Thermoplastic composite material damage detection method based on TimeMixer framework MLP

Abstract

According to the thermoplastic composite material damage detection method based on the TimeMixer framework MLP, a multi-source sensor acquisition system is deployed, micro-vibration sensors, temperature sensors, acoustic emission sensors, ultrasonic phase sensors and the like are installed in specific areas of the composite material, and synchronous sampling of data of the four types of sensors is achieved. Feature data preprocessing is carried out on the collected data, for example, wavelet transformation is carried out on micro-vibration signals to extract frequency spectrum energy and Higuchi fractal dimensions and the like; an MLP model based on a TimeMixer architecture is constructed, multi-source data are integrated into a unified tensor format through an input layer, a multi-scale hybrid module extracts multi-scale damage features, and a multi-task output head realizes damage level evaluation and future trend prediction, so that the problems of damage monitoring and maintenance of the thermoplastic composite material are effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of thermoplastic composite material damage detection, and in particular to a thermoplastic composite material damage detection method based on a TimeMixer architecture MLP. Background Art

[0002] Thermoplastic composites, with their exceptional properties such as lightweight, high strength, and recyclability, have been widely used in a wide range of fields, including aerospace, automotive manufacturing, and energy. However, due to their complex service environments, such as mechanical loads, temperature fluctuations, and chemical corrosion, thermoplastic composites are prone to damage during use. If these damages are not detected and addressed promptly, they can lead to structural failure and serious safety accidents. Therefore, effective health monitoring of thermoplastic composites is crucial.

[0003] At present, health monitoring technologies for thermoplastic composites are mainly divided into the following categories:

[0004] Nondestructive testing technologies: such as ultrasonic testing and radiographic testing. Ultrasonic testing detects internal defects by analyzing the propagation characteristics of ultrasonic waves in materials. However, its accuracy is limited for detecting tiny cracks or complex structures, and its detection speed is slow, making real-time monitoring difficult. Although radiographic testing can provide intuitive images of internal structures, it poses radiation hazards, requires high operator requirements, and has expensive equipment, making it unsuitable for large-scale online monitoring.

[0005] Single-sensor monitoring technologies, such as strain gauges for strain measurement and thermocouples for temperature measurement, only capture changes in a specific physical quantity and fail to fully reflect the complex damage state within a composite material. For example, it's difficult to determine potential damage caused by temperature fluctuations based solely on strain changes. Furthermore, single sensors are susceptible to interference from environmental factors, leading to inaccurate monitoring results.

[0006] Traditional machine learning monitoring technology: Some studies have attempted to use traditional machine learning algorithms to monitor composite material damage. However, traditional machine learning methods face numerous challenges when processing multi-source, heterogeneous data. For one thing, they struggle to effectively integrate data collected by different sensor types, resulting in low information utilization. Furthermore, traditional machine learning models have limited ability to model complex nonlinear relationships, making them unable to accurately capture the complex characteristics of thermoplastic composite material damage processes, thus affecting monitoring accuracy and reliability.

[0007] Furthermore, most existing health monitoring systems lack effective early warning mechanisms and system integration. In terms of early warning, either the thresholds are poorly set, prone to false alarms or missed alerts, or they lack a tiered early warning strategy, failing to provide targeted treatment recommendations based on the severity of damage. Regarding system integration, damage data management is haphazard, predictive maintenance functionality is incomplete, and integration with advanced technologies like digital twins is insufficient, making it difficult to meet the requirements for full lifecycle health management of thermoplastic composites in practical engineering projects.

[0008] In summary, the existing health monitoring technology for thermoplastic composites has many shortcomings in terms of accuracy, real-time performance, multi-source data fusion, and system integration. There is an urgent need for a more efficient, comprehensive, and intelligent health monitoring method to solve these problems. Summary of the Invention

[0009] To solve the above problems, the present invention proposes a thermoplastic composite damage detection method based on the TimeMixer architecture MLP. The specific steps are as follows, which are characterized by:

[0010] Step 1: Deploy a multi-source sensor acquisition system. Install micro-vibration sensors, temperature sensor arrays, acoustic emission sensors, and ultrasonic phase sensors at the composite structure's joints, support structures, geometric mutations, and external load areas to ensure coverage of stress concentration areas and potential damage areas. By integrating a multi-channel data acquisition module, synchronized sampling of data from the four types of sensors is achieved.

