Deep learning-based method for predicting low-temperature brittleness of timber members

Abstract

The application relates to a deep learning-based low-temperature brittleness prediction method for wood components. The method comprises the following steps: acquiring temperature and humidity time series data of an environment in which a target wood component is located and performing standardization preprocessing to generate a standardized environment data sequence; based on the standardized environment data sequence, a temperature and humidity coupling derived index is calculated, and a multi-dimensional feature sequence is generated in combination with original time series statistical features of the standardized environment data sequence; the multi-dimensional feature sequence is intercepted according to a first preset sliding time window to generate a plurality of feature matrices, and each feature matrix is assigned a label to generate a deep space-time sequence sample; the deep space-time sequence sample is input into a trained deep space-time prediction model to generate a continuous risk probability curve of the target wood component occurring brittle fracture in a preset future time period. The method can deeply integrate time series environment data, scientifically quantify the temperature and humidity coupling effect, and significantly improve the timeliness and accuracy of risk early warning.

Description

Technical Field

[0001] This invention belongs to the field of low-temperature brittleness prediction technology for wooden components, and in particular relates to a method for predicting low-temperature brittleness of wooden components based on deep learning. Background Technology

[0002] With the continuous development of the fields of timber-framed architecture and historical heritage preservation, the need for monitoring the long-term performance and safety status of critical load-bearing timber components is becoming increasingly prominent. In complex natural environments, especially those with alternating or coexisting low temperatures and high humidity, the mechanical properties of timber components undergo significant changes, exhibiting reduced toughness and increased brittleness, posing a risk of sudden fracture. To address this challenge, current methods primarily rely on regular on-site inspections combined with environmental monitoring using temperature and humidity sensors, aiming to assess the safety status of components through observation and measurement.

[0003] In traditional techniques, assessing the risk of low-temperature brittleness in wooden components largely relies on the experience of technicians. The common practice is to set fixed temperature or relative humidity thresholds; when sensor readings exceed these thresholds, an alarm is triggered or simple protective measures are initiated. Additionally, attempts are made to utilize classical mechanical models or statistical empirical formulas based on limited historical data to approximate the impact of specific environmental conditions on wood properties, aiming to achieve a preliminary risk assessment.

[0004] However, these current methods or traditional approaches have significant limitations: empirical judgment and fixed threshold methods cannot quantify the complex synergistic coupling effect between the two key environmental factors of temperature and humidity, and research shows that this coupling effect greatly accelerates the embrittlement process of wood; manual inspection and simple threshold alarms have significant lag, making it impossible to predict risks in advance and providing sufficient early warning time before brittle fracture occurs; wood embrittlement is a nonlinear process affected by the cumulative effect of historical loads, and existing methods are unable to automatically learn and extract the deep-seated patterns and key characteristics that lead to the evolution of embrittlement risk from long-term, sequential monitoring data, resulting in insufficient prediction accuracy. Summary of the Invention

[0005] Therefore, it is necessary to provide a method that can deeply integrate time-series environmental data, scientifically quantify the temperature and humidity coupling effect, and achieve advanced and accurate prediction of the low-temperature brittleness risk of wooden components to address the above-mentioned technical problems.

[0006] Firstly, this application provides a deep learning-based method for predicting the low-temperature brittleness of wooden components, including:

[0007] Acquire the time-series temperature and humidity data of the environment in which the target wooden component is located, perform standardized preprocessing on the temperature and humidity time-series data, and generate a standardized environmental data sequence; the temperature and humidity time-series data includes historical temperature and humidity time-series data and real-time monitoring temperature and humidity time-series data;

[0008] Based on standardized environmental data sequences, temperature and humidity coupled derivative indices are calculated, and multidimensional feature sequences are generated based on the original time-series statistical characteristics of the temperature and humidity coupled derivative indices and standardized environmental data sequences; among them, multidimensional feature sequences are used to characterize the joint driving force of the environment on wood embrittlement.

[0009] The multidimensional feature sequence is extracted according to the first preset sliding time window, multiple feature matrices are generated, and a label is assigned to each feature matrix to obtain a labeled feature matrix; where the label is the fragile risk state corresponding to the future time.

[0010] Arrange all labeled feature matrices in chronological order to generate deep spatiotemporal sequence samples;

[0011] The deep spatiotemporal sequence samples are input into the trained deep spatiotemporal prediction model to generate a continuous risk probability curve of brittle fracture of the target wooden component within a preset future time period; the encoder of the trained deep spatiotemporal prediction model includes an embedding layer, an encoder, and a decoder.

