, and maintenance strategy formulation. This process ensures that the overall trend of the original signal is preserved while efficiently identifying and handling local anomalies, thereby ensuring the accuracy and reliability of the analysis results.

[0190] This invention enables the prediction of structural response monitoring sequences with non-stationary characteristics on a quarterly basis, such as... Figures 5 to 8 The image shows the prediction performance of the NSFeatFocusNet non-stationary feature focusing network model on the cable force monitoring sequence for different seasons. Taking the fourth quarter of 2020 as the experimental subject, the mean absolute error (MAE) of the prediction by this invention is 2.216, the root mean square error (RMSE) is 2.867, and the coefficient of determination (R2) is 0.944.

[0191] The NSFeatFocusNet non-stationary feature focusing network model is used to predict structural response monitoring sequences by using multi-source monitoring sequences as features.

[0192] Example 2:

[0193] This embodiment provides a feature prediction device for non-stationary time series data, the device comprising:

[0194] The acquisition module is used to acquire multi-source monitoring sequences, which include structural response monitoring sequences and environmental load sequences;

[0195] The decomposition module is used to calculate waveform coefficients based on the time series and preset wavelet basis sequences in the multi-source monitoring sequence. The waveform coefficients are then used to perform optimal wavelet decomposition on the multi-source monitoring sequence at different decomposition levels to obtain multiple low-frequency components and multiple high-frequency components.

[0196] To clarify the specific steps for obtaining the selected decomposition module, the steps are as follows:

[0197] The acquisition unit is used to acquire multiple decomposition levels.

[0198] The first calculation unit is used to calculate waveform coefficients based on the time series and the wavelet basis sequence.

[0199] The selection unit is used to discretize the waveform coefficients, perform waveform similarity analysis on the time series and the wavelet basis sequence using the discretized waveform coefficients, and select the optimal wavelet basis based on the analysis results to obtain the optimal wavelet basis.

[0200] The decomposition unit is used to perform wavelet decomposition on the multi-source monitoring sequence according to the optimal wavelet basis to obtain multiple initial low-frequency components and multiple initial high-frequency components.

[0201] The first processing unit is used to process multiple initial high-frequency components according to the structural response monitoring sequence under different decomposition levels to obtain periodic components, and to optimize multiple initial low-frequency components and multiple initial high-frequency components through the periodic components to obtain multiple low-frequency components and multiple high-frequency components.

[0202] To clarify the specific acquisition method and steps of the first processing unit, the following are included:

[0203] The superposition subunit is used to superimpose multiple initial high-frequency components according to the structural response monitoring sequence under different decomposition levels to obtain periodic components;

[0204] The second calculation subunit is used to calculate the dynamic time bending distance based on the periodic component and the environmental load sequence.

[0205] The first processing subunit is used to perform weighted processing on the preset minimum decomposition layer number according to the dynamic time bending distance and the preset entropy value method to obtain the optimal decomposition layer number.

[0206] The selection subunit is used to optimize the selection of multiple initial low-frequency components and multiple initial high-frequency components according to the optimal decomposition layer number, so as to obtain multiple low-frequency components and multiple high-frequency components.

[0207] The first prediction module is used to input multiple low-frequency components into a preset non-stationary attention mechanism for sequence stationarity processing and prediction to obtain low-frequency prediction results.

[0208] The second prediction module is used to predict multiple high-frequency components based on the attention mechanism of a preset sequence model to obtain high-frequency prediction results. The attention mechanism includes a feature attention mechanism and a temporal attention mechanism.

[0209] To clarify the specific acquisition method and steps of the second prediction module, the following are included:

[0210] The second processing unit is used to integrate the multiple high-frequency components to obtain a high-frequency component set.

[0211] An iterative unit is used to calculate weighted high-frequency components on the high-frequency component set according to the convolutional long short-term memory unit in the feature attention mechanism, and input the weighted high-frequency components into the convolutional long short-term memory unit to iterate the time series information to obtain the encoder hidden state.