[0011] Step 2: Feature data preprocessing: perform wavelet transform on the micro-vibration signal to extract spectral energy and Higuchi fractal dimension, calculate the local temperature mutation rate for the temperature gradient, extract event count, energy integral and ringing number time domain parameters for the acoustic emission signal, and extract phase difference features for the ultrasonic wave;

[0012] Step 3: Build a TimeMixer architecture MLP model. The input layer integrates the four sensor data types into a unified tensor format, including vibration, acoustic emission, temperature, and ultrasonic feature data. The multiscale mixing module then uses Past Decomposable Mixing (PDM) to decompose historical data into seasonal and trend components, extracting multiscale damage features through bidirectional information flow. Finally, a multi-task output head outputs the mild, moderate, and severe damage levels of the thermoplastic composite material, as well as the damage area growth over the next 10 steps.

[0013] Step 4: A real-time monitoring and intelligent early warning system runs a lightweight TimeMixer model on the edge to extract and preliminarily classify damage features in real time. A threshold warning is triggered when the damage level is moderate or higher. The full TimeMixer model is deployed on the cloud to deeply analyze uploaded data and output high-precision damage location and trend predictions.

[0014] Step 5: System integration: Build a thermoplastic composite health monitoring platform that integrates the damage database, predictive maintenance module, and digital twin interface to support damage data retrieval.

[0015] The present invention provides a thermoplastic composite material damage detection method based on the TimeMixer architecture MLP, which has beneficial effects. The technical effects of the present invention are:

[0016] 1. The present invention achieves comprehensive coverage of stress concentration areas and potential damage areas of composite materials by deploying micro-vibration sensors, temperature sensor arrays, acoustic emission sensors, and ultrasonic phase sensors at key locations such as connection joints, support structures, geometric mutations, and external load areas of composite materials.

[0017] 2. The present invention performs targeted feature data preprocessing on different types of sensor signals. For example, wavelet transform is used to extract spectral energy and Higuchi fractal dimension from micro-vibration signals, effectively capturing the dynamic response of microcracks within the material and accurately extracting damage characteristics from high-frequency vibration signals. The local temperature mutation rate is calculated from the temperature gradient to quantify the mutation characteristics of the material surface temperature field and accurately identify areas of abnormal heat conduction. Time domain parameters such as event count, energy integral, and ringing number are extracted from acoustic emission signals to quantify the energy release pattern within the material. Phase difference characteristics are extracted from ultrasonic waves to reflect abnormalities in the material's internal structure.

[0018] 3. The MLP model based on the TimeMixer architecture constructed by the present invention has multi-scale hybrid analysis and multi-task prediction capabilities. The input layer integrates multi-source sensor data into a unified tensor format, and injects it through the embedding layer and position encoding to effectively retain data features and time sequence information. The multi-scale hybrid module uses PDM to decompose historical data into seasonal and trend components, and extracts multi-scale damage features through bidirectional information flow. This multi-scale analysis method for time series data can fully explore the changing laws of material damage at different time scales. The multi-task output head can not only accurately evaluate the mild, moderate, and severe damage levels of thermoplastic composites, but also predict the growth trend of the damage area in the next 10 steps. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the present invention;

[0020] Figure 2 This is the MLP network structure diagram of the present invention. DETAILED DESCRIPTION

[0021] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0022] The present invention relates to health monitoring of thermoplastic composite materials. Data is acquired through a multi-source sensor acquisition system, and after feature preprocessing, a TimeMixer architecture MLP model is constructed for analysis and prediction. A real-time monitoring and intelligent early warning system is set up to provide early warnings based on damage classification. The system integrates and builds a monitoring platform to achieve damage data management, effectively improving the level of composite material damage monitoring and maintenance. The invention flow chart is as follows Figure 1 As shown, the steps of the present invention are described in detail below.