[0012] Furthermore, based on the standardized environmental data sequence, a temperature-humidity coupled derivative index was calculated, and based on the original time-series statistical characteristics of the temperature-humidity coupled derivative index and the standardized environmental data sequence, a multidimensional feature sequence was generated. This multidimensional feature sequence is used to characterize the combined driving force of the environment on wood embrittlement, including:

[0013] Based on the temperature and relative humidity values ​​in the standardized environmental data sequence, the absolute humidity estimate for each time point is calculated, and all absolute humidity estimates are arranged in chronological order to generate the first time-series feature sequence.

[0014] The temperature and relative humidity values ​​are multiplied point by point to generate a temperature and humidity interaction term for each time point. The temperature and humidity interaction term for the current time point is combined with the temperature and humidity interaction term for the preset historical lag time to obtain the temperature and humidity interaction feature group for the current time point.

[0015] Arrange all temperature and humidity interaction feature groups in chronological order to generate a second time-series feature sequence;

[0016] Based on temperature and relative humidity values, the embrittlement driving force index at each time point was calculated; the embrittlement driving force index is used to reflect the degree of synergistic stress of low temperature and high humidity.

[0017] Arrange all embrittlement driving force indices in chronological order to generate an embrittlement driving force index sequence;

[0018] Within the second preset sliding time window, the statistical characteristics of the standardized environmental data sequence, the first time-series feature sequence, the second time-series feature sequence, and the embrittlement driving force index sequence are calculated respectively; wherein, the statistical characteristics include the mean, standard deviation, maximum value, and minimum value;

[0019] The statistical features within the second preset sliding time window are combined to generate a high-dimensional feature vector corresponding to the second preset sliding time window.

[0020] Arrange all high-dimensional feature vectors in chronological order to generate a multi-dimensional feature sequence.

[0021] Furthermore, based on temperature and relative humidity values, the embrittlement driving force index was calculated at each time point; wherein, the embrittlement driving force index is used to reflect the degree of synergistic stress of low temperature and high humidity, including:

[0022] Based on temperature and relative humidity values, the embrittlement driving force index at each time point is calculated using the following formula:

[0023]

[0024] in, The embrittlement driving force index, The preset reference temperature for the onset of embrittlement effect. This is an adjustable scaling factor. This is the relative humidity value. This is the temperature value.

[0025] Furthermore, the deep spatiotemporal sequence samples are input into the trained deep spatiotemporal prediction model to generate a continuous risk probability curve for brittle fracture of the target wooden component within a preset future time period; wherein, the encoder of the trained deep spatiotemporal prediction model includes an embedding layer, an encoder, and a decoder, including:

[0026] The feature matrix from the deep spatiotemporal sequence sample is input into the embedding layer of the trained deep spatiotemporal prediction model to generate a high-dimensional feature embedding vector, which is arranged continuously according to the time step.

[0027] The high-dimensional feature embedding vector is input into the encoder of the trained deep spatiotemporal prediction model to generate the hidden state vector at each time step, and the hidden state vector at the last time step is determined as the latent variable; where the encoder is constructed based on a long short-term memory network; the latent variable is used to characterize the historical embrittlement state and the current embrittlement state of the wood.

[0028] Based on latent variables, the decoder of a trained deep spatiotemporal prediction model is used to extrapolate future brittleness risks and generate a continuous risk probability curve for the brittle fracture of the target wooden component within a preset future time period.

[0029] Furthermore, based on latent variables, the decoder of the trained deep spatiotemporal prediction model is used to extrapolate future brittleness risks, generating a continuous risk probability curve for the target wooden component to undergo brittle fracture within a preset future time period, including:

[0030] Based on latent variables, the hidden states and cell states of the decoder of the trained deep spatiotemporal prediction model are initialized; the decoder is constructed based on a long short-term memory network.

[0031] Based on the initialized hidden state and cell state, at each prediction time step of the decoder, the attention weights between each element in the current decoder hidden state and the hidden state sequence of all time steps of the encoder are calculated.

[0032] Based on the attention weights, the elements in the hidden state sequence are summed in a weighted manner to generate the current prediction time step context vector of the decoder.

[0033] Based on the context vector of the current prediction time step and the current decoder hidden state, the fragility risk probability of the current prediction time step is calculated.

[0034] Based on the current predicted time step fragility risk probability, repeat the steps from calculating attention weights to calculating the current predicted time step fragility risk probability, and sequentially generate the fragility risk probability for each time step within the preset future time period.

[0035] Connect all the fragile risk probabilities to generate a continuous risk probability curve.