[0212] The second computational unit is used to calculate based on the hidden state of the previous time step in the decoder and the unit state in the decoder, and to obtain the intermediate semantic vector by weighted summation of the hidden states in the encoder through the temporal attention mechanism.

[0213] The update unit is used to update the hidden state in the decoder based on the intermediate semantic vector of the previous time step and the output of the sequence model. The updated hidden state and the intermediate semantic vector are concatenated and input into the fully connected layer in the temporal attention mechanism for prediction to obtain high-frequency prediction results.

[0214] A construction module is used to reconstruct the time-series feature prediction results based on the low-frequency prediction results and the high-frequency prediction results.

[0215] It should be noted that the specific manner in which each module performs its operation in the apparatus described in the above embodiments has been described in detail in the embodiments of the method, and will not be elaborated here.

[0216] Example 3:

[0217] Corresponding to the above method embodiments, this embodiment also provides a feature prediction device for non-stationary time series data. The feature prediction device for non-stationary time series data described below can be referred to in correspondence with the feature prediction method for non-stationary time series data described above.

[0218] Figure 9 This is a block diagram illustrating a feature prediction device 800 for non-stationary time-series data according to an exemplary embodiment. Figure 9 As shown, the feature prediction device 800 for non-stationary time-series data may include a processor 801 and a memory 802. The feature prediction device 800 for non-stationary time-series data may also include one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0219] The processor 801 controls the overall operation of the feature prediction device 800 for non-stationary time-series data to complete all or part of the steps in the aforementioned feature prediction method for non-stationary time-series data. The memory 802 stores various types of data to support the operation of the feature prediction device 800 for non-stationary time-series data. This data may include, for example, instructions for any application or method operating on the feature prediction device 800 for non-stationary time-series data, as well as application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as a keyboard, mouse, and buttons. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the feature prediction device 800 for non-stationary time-series data and other devices. Wireless communication includes Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, or an NFC module.

[0220] In an exemplary embodiment, the feature prediction device 800 for non-stationary time-series data may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the feature prediction method for non-stationary time-series data described above.

[0221] Example 4:

[0222] Corresponding to the above method embodiments, this embodiment also provides a medium. The medium described below can be referred to in conjunction with the feature prediction method for non-stationary time series data described above.

[0223] A medium storing a computer program, which, when executed by a processor, implements the steps of the feature prediction method for non-stationary time-series data described in the above method embodiments.

[0224] The medium can specifically be any medium capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0225] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0226] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A feature prediction method for non-stationary time series data, characterized in that, include: Acquire multi-source monitoring sequences, which include structural response monitoring sequences and environmental load sequences; Waveform coefficients are calculated based on the time series and preset wavelet basis sequences in the multi-source monitoring sequence. The waveform coefficients are then used to perform optimal wavelet decomposition on the multi-source monitoring sequence at different decomposition levels to obtain multiple low-frequency components and multiple high-frequency components. Multiple low-frequency components are input into a preset non-stationary attention mechanism for sequence stationarity processing and prediction to obtain low-frequency prediction results. The high-frequency components are predicted based on the attention mechanism of the preset sequence model to obtain the high-frequency prediction result. The attention mechanism includes a feature attention mechanism and a temporal attention mechanism. Based on the low-frequency prediction results and the high-frequency prediction results, the temporal feature prediction results are obtained by reconstruction.