[0023] Step 1: Deploy a multi-source sensor acquisition system. Install micro-vibration sensors, temperature sensor arrays, acoustic emission sensors, and ultrasonic phase sensors at the joints, support structures, geometric mutations, and external load areas of the composite structure to ensure coverage of stress concentration areas and potential damage areas. By integrating a multi-channel data acquisition module, synchronized sampling of the four types of sensor data is achieved.

[0024] Step 2: Feature data preprocessing: perform wavelet transform on the micro-vibration signal to extract spectral energy and Higuchi fractal dimension, calculate the local temperature mutation rate for the temperature gradient, extract event count, energy integral and ringing number time domain parameters for the acoustic emission signal, and extract phase difference features for the ultrasonic wave;

[0025] Step 2.1: Process the micro-vibration signal to extract the damage characteristics in the high-frequency vibration signal and capture the dynamic response of the micro-cracks inside the material.

[0026] Step 2.1.1 Wavelet transform to extract spectrum energy

[0027] The vibration signal x(n) is decomposed into multiple scale sub-bands through three layers using discrete wavelet transform:

[0028]

[0029] Among them, W j (k) is the wavelet coefficient at the kth time point after the jth layer wavelet decomposition, ψ j,k (n) is the db4 wavelet basis function, j is the number of decomposition levels, and k is the time index.

[0030] Sum the squares of the wavelet coefficients of each subband to calculate the spectral energy E j :

[0031]

[0032] Among them, E j is the spectral energy of the j-th wavelet subband.

[0033] Step 2.1.2 Higuchi fractal dimension calculation

[0034] The fractal dimension D is obtained by constructing a virtual curve and calculating the logarithmic relationship between its length and the scaling factor:

[0035]

[0036] Where L(m) is the length of the curve of the reconstructed time series when the scaling factor is m, m is the scaling factor, is the step size change of the i-th segment in the reconstructed time series.

[0037] Step 2.2 Temperature gradient treatment

[0038] Quantify the sudden change characteristics of the material surface temperature field and identify the abnormal heat conduction area. Calculate the ratio of the temperature difference ΔT at adjacent moments to the time step Δt:

[0039]

[0040] Among them, R t is the local temperature mutation rate, T t+1 is the temperature value at time t+1, T t is the temperature value at time t, and Δt is the sampling interval.

[0041] Step 2.3 Acoustic emission signal processing

[0042] Extract the time domain characteristics of acoustic emission events and quantify the energy release patterns inside the material.

[0043] Step 2.3.1 Event Counting and Energy Integration

[0044] Count the number of acoustic emission events per unit time N:

[0045] N=count(A(t)>10mV)

[0046] Where A(t) is the amplitude of the acoustic emission signal at time t. The above formula represents the number of events with an amplitude greater than 10mV and the energy E of a single event is calculated as follows:

[0047]

[0048] Among them, t1 and t2 are the start and end times of the event.

[0049] Step 2.3.2 Time domain parameter extraction

[0050] Duration τ: The time difference between the start and end of the event τ = t2-t1.

[0051] Ring count C: Detects the number of rings by zero-crossing rate:

[0052]

[0053] Among them, sign is the sign function.

[0054] Step 2.4 Ultrasonic phase difference processing

[0055] Step 2.4.1 Full Matrix Capture

[0056] The phased array ultrasonic probe transmits and receives signals to generate full matrix data A ij :

[0057] A ij (t) = Receive(i)·Transmit(j)

[0058] Among them, A ij (t) is the echo signal between the i-th receiving element and the j-th transmitting element of the phased array probe, t is time, Receive(i) represents the signal received by the i-th probe element, and Transmit(j) represents the signal transmitted by the j-th probe element.

[0059] Step 2.4.2 Phase difference calculation

[0060] Step 2.4.2.1 Fourier transform

[0061] For each echo signal A ij (t) Perform fast Fourier transform to extract its phase at a specific frequency

[0062] φ ij =∠(F{A ij (t)})

[0063] Among them, φ ij is the phase angle of the echo signal at a specific frequency, ∠ represents the complex phase angle, and F is the Fourier transform operator.

[0064] Step 2.4.2.2 Phase difference calculation

[0065] By comparing the phase difference Δφ of echoes from different paths, the abnormality of the internal structure of the material can be reflected:

[0066] Δφ=φ ij -φ ref

[0067] Among them, φ ref is the phase reference of the defect-free reference sample.