[0036] Secondly, this application also provides a deep learning-based device for predicting the low-temperature brittleness of wooden components, including:

[0037] The data acquisition module is used to acquire the temperature and humidity time series data of the environment in which the target wooden component is located, and to perform standardized preprocessing on the temperature and humidity time series data to generate a standardized environmental data sequence; among which, the temperature and humidity time series data includes historical temperature and humidity time series data and real-time monitoring temperature and humidity time series data;

[0038] The feature calculation module is used to calculate the temperature and humidity coupled derivative index based on the standardized environmental data sequence, and generate a multidimensional feature sequence based on the original time-series statistical features of the temperature and humidity coupled derivative index and the standardized environmental data sequence; the multidimensional feature sequence is used to characterize the joint driving force of the environment on wood embrittlement.

[0039] The matrix generation module is used to extract multi-dimensional feature sequences according to a first preset sliding time window, generate multiple feature matrices, and assign labels to each feature matrix to obtain labeled feature matrices; wherein, the label is the fragile risk state corresponding to the future time.

[0040] The sample generation module is used to arrange all labeled feature matrices in chronological order to generate deep spatiotemporal sequence samples.

[0041] The prediction curve generation module is used to input deep spatiotemporal sequence samples into a trained deep spatiotemporal prediction model to generate a continuous risk probability curve of brittle fracture of the target wooden component within a preset future time period; wherein, the encoder of the trained deep spatiotemporal prediction model includes an embedding layer, an encoder, and a decoder.

[0042] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores at least one instruction, at least one program, a code set, or an instruction set, and the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement any of the deep learning-based low-temperature brittleness prediction methods for wooden components described in the embodiments of this application.

[0043] Fourthly, this application also provides a computer-readable storage medium storing at least one piece of program code, which is loaded and executed by a processor to implement the personalized learning path generation method based on learning behavior profiles as described in any of the embodiments of this application.

[0044] The aforementioned deep learning-based method for predicting the low-temperature brittleness of wooden components collects multi-source heterogeneous data and performs standardization processing. It then calculates coupled features such as absolute humidity, temperature-humidity interaction term, and embrittlement driving force index. A sliding window is used to construct labeled spatiotemporal sequence samples, combined with a time-ordered dataset partitioning strategy. Through a trained deep spatiotemporal prediction model, it deduces future continuous risk probability curves. This method deeply integrates temporal environmental data and scientifically quantifies the temperature-humidity coupling effect, achieving a leap from passively responding to threshold alarms to proactively predicting risk trends, significantly improving the timeliness and accuracy of risk warnings. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1This is a flowchart illustrating a deep learning-based method for predicting the low-temperature brittleness of wooden components in one embodiment.

[0047] Figure 2 This is a flowchart illustrating the steps of calculating temperature and humidity coupled derivative indices based on standardized environmental data sequences in one embodiment, and generating multidimensional feature sequences based on the original time-series statistical characteristics of the temperature and humidity coupled derivative indices and standardized environmental data sequences.

[0048] Figure 3 This is a schematic diagram of a deep learning-based low-temperature brittleness prediction device for wooden components in one embodiment. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0050] In one embodiment, a deep learning-based method for predicting the low-temperature brittleness of wooden components is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. Figure 1 As shown, in this embodiment, the method includes the following steps:

[0051] Step S101: Obtain the temperature and humidity time series data of the environment where the target wooden component is located, perform standardized preprocessing on the temperature and humidity time series data, and generate a standardized environmental data sequence; wherein, the temperature and humidity time series data includes historical temperature and humidity time series data and real-time monitoring temperature and humidity time series data.

[0052] Among them, target wooden components refer to wooden components that are clearly designated as operation objects, research objects, or control objects in specific scenarios (such as structural inspection); standardized preprocessing is a key data regularization process, and its primary operations are data cleaning and data alignment.

[0053] For example, historical time-series data is obtained from the structural health monitoring database of the target wooden component. This data is recorded by high-precision digital temperature and humidity sensors deployed near key parts of the component, and its data fields include at least a timestamp, Celsius temperature value, and percentage relative humidity value. Simultaneously, monitoring data streams uploaded in real time from sensor nodes at the same or equivalent locations at a preset frequency (e.g., every 10 minutes) are received. The acquired raw temperature and humidity time-series data is filtered according to a preset physical reasonableness range to remove outliers caused by momentary sensor failures or transmission interference. The temperature and humidity data streams from different sensors are time-aligned based on the timestamp synchronized by a high-precision network time protocol, ensuring that each moment has a corresponding temperature and humidity pairing value. For short-term data gaps caused by brief sensor failures or communication packet loss, a combination of time-series-based seasonal decomposition and linear interpolation can be used to fill in the gaps and maintain the continuity of the sequence. All processed data is resampled at uniform time intervals (e.g., hourly) to obtain a continuous, complete, and time-aligned standardized environmental data sequence.