2. The feature prediction method for non-stationary time series data according to claim 1, characterized in that, Waveform coefficients are calculated based on the time series and preset wavelet basis sequences in the multi-source monitoring sequence. These waveform coefficients are then used to perform optimal wavelet decomposition on the multi-source monitoring sequence at different decomposition levels, yielding multiple low-frequency components and multiple high-frequency components, including: Obtain multiple decomposition levels; The waveform coefficients are obtained by calculation based on the time series and the wavelet basis sequence; The waveform coefficients are discretized, and the time series and the wavelet basis sequence are analyzed for waveform similarity using the discretized waveform coefficients. Based on the analysis results, the optimal wavelet basis is selected to obtain the optimal wavelet basis. The multi-source monitoring sequence is decomposed using the optimal wavelet basis to obtain multiple initial low-frequency components and multiple initial high-frequency components. The initial high-frequency components are processed according to the structural response monitoring sequence under different decomposition levels to obtain periodic components. The initial low-frequency components and the initial high-frequency components are then optimized using the periodic components to obtain multiple low-frequency components and multiple high-frequency components.

3. The feature prediction method for non-stationary time series data according to claim 2, characterized in that, The initial high-frequency components are processed according to the structural response monitoring sequence under different decomposition levels to obtain periodic components. These periodic components are then used to optimize the initial low-frequency components and the initial high-frequency components, resulting in multiple low-frequency components and multiple high-frequency components, including: Multiple initial high-frequency components are superimposed based on the structural response monitoring sequences under different decomposition levels to obtain periodic components; The dynamic time bending distance is calculated based on the periodic components and the environmental load sequence. The optimal number of decomposition layers is obtained by weighting the preset minimum decomposition layer based on the dynamic time bending distance and the preset entropy value method. Based on the optimal decomposition layer, multiple initial low-frequency components and multiple initial high-frequency components are optimized and selected to obtain multiple low-frequency components and multiple high-frequency components.

4. The feature prediction method for non-stationary time series data according to claim 1, characterized in that, The sequence model includes an encoder and a decoder. Based on a preset sequence model attention mechanism, multiple high-frequency components are predicted to obtain high-frequency prediction results. The attention mechanism includes a feature attention mechanism and a temporal attention mechanism, comprising: The multiple high-frequency components are integrated to obtain a high-frequency component set. The high-frequency component set is calculated by the convolutional long short-term memory unit in the feature attention mechanism, and the weighted high-frequency component is input into the convolutional long short-term memory unit for time series information iteration to obtain the encoder hidden state. The intermediate semantic vector is obtained by calculating the hidden state of the previous time step in the decoder and the unit state in the decoder, and by weighting and summing the hidden states in the encoder through the temporal attention mechanism. The hidden state in the decoder is updated based on the intermediate semantic vector from the previous time step and the output of the sequence model. The updated hidden state and the intermediate semantic vector are then concatenated and input into the fully connected layer of the temporal attention mechanism for prediction, resulting in a high-frequency prediction result.

5. The feature prediction method for non-stationary time-series data according to claim 4, wherein the high-frequency component set is calculated by weighting the high-frequency components according to the convolutional long short-term memory unit in the feature attention mechanism, and the weighted high-frequency components are input into the convolutional long short-term memory unit for time-series information iteration to obtain the encoder hidden state, including: The hidden state of the current time step is obtained by calculating the high-frequency component set and the previous time step hidden state in the encoder based on the convolutional long short-term memory unit. The attention weights at the current time step are obtained by performing a convolution operation based on the hidden state of the previous time step in the encoder and the first component in the set of high-frequency components. The importance of the structural response monitoring sequence is obtained by calculating the attention weights at the current moment using a normalized exponential function. Weighted high-frequency components are constructed based on the importance of the structural response monitoring sequence and the preset structural response characteristics at the current moment; The weighted high-frequency components are input into the convolutional long short-term memory unit to iterate the time-series information of the hidden state at the current time step, thereby obtaining the encoder hidden state.

6. The feature prediction method for non-stationary temporal data according to claim 4, comprising calculating based on the hidden state of the previous time step in the decoder and the unit state in the decoder, and weighting and summing the hidden states in the encoder through the temporal attention mechanism to obtain an intermediate semantic vector, including: The attention weights of the hidden states are calculated based on the hidden state of the previous time step in the decoder and the unit states in the decoder. The influence of the hidden state is obtained by calculating the attention weights of the hidden state using a normalized exponential function. The intermediate semantic vector is obtained by weighting and summing the influence of the temporal attention mechanism on the hidden state and the hidden state in the encoder.