[0068] Step 3: Build a TimeMixer architecture MLP model. The input layer integrates the four types of sensor data into a unified tensor format, including feature data of vibration, acoustic emission, temperature, and ultrasound. Then, the multi-scale mixing module uses Past Decomposable Mixing (PDM) to decompose historical data into seasonal and trend components, and extracts multi-scale damage features through bidirectional information flow. Finally, a multi-task output head is used to output the mild, moderate, and severe grades of thermoplastic composite materials and the damage area growth in the next 10 steps. The MLP network structure is shown in the figure below. Figure 2 shown.

[0069] This step designs a multi-layer perceptron (MLP) model based on the TimeMixer architecture to implement multi-scale hybrid analysis and multi-task prediction of multi-source sensor data. The model is constructed from three parts: input layer integration, multi-scale hybrid module, and multi-task output head. The specific process is as follows:

[0070] Step 3.1 Input layer integration: unified tensor format and embedding layer

[0071] The micro-vibration, temperature, acoustic emission, and ultrasonic sensor feature data extracted in step 2 are converted into a unified tensor format and mapped to the dimensions required by the model through the embedding layer.

[0072] Step 3.1.1 Data format unification

[0073] Organize the multi-source data features preprocessed in step 2 into a time series tensor τ∈R T×F , where T is the time step, F is the feature dimension, and F=64.

[0074]

[0075] Among them, x t,f represents the f-th eigenvalue at the t-th time step.

[0076] Step 3.1.2 Embedding layer projection

[0077] The original features are mapped to the model hidden layer dimension D through the fully connected layer MLP, D = 128:

[0078] H0=ReLU(W embed ·τ+b embed )

[0079] Among them, H0 is the hidden layer feature after the embedding layer projection, W embed , b embed is a learnable parameter and ReLU is the ReLU function.

[0080] Step 3.1.3 Positional encoding injection

[0081] Add sine-cosine position encoding to a time series to preserve temporal order information:

[0082]

[0083] Among them, PE t,2i is a sinusoidal position code, PE t,2i+1 is the cosine position encoding, t is the time step, i is the feature dimension index, and D is the model hidden layer dimension.

[0084] The position coding injection H'0 formula is:

[0085] H'0=H0+PE

[0086] Step 3.2 Multiscale Hybrid Module

[0087] Decomposition, mixing and prediction of multi-scale time series are achieved through PDM and Future Multipredictor Mixing (FMM).

[0088] Step 3.2.1 PDM multi-scale decomposition and bidirectional mixing

[0089] Step 3.2.1.1 Multi-scale downsampling

[0090] Perform multi-scale pooling on the input sequence H'0 to generate a sequence of N scales Where s represents the scale level, N = 4, the formula is:

[0091]

[0092] in, is the pooling sequence of the i-th scale, s is the scale level, k i is the pooling kernel, and AvgPool is the average pooling function.

[0093] Step 3.2.1.2 Seasonality-Trend Decomposition

[0094] For each scale sequence The sliding average method is used to decompose the seasonal component S of the i-th scale i and trend component T i :

[0095]

[0096] Wherein, w is the sliding window size, w=24, and MovingAvg is the sliding average operation.

[0097] Step 3.2.1.3 Two-way hybrid strategy

[0098] The seasonal component aggregates seasonal characteristics step by step from fine-grained to coarse-grained:

[0099]

[0100] in, is the mixed seasonal feature, i represents the time step index, MLP seasonal For a multi-layer perceptron, the trend component gradually optimizes the trend characteristics from coarse granularity to fine granularity:

[0101]

[0102] in, is the trend feature after mixing, MLP trend It is a multi-layer perceptron.

[0103] Step 3.2.1.4 Multi-scale fusion output

[0104] The mixed seasonal and trend components are spliced ​​together to obtain the multi-scale fusion feature H multi-scale :

[0105] H multi-scale =Concat(S mixed ,T mixed )

[0106] Among them, Concat is a concatenation operation.