[0054] Among them, the structural health monitoring database is the core data management system supporting structural health monitoring work. It is mainly used to integrate, store, manage, and analyze various monitoring data collected throughout the entire life cycle of the engineering structure, providing data support for structural condition assessment, damage early warning, and operation and maintenance decisions. A high-precision digital temperature and humidity sensor is an electronic component that can accurately collect ambient temperature and humidity data and directly output digital signals. The preset physical rationality range refers to the boundary interval that conforms to physical logic, pre-defined for the parameters, behaviors, or phenomena of the research object based on known physical laws, objective constraints, or practical needs. For example, the upper and lower bounds of temperature and humidity defined based on wood survival climatology. A sensor is a device that can convert non-electrical physical or chemical quantities (such as temperature and humidity) into measurable electrical signals (such as voltage and current). Seasonal decomposition and linear interpolation based on time series are also included. The combined approach refers to first breaking down the inherent patterns of a time series, and then addressing the data missing issues in a targeted manner. This approach is suitable for time series data with seasonal fluctuations and a small number of missing values ​​(such as quarterly temperatures). Time alignment is an operation that uses technical or manual means to ensure that the time dimensions of multiple independent time-related objects (such as data and events) are consistent or matched. It is used to eliminate time deviations and ensure that information from different objects at the same time point can correspond and be correlated. Interpolation imputation refers to constructing a continuous (or approximately continuous) distribution pattern of the variable based on the known data points of the variable containing the missing value through a mathematical model, and then calculating a reasonable value for the missing position to achieve data integrity restoration. Resampling refers to adjusting the sampling frequency or data density of data (or signals) through a specific algorithm. Preset frequency refers to a fixed frequency value or frequency range that is pre-set and has a specific function or purpose.

[0055] Step S102: Based on the standardized environmental data sequence, the temperature and humidity coupled derivative index is calculated, and based on the original time-series statistical characteristics of the temperature and humidity coupled derivative index and the standardized environmental data sequence, a multidimensional feature sequence is generated; wherein, the multidimensional feature sequence is used to characterize the joint driving force of the environment on wood embrittlement.

[0056] Among them, the temperature and humidity coupled derivative index is a derivative parameter that is calculated based on two basic environmental parameters, temperature and humidity, through specific logic, formulas or models, and can more accurately reflect the comprehensive environmental effects.

[0057] For example, based on the temperature and relative humidity at each time point in the standardized environmental data sequence, key environmental coupling-derived indicators such as absolute humidity, temperature-humidity interaction term, and embrittlement driving force index are calculated for each time point, resulting in absolute humidity sequence, temperature-humidity interaction term sequence, and embrittlement driving force index sequence. Within multiple sliding windows of preset time lengths, a series of statistical characteristics, including but not limited to mean, standard deviation, maximum value, minimum value, skewness, and kurtosis, are calculated for the original temperature and relative humidity sequences and the calculated absolute humidity sequence, temperature-humidity interaction term sequence, and embrittlement driving force index sequence. All basic observation values, coupled-derived indicator values, and their statistical characteristic values ​​under different time windows corresponding to each time point are concatenated according to feature dimensions to generate a high-dimensional multidimensional feature sequence. Each row vector of this sequence is used to characterize the joint driving force of the current moment and recent historical environment on wood embrittlement. Here, the preset time length refers to the size of the pre-set sliding window, such as 10 seconds or 5 minutes.

[0058] Step S103: Extract the multidimensional feature sequence according to the first preset sliding time window, generate multiple feature matrices, and assign a label to each feature matrix to obtain a labeled feature matrix; wherein, the label is the fragile risk state corresponding to the future time.

[0059] The first preset sliding time window is a core technical concept in scenarios such as time series data processing and stream computing. Essentially, it is to predefine a data observation box with a fixed time length and let this box slide continuously along the time axis according to preset rules (such as a fixed step size), thereby realizing the segmentation, filtering and analysis of dynamic data. The term "first" is only used to distinguish similar objects and does not need to be used to describe a specific order or sequence, nor should it be understood as indicating or implying relative importance.

[0060] For example, two key time parameters are predefined: historical observation duration. (e.g., 168 hours, or one week) and future forecast duration (For example, 24 hours). Starting from the beginning of the multidimensional feature sequence, slide a length of [length missing] with a fixed step size (e.g., 1 hour). The time window is defined. For each window, the feature vectors of all time steps within the window are extracted and stacked to form a two-dimensional feature matrix with dimensions of . ,in, This represents the total number of features. For example, a label is assigned to this feature matrix based on the historical maintenance and event log database of the target wooden component: by querying at the end of the current window. The future Within the past hour, has there been any documented confirmed brittle fracture or serious damage event recorded for the target wooden component? If so, the label of this sample can be... A "1" label represents a high-risk positive sample; if none is found, a "0" label represents a safe negative sample. For cases with only minor damage reports, an intermediate label value (e.g., 0.5) can be introduced to characterize medium risk. By traversing the entire feature sequence, a large number of such labeled feature matrix sample pairs are generated. Each of them All are associated with an indicator of their future. Hourly risk status label .