7. A feature prediction device for non-stationary time series data, characterized in that, include: The acquisition module is used to acquire multi-source monitoring sequences, which include structural response monitoring sequences and environmental load sequences; The decomposition module is used to calculate waveform coefficients based on the time series and preset wavelet basis sequences in the multi-source monitoring sequence. The waveform coefficients are then used to perform optimal wavelet decomposition on the multi-source monitoring sequence at different decomposition levels to obtain multiple low-frequency components and multiple high-frequency components. The first prediction module is used to input multiple low-frequency components into a preset non-stationary attention mechanism for sequence stationarity processing and prediction to obtain low-frequency prediction results. The second prediction module is used to predict multiple high-frequency components based on the attention mechanism of a preset sequence model to obtain high-frequency prediction results. The attention mechanism includes a feature attention mechanism and a temporal attention mechanism. A construction module is used to reconstruct the time-series feature prediction results based on the low-frequency prediction results and the high-frequency prediction results.

8. The feature prediction device for non-stationary time series data according to claim 7, characterized in that, The selection decomposition module includes: The acquisition unit is used to acquire multiple decomposition levels. The first calculation unit is used to calculate waveform coefficients based on the time series and the wavelet basis sequence. The selection unit is used to discretize the waveform coefficients, perform waveform similarity analysis on the time series and the wavelet basis sequence using the discretized waveform coefficients, and select the optimal wavelet basis based on the analysis results to obtain the optimal wavelet basis. The decomposition unit is used to perform wavelet decomposition on the multi-source monitoring sequence according to the optimal wavelet basis to obtain multiple initial low-frequency components and multiple initial high-frequency components. The first processing unit is used to process multiple initial high-frequency components according to the structural response monitoring sequence under different decomposition levels to obtain periodic components, and to optimize multiple initial low-frequency components and multiple initial high-frequency components through the periodic components to obtain multiple low-frequency components and multiple high-frequency components.

9. The feature prediction device for non-stationary time series data according to claim 8, characterized in that, The first processing unit includes: The superposition subunit is used to superimpose multiple initial high-frequency components according to the structural response monitoring sequence under different decomposition levels to obtain periodic components; The second calculation subunit is used to calculate the dynamic time bending distance based on the periodic component and the environmental load sequence. The first processing subunit is used to perform weighted processing on the preset minimum decomposition layer number according to the dynamic time bending distance and the preset entropy value method to obtain the optimal decomposition layer number. The selection subunit is used to optimize the selection of multiple initial low-frequency components and multiple initial high-frequency components according to the optimal decomposition layer number, so as to obtain multiple low-frequency components and multiple high-frequency components.

10. The feature prediction device for non-stationary time series data according to claim 7, characterized in that, The second prediction module includes: The second processing unit is used to integrate the multiple high-frequency components to obtain a high-frequency component set. An iterative unit is used to calculate weighted high-frequency components on the high-frequency component set according to the convolutional long short-term memory unit in the feature attention mechanism, and input the weighted high-frequency components into the convolutional long short-term memory unit to iterate the time series information to obtain the encoder hidden state. The second computational unit is used to calculate based on the hidden state of the previous time step in the decoder and the unit state in the decoder, and to obtain the intermediate semantic vector by weighted summation of the hidden states in the encoder through the temporal attention mechanism. The update unit is used to update the hidden state in the decoder based on the intermediate semantic vector of the previous time step and the output of the sequence model. The updated hidden state and the intermediate semantic vector are concatenated and input into the fully connected layer in the temporal attention mechanism for prediction to obtain high-frequency prediction results.

Images