[0107] Step 3.2.2 FMM multi-predictor fusion

[0108] Step 3.2.2.1 Multi-scale prediction

[0109] Features at each scale Independently train a predictor Pred i , output the prediction results at this scale

[0110]

[0111] The predictor formula can be expressed as:

[0112]

[0113] Among them, Linear represents the linear layer operation, and are the learnable weights and biases of the s-th scale predictor.

[0114] Step 3.2.2.2 Weighted Fusion

[0115] Through the learnable weight α i Fusion of multi-scale prediction results to generate the final prediction

[0116]

[0117] Where N is the scale level.

[0118] Step 3.3 Multi-task output head: damage assessment, localization and trend prediction

[0119] Based on multi-scale mixed features, multi-task learning of damage level assessment and future trend prediction is achieved.

[0120] Step 3.3.1 Use Softmax classifier to evaluate damage level

[0121] Output the damage level through the fully connected layer: probability distribution of mild, moderate, and severe:

[0122]

[0123] Among them, z k is the linear layer output of injury registration k, P(Class=k) is the class probability output by the Softmax classifier, and k represents the specific injury level, such as mild, moderate, or severe. and are the weights and biases of the fully connected layer. The loss function is the cross entropy loss L cls :

[0124]

[0125] Among them, p k is the predicted probability distribution, y k is the true label.

[0126] Step 3.3.2 Trend Prediction

[0127] Multi-scale prediction results using the FMM module Generate area growth damage trend for the next L = 10 steps:

[0128]

[0129] in, is the damage area growth prediction for the next L steps, FMM is the feature-level multi-scale hybrid module, and the loss function is the Mahalanobis distance loss L trend :

[0130]

[0131] Among them, y t and denote the true value and predicted value at time step t, respectively.

[0132] Step 3.3.3 Multi-task joint training

[0133] By jointly optimizing the multi-task objectives through weighted summation, the total loss function L total Expressed as:

[0134] L total =λ1L cls +λ2L trend

[0135] Step 4: A real-time monitoring and intelligent early warning system runs a lightweight TimeMixer model on the edge to extract and preliminarily classify damage features in real time. A threshold warning is triggered when the damage level is moderate or higher. The full TimeMixer model is deployed on the cloud to deeply analyze uploaded data and output high-precision damage location and trend predictions.

[0136] Graded early warning strategy:

[0137] Minor damage: marked as "observation level", data recorded and regular review recommended;

[0138] Moderate damage: Triggers a "warning level" warning, notifying maintenance personnel to conduct a local inspection;

[0139] Severe damage: Activate an "emergency level" alarm, automatically shut down the equipment and recommend a repair plan.

[0140] Step 5: System integration: Build a thermoplastic composite health monitoring platform that integrates the damage database, predictive maintenance module, and digital twin interface to support damage data retrieval.

[0141] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent variation based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A thermoplastic composite damage detection method based on the TimeMixer architecture MLP, comprising the following steps, characterized by: Step 1: Deploy a multi-source sensor acquisition system. Install micro-vibration sensors, temperature sensor arrays, acoustic emission sensors, and ultrasonic phase sensors at the composite structure's joints, support structures, geometric mutations, and external load areas to ensure coverage of stress concentration areas and potential damage areas. By integrating a multi-channel data acquisition module, synchronized sampling of data from the four types of sensors is achieved. Step 2: Feature data preprocessing: perform wavelet transform on the micro-vibration signal to extract spectral energy and Higuchi fractal dimension, calculate the local temperature mutation rate for the temperature gradient, extract event count, energy integral and ringing number time domain parameters for the acoustic emission signal, and extract phase difference features for the ultrasonic wave; Step 3: Build a TimeMixer architecture MLP model. The input layer integrates the four sensor data types into a unified tensor format, including vibration, acoustic emission, temperature, and ultrasonic feature data. The multiscale mixing module then uses Past Decomposable Mixing (PDM) to decompose historical data into seasonal and trend components, extracting multiscale damage features through bidirectional information flow. Finally, a multi-task output head outputs the mild, moderate, and severe damage levels of the thermoplastic composite material, as well as the damage area growth over the next 10 steps. Step 4: Real-time monitoring and intelligent early warning system: A lightweight TimeMixer model runs on the edge to extract damage features and perform preliminary classification in real time. A threshold warning is triggered when the damage level is moderate or higher. Deploy the complete TimeMixer model on the cloud, deeply analyze the uploaded data, and output high-precision damage location and trend prediction; Step 5: System integration: Build a thermoplastic composite health monitoring platform that integrates the damage database, predictive maintenance module, and digital twin interface to support damage data retrieval.