[0061] Step S104: Arrange all labeled feature matrices in chronological order to generate deep spatiotemporal sequence samples.

[0062] For example, all labeled feature matrix samples are strictly ordered chronologically according to the end timestamps of their corresponding time windows to ensure the temporal causality of the dataset. This can be achieved using a proportional partitioning method based on chronological order, dividing the sorted sample sequence into three non-overlapping datasets: a training set, a validation set, and a test set. After partitioning, the samples (feature matrices and labels) of the training and validation sets are organized into data loading formats that conform to the requirements of deep learning frameworks, such as constructing iterable Dataset objects. The final result is a structured, time-partitioned deep spatiotemporal sequence sample set, where each sample contains a feature matrix representing a past environmental sequence and a label representing future risk.

[0063] Among them, the time-sequential proportional division method is a method of dividing or allocating target objects (such as data) into stages based on the time dimension and according to a specific ratio. Combining the time progression rhythm and preset proportional rules, it ensures that the division of each stage conforms to the linear law of time and meets the quantitative allocation requirements. For example, the samples of the first 70% of the time period are used as the training set, the middle 15% as the validation set, and the last 15% as the test set. The core requirements for data loading in deep learning frameworks are structured, batch-compatible, iterable, and compatible with tensor formats. The core goal is to efficiently feed data to the model for training / inference. The iterable Dataset object is a core component that encapsulates a data set and supports reading data one by one / in batch. Its feature is that it implements an iterable protocol and can obtain data samples one by one through loops or framework interfaces.

[0064] Step S105: Input the deep spatiotemporal sequence samples into the trained deep spatiotemporal prediction model to generate a continuous risk probability curve of the target wooden component undergoing brittle fracture within a preset future time period; wherein, the encoder of the trained deep spatiotemporal prediction model includes an embedding layer, an encoder, and a decoder.

[0065] The trained deep spatiotemporal prediction model is obtained through the following method: acquiring historical monitoring data, i.e., time-series temperature and humidity data with timestamps, and processing it according to the exact same process as steps S101 (standardization), S102 (multidimensional feature construction), and S103 (sliding window extraction) to generate a historical feature matrix. Using the future prediction time step corresponding to the historical feature matrix as a benchmark, matching historical brittle fracture event records, and constructing a training sample set. If brittle fracture occurs within the future time step, a binary label is assigned. (Positive sample); if no breakage occurs, label with a binary tag. (Negative samples); Based on the preset encoder-decoder architecture parameters, an initial deep spatiotemporal prediction model of the encoder-decoder architecture is constructed. The encoder front end has an embedded layer for linear transformation and nonlinear activation of the input matrix. The training sample set is input into the initial model for forward propagation to obtain the risk probability value predicted by the model. The weighted cross-entropy loss function is used to calculate the weighted cross-entropy loss value between the risk probability value predicted by the model and the true binary risk label, which serves as the main loss function for model training. The formula for calculating the weighted cross-entropy loss function is:

[0066]

[0067] The weighted cross-entropy loss function is... For the sample size, and These are the preset weight coefficients for the minority class positive samples and the majority class negative samples, respectively. ; The risk probability value predicted by the model. To represent the true binary risk label, the sum of squared differences between the predicted values ​​of the future continuous risk probability curve at adjacent time points is calculated and used as the loss term for time series smoothness regularization. The weighted cross-entropy loss value With time series smoothness regularization loss The total loss is obtained by weighted summation. ,in The smoothing coefficient is preset; the total loss is minimized using the backpropagation algorithm and gradient descent optimizer. The parameters of the initial deep spatiotemporal prediction model are updated until the model's performance on the validation set converges, resulting in a pre-trained deep spatiotemporal prediction model. The encoder-decoder architecture is the core framework for sequence transformation tasks in deep learning, essentially a two-step logic of first understanding the input and then generating the output. Forward propagation is the core process for neural networks to calculate output results in deep learning; it essentially starts from the input data, passes information layer by layer, and calculates to finally obtain the prediction result. The backpropagation algorithm (BP algorithm) is the core algorithm for deep learning and neural network training, used to calculate results by working backwards. The algorithm calculates the error and iteratively adjusts the network parameters (weights and biases) to gradually bring the neural network's predictions closer to the true labels, ultimately achieving model learning. The gradient descent optimizer is the most basic and core parameter optimization algorithm in machine learning (especially deep learning). Its core objective is to minimize the model's loss function, that is, to minimize the error between the model's predicted value and the true value. Essentially, it achieves optimization by gradually adjusting the model parameters along the downhill path of the loss function. The preset encoder-decoder architecture parameters refer to the hyperparameters and structural configurations that are pre-set and not learned autonomously by the model. They are used to define the architecture framework and constrain the training process, directly affecting model performance and efficiency, such as the hidden layer dimension and the embedding layer output dimension.

[0068] For example, deep spatiotemporal sequence samples are input into a trained deep spatiotemporal prediction model, and processed by an encoder and a decoder to generate a continuous risk probability curve of brittle fracture of the target wooden component within a preset future time period.

[0069] In this embodiment, multi-source heterogeneous data is collected and standardized, and then coupled features such as absolute humidity, temperature-humidity interaction term, and embrittlement driving force index are calculated. A sliding window is used to construct labeled spatiotemporal sequence samples, combined with a time-ordered dataset partitioning strategy. A trained deep spatiotemporal prediction model is then used to deduce future continuous risk probability curves. This achieves a leap from passively responding to threshold alarms to proactively predicting risk trends, significantly improving the timeliness and accuracy of risk warnings.

[0070] In one embodiment, such as Figure 2 As shown, based on standardized environmental data sequences, temperature and humidity coupled derivative indices are calculated, and based on the original time-series statistical characteristics of the temperature and humidity coupled derivative indices and standardized environmental data sequences, multidimensional feature sequences are generated. These multidimensional feature sequences characterize the combined driving force of the environment on wood embrittlement, including:

[0071] Step S201: Based on the temperature and relative humidity values ​​in the standardized environmental data sequence, calculate the estimated absolute humidity value at each time point, and arrange all the estimated absolute humidity values ​​in chronological order to generate the first time-series feature sequence.

[0072] For example, a standardized environmental data sequence is received, which includes temperature values ​​strictly aligned to timestamps. and relative humidity value For each time point in the sequence Calculate its estimated absolute humidity value Its physical meaning is the mass of water vapor contained in a unit volume of air. Based on temperature... Calculate the saturated vapor pressure (unit: degrees Celsius). (Unit: hectopascals), the calculation formula is:

[0073]

[0074] Combined with relative humidity (Percentage) to obtain the actual water vapor pressure The calculation formula is:

[0075]

[0076] Then convert it to absolute humidity. The conversion formula is the ideal gas law:

[0077]

[0078] The unit is usually grams per cubic meter. Perform the above calculation for each time point of the input sequence and obtain all the estimated absolute humidity values. Arrange them according to their original time sequence to generate a new time series of the same length as the original temperature and humidity series, and define this series as the first time series feature sequence.

[0079] Step S202: Perform point-by-point multiplication of temperature and relative humidity values ​​to generate temperature and humidity interaction terms for each time point, and combine the temperature and humidity interaction terms of the current time point with the temperature and humidity interaction terms of the preset historical lag time to obtain the temperature and humidity interaction feature group of the current time point.

[0080] For example, temperature values ​​are read from a standardized environmental data sequence. and relative humidity value For each identical point in time... Perform point-by-point multiplication, that is, calculate Generate the original values ​​of the temperature and humidity interaction term at that point in time. A pre-defined set of historical lag times is used, for example... These represent times taken 6 hours, 12 hours, and 24 hours ago, respectively. The current time point is extracted from the calculated time series of temperature and humidity interaction terms. And the values ​​of the interactive items corresponding to these preset lag times. These belong to different times but revolve around the same current time. The interaction terms are combined into feature groups, for example, they can be represented as... This feature set is the temperature and humidity interaction feature set at the current time point.

[0081] Step S203: Arrange all temperature and humidity interaction feature groups in chronological order to generate a second time-series feature sequence.

[0082] For example, according to time points Starting from morning to night, all generated temperature and humidity interaction feature groups are traversed. Since each feature group is essentially a vector with a dimension equal to 1 plus the preset number of historical lag times, these feature group vectors are treated as new, composite "observations." These vectors are then arranged sequentially according to the indices of the original time series, thus forming a new, multi-dimensional time series data structure. This newly generated sequence is the second time series feature sequence.

[0083] Step S204: Based on the temperature and relative humidity values, calculate the embrittlement driving force index at each time point; wherein, the embrittlement driving force index is used to reflect the degree of synergistic stress of low temperature and high humidity.

[0084] The design principle of the embrittlement driving force index is that the low-temperature embrittlement of wood is significantly aggravated when the temperature is below a certain critical point and a certain amount of humidity is present.

[0085] For example, based on temperature values ​​in a standardized environmental data sequence and relative humidity value Calculate each time point Embrittlement driving force index The formula for calculating this index is as follows:

[0086]

[0087] in, The embrittlement driving force index, The preset reference temperature for the onset of embrittlement effect. This is an adjustable scaling factor. This is the relative humidity value. This is the temperature value.