2. The thermoplastic composite material damage detection method based on the TimeMixer architecture MLP according to claim 1 is characterized by: The feature data preprocessing in step 2 can be expressed as: Step 2.1: Process the micro-vibration signal to extract the damage characteristics in the high-frequency vibration signal and capture the dynamic response of the micro-cracks inside the material; Step 2.1.1 Wavelet transform to extract spectrum energy The vibration signal x(n) is decomposed into multiple scale sub-bands through three layers using discrete wavelet transform: Among them, W j (k) is the wavelet coefficient at the kth time point after the jth layer wavelet decomposition, ψ j,k (n) is the db4 wavelet basis function, j is the number of decomposition levels, and k is the time index; Sum the squares of the wavelet coefficients of each subband to calculate the spectral energy E j : Among them, E j is the spectral energy of the j-th wavelet subband; Step 2.1.2 Higuchi fractal dimension calculation The fractal dimension D is obtained by constructing a virtual curve and calculating the logarithmic relationship between its length and the scaling factor: Where L(m) is the length of the curve of the reconstructed time series when the scaling factor is m, m is the scaling factor, is the change in the step size of the i-th segment in the reconstructed time series; Step 2.2 Temperature gradient treatment Quantify the sudden change characteristics of the material surface temperature field and identify the abnormal heat conduction area; calculate the ratio of the temperature difference ΔT at adjacent moments to the time step Δt: Among them, R t is the local temperature mutation rate, T t+1 is the temperature value at time t+1, T t is the temperature value at time t, Δt is the sampling interval; Step 2.3 Acoustic emission signal processing Extract the time domain characteristics of acoustic emission events and quantify the energy release law inside the material; Step 2.3.1 Event Counting and Energy Integration Count the number of acoustic emission events per unit time N: N=count(A(t)>10mV) Where A(t) is the amplitude of the acoustic emission signal at time t. The above formula represents the number of events with an amplitude greater than 10mV and the energy E of a single event is calculated as follows: Among them, t1 and t2 are the start and end time of the event; Step 2.3.2 Time domain parameter extraction Duration τ: the time difference between the start and end of the event τ = t2-t1; Ring count C: Detects the number of rings by zero-crossing rate: Among them, sign is the sign function; Step 2.4 Ultrasonic phase difference processing Step 2.4.1 Full Matrix Capture The phased array ultrasonic probe transmits and receives signals to generate full matrix data A ij : A ij (t)=Receive(i)·Transmit(j) Among them, A ij (t) is the echo signal between the i-th receiving element and the j-th transmitting element of the phased array probe, t is time, Receive(i) represents the signal received by the i-th probe element, and Transmit(j) represents the signal transmitted by the j-th probe element; Step 2.4.2 Phase difference calculation Step 2.4.2.1 Fourier transform For each echo signal A ij (t) Perform fast Fourier transform to extract its phase at a specific frequency φ ij =∠(F{A ij (t)}) Among them, φ ij is the phase angle of the echo signal at a specific frequency, ∠ represents the complex phase angle, and F is the Fourier transform operator; Step 2.4.2.2 Phase difference calculation By comparing the phase difference Δφ of echoes from different paths, the abnormality of the internal structure of the material can be reflected: Δφ=φ ij -f ref Among them, φ ref is the phase reference of the defect-free reference sample.