[0088] Step S205: Arrange all embrittlement driving force indices in chronological order to generate an embrittlement driving force index sequence.

[0089] For example, a brittleness driving force index sequence is generated by arranging all the brittleness driving force indices in their original chronological order.

[0090] Step S206: Within the second preset sliding time window, calculate the statistical characteristics of the standardized environmental data sequence, the first time-series feature sequence, the second time-series feature sequence, and the embrittlement driving force index sequence, respectively; wherein, the statistical characteristics include the mean, standard deviation, maximum value, and minimum value.

[0091] Here, the second preset sliding time window refers to one or more pre-set second preset sliding time window lengths, for example... Hour, The hourly values ​​are used to capture short-term and medium-term impact patterns, respectively; the mean reflects the average level within the window period; the standard deviation reflects the volatility or dispersion within the window period; the maximum value reflects the extreme upper limit within the window period; and the minimum value reflects the extreme lower limit within the window period.

[0092] For example, consider the original temperature sequence and original relative humidity sequence, the first time-series feature sequence (absolute humidity sequence), the second time-series feature sequence (temperature-humidity interaction vector sequence), and the embrittlement driving force index sequence in a standardized environmental data sequence. For the current time point to be processed... ,by Backtracking backwards from the end point of the window. For each hour (e.g., 24 hours), all data points of the corresponding sequence are extracted within that window. A set of predefined statistical characteristics are calculated on this window of data, including at least the mean, standard deviation, maximum, and minimum values. For example, for the original temperature sequence, the average temperature over the past 24 hours is calculated. Temperature standard deviation Maximum temperature and lowest temperature For each preset window length For each of the aforementioned input sequences, this truncation and computation process is performed independently and repeatedly, thus for each time point... Generate a large set of statistical features, which are used to characterize the cutoff time from different perspectives and time scales. The environmental conditions and historical evolution.

[0093] Step S207: Combine the statistical features within the second preset sliding time window to generate a high-dimensional feature vector corresponding to the second preset sliding time window.

[0094] For example, regarding the current time point and a specific second preset sliding time window Multiple sets of statistical characteristics (mean, standard deviation, etc.) from different source sequences have been calculated from different source sequences (original temperature and humidity, absolute humidity, components of the temperature-humidity interaction vector, and embrittlement driving force index). Those belonging to the same time point... and the same sliding window All statistical characteristic values ​​are concatenated in a preset order. For example, the temperature series is arranged first. Next, we have the four statistics for the relative humidity sequence, followed by the four statistics for the absolute humidity sequence. Then, we have the mean, standard deviation, etc., of each component of the temperature-humidity interaction vector within the current window (statistics need to be calculated separately for each dimension of the vector). Finally, we have the four statistics for the embrittlement driving force index. After concatenating all the scalar values, we obtain a one-dimensional, high-dimensional feature vector. This vector is the second preset sliding time window. The corresponding high-dimensional feature vector.

[0095] Step S208: Arrange all high-dimensional feature vectors in chronological order to generate a multi-dimensional feature sequence.

[0096] For example, for each time point It is possible to generate multiple high-dimensional feature vectors if multiple sliding windows of different lengths are defined. Each window will generate a vector. These feature vectors, organized by time points, are then integrated into a continuous temporal sequence. First, the final composition of the feature vectors is determined: for example, if two window lengths are used... and , each time point The feature vectors of the two corresponding windows are then concatenated to generate the final representative feature vector at that time point. According to time index Extract the final feature vector corresponding to each time point in sequence from start to finish. These features are then processed and arranged into a sequence. This sequence is the final generated multidimensional feature sequence. Each time step of this sequence is a high-dimensional feature vector. The dimension of the vector is determined by all the original features, coupled derived features, sliding window length, and types of statistical features selected.

[0097] In this embodiment, based on temperature and relative humidity values ​​from a standardized environmental data sequence, an absolute humidity estimation sequence, a temperature-humidity interaction term sequence, and a brittleness driving force index sequence are calculated for each time point. Within a second preset sliding time window, the statistical characteristics of these sequences are calculated and combined to generate a high-dimensional feature vector corresponding to the second preset sliding time window. All high-dimensional feature vectors are arranged in chronological order to generate a multi-dimensional feature sequence. This effectively overcomes the limitations of traditional methods that rely solely on raw temperature or humidity thresholds, and constructs a data-driven feature representation that profoundly reflects how complex environmental factors jointly drive the wood brittleness process from both physical mechanisms and statistical perspectives.