3. The thermoplastic composite material damage detection method based on the TimeMixer architecture MLP according to claim 1 is characterized in that: The TimeMixer architecture MLP model constructed in step 3 can be expressed as follows: This step designs a multi-layer perceptron (MLP) model based on the TimeMixer architecture to implement multi-scale hybrid analysis and multi-task prediction of multi-source sensor data. The model is constructed through three parts: input layer integration, multi-scale hybrid module, and multi-task output head. The specific process is as follows: Step 3.1 Input layer integration: unified tensor format and embedding layer Convert the micro-vibration, temperature, acoustic emission, and ultrasonic sensor feature data extracted in step 2 into a unified tensor format and map them to the dimensions required by the model through an embedding layer; Step 3.1.1 Data format unification Organize the multi-source data features preprocessed in step 2 into a time series tensor τ∈R T×F , where T is the time step, F is the feature dimension, F = 64; Among them, x t,f represents the fth eigenvalue at the tth time step; Step 3.1.2 Embedding layer projection The original features are mapped to the model hidden layer dimension D through the fully connected layer MLP, D = 128: H0=ReLU(W embed ·τ+b embed ) Among them, H0 is the hidden layer feature after the embedding layer projection, W embed , b embed is a learnable parameter, ReLU is the ReLU function; Step 3.1.3 Positional encoding injection Add sine-cosine position encoding to a time series to preserve temporal order information: Among them, PE t,2i is a sinusoidal position code, PE t,2i+1 is the cosine position encoding, t is the time step, i is the feature dimension index, and D is the model hidden layer dimension; The position coding injection H'0 formula is: H'0=H0+PE Step 3.2 Multiscale Hybrid Module Decomposition, mixing and prediction of multi-scale time series are achieved through PDM and FMM; Step 3.2.1 PDM multi-scale decomposition and bidirectional mixing Step 3.2.1.1 Multi-scale downsampling Perform multi-scale pooling on the input sequence H'0 to generate a sequence of N scales Where s represents the scale level, N = 4, the formula is: in, is the pooling sequence of the i-th scale, s is the scale level, k i is the pooling kernel, AvgPool is the average pooling function; Step 3.2.1.2 Seasonality-Trend Decomposition For each scale sequence The sliding average method is used to decompose the seasonal component S of the i-th scale i and trend component T i : Where w is the sliding window size, w=24, and MovingAvg is the sliding average operation; Step 3.2.1.3 Two-way hybrid strategy The seasonal component aggregates seasonal characteristics step by step from fine-grained to coarse-grained: in, is the mixed seasonal feature, i represents the time step index, MLP seasonal For a multi-layer perceptron, the trend component gradually optimizes the trend characteristics from coarse granularity to fine granularity: in, is the trend feature after mixing, MLP trend is a multi-layer perceptron; Step 3.2.1.4 Multi-scale fusion output The mixed seasonal and trend components are spliced ​​together to obtain the multi-scale fusion feature H multi-scale : H multi-scale =Concat(S mixed ,T mixed ) Among them, Concat is a concatenation operation; Step 3.2.2 FMM multi-predictor fusion Step 3.2.2.1 Multi-scale prediction Features at each scale Independently train a predictor Pred i , output the prediction results at this scale The predictor formula can be expressed as: Among them, Linear represents the linear layer operation, and are the learnable weights and biases of the s-th scale predictor; Step 3.2.2.2 Weighted Fusion Through the learnable weight α i Fusion of multi-scale prediction results to generate the final prediction Where N is the scale series; Step 3.3 Multi-task output head: damage assessment, localization and trend prediction Based on multi-scale mixed features, multi-task learning of damage level assessment and future trend prediction is achieved; Step 3.3.1 Use Softmax classifier to evaluate damage level Output the damage level through the fully connected layer: probability distribution of mild, moderate, and severe: Among them, z k is the linear layer output of injury registration k, P(Class=k) is the class probability output by the Softmax classifier, and k represents the specific injury level, such as mild, moderate, or severe. and is the weight and bias of the fully connected layer; the loss function is the cross entropy loss L cls : Among them, p k is the predicted probability distribution, y k is the true label; Step 3.3.2 Trend Prediction Multi-scale prediction results using the FMM module Generate area growth damage trend for the next L = 10 steps: in, is the damage area growth prediction for the next L steps, FMM is the feature-level multi-scale hybrid module, and the loss function is the Mahalanobis distance loss L trend : Among them, y t and Represent the true value and predicted value of time step t respectively; Step 3.3.3 Multi-task joint training By jointly optimizing the multi-task objectives through weighted summation, the total loss function L total Expressed as: L total =λ1L cls +λ2L trend 。

Images