[0098] In one embodiment, the embrittlement driving force index is calculated based on temperature and relative humidity values ​​at each time point; wherein the embrittlement driving force index is used to reflect the degree of synergistic stress of low temperature and high humidity, including:

[0099] Step S301: Based on the temperature and relative humidity values, the embrittlement driving force index at each time point is calculated using the following formula:

[0100]

[0101] in, The embrittlement driving force index, The preset reference temperature for the onset of embrittlement effect. This is an adjustable scaling factor. This is the relative humidity value. This is the temperature value.

[0102] in, It is a preset reference temperature for the onset of embrittlement effect. This parameter is not fixed, but can be set and calibrated based on the tree species, age and historical performance data of the target wood component through expert knowledge or preliminary data analysis. Its physical meaning is to identify a temperature threshold. When the ambient temperature is lower than this threshold, the contribution of low temperature to wood embrittlement is considered to become significant. The function is used to ensure that the result is non-negative, which means that only when... This term only takes a positive value when the temperature is low, and the lower the temperature, the lower the value. The larger the difference, the greater the "low-temperature stress intensity" it represents; if If the value is zero, it means that the embrittlement driving force caused by low temperature is not considered at the current temperature. This is used to convert the relative humidity percentage value into a coefficient between 0 and 1, which is used to characterize the intensity of humidity stress; This is an adjustable scaling factor used to balance the magnitude of the index across the overall characteristics.

[0103] For example, temperature values ​​based on standardized environmental data sequences and relative humidity value Execution of embrittlement driving force index The calculation is as follows:

[0104]

[0105] For each time point in the standardized environmental data sequence The above calculation process is executed independently and sequentially, ultimately outputting a new time series of the same length as the input sequence, namely the embrittlement driving force index time series. Each scalar value in the sequence This is used to intuitively and quantitatively reflect the magnitude of the theoretical combined driving force of environmental conditions at a given moment on inducing brittle fracture in wood.

[0106] In this embodiment, the embrittlement driving force index is calculated at each time point using the temperature and relative humidity values ​​of the standardized environmental data sequence. This allows for the quantification of the nonlinear synergistic effect of the two environmental stress factors, low temperature and high humidity, in a product form. The complex material environmental response mechanism is condensed into a calculable scalar index, greatly enhancing the interpretability of the features.

[0107] In one embodiment, a deep spatiotemporal sequence sample is input into a trained deep spatiotemporal prediction model to generate a continuous risk probability curve of brittle fracture of the target wooden component within a preset future time period; wherein, the encoder of the trained deep spatiotemporal prediction model includes an embedding layer, an encoder, and a decoder, comprising:

[0108] Step S401: Input the feature matrix in the deep spatiotemporal sequence sample into the embedding layer of the trained deep spatiotemporal prediction model to generate a high-dimensional feature embedding vector, which is arranged continuously according to the time step.

[0109] The embedding layer is essentially one or more fully connected neural networks located at the model input. Its technical principle lies in the fact that the original high-dimensional feature vector may contain redundant information, and different features may have differences in dimensions and numerical ranges. Directly inputting these features into a complex temporal network may affect learning efficiency and stability. This embedding layer, through its learnable weight parameters, optimizes the input feature vector (with dimensions of 1000-12000) at each time step. Perform a nonlinear transformation and dimensionality reduction mapping to project it into a more representative and dimensionally optimized continuous vector space, namely the latent space; the learnable weight parameters refer to a key tool in the embedding layer, namely the weight matrix. Learnable means that this weight matrix is ​​not fixed and will be continuously adjusted during model training.

[0110] For example, the feature matrix from deep spatiotemporal sequence samples is input into the embedding layer of a pre-trained deep spatiotemporal prediction model. Through its learnable weight parameters, a nonlinear transformation and dimensionality reduction mapping is performed on the feature vector at each time step of the input, projecting it into a more representative and dimensionally optimized continuous vector space, i.e., the latent space. This mapping process can be described as follows: ,in, It is the first The original feature vectors at each time step, and These are the weight matrix and bias vector of the embedding layer. It is a non-linear activation function. It is the output high-dimensional feature embedding vector. For the input feature matrix... This operation is performed step by step in each time step, and the final output is a result of... A new sequence is formed by arranging high-dimensional feature embedding vectors in their original time order. This sequence represents the original features in the abstract space learned by the model. Nonlinear transformation is a data transformation that breaks linear relationships. The essence of data transformation is to re-represent data using new rules. The core of nonlinear transformation is that the transformation rules do not satisfy superposition and homogeneity (the two core properties of linear transformation), meaning that the input and output of data are not a simple "proportion + superposition" relationship. Dimensionality reduction is the compression from a high-dimensional space to a low-dimensional space. Dimensionality can be understood as the number of features in the data. For example, describing a person using "height, weight, age, and blood pressure" is 4-dimensional data. The core of dimensionality reduction is to reduce the number of features (reduce dimensionality) while preserving core data information (such as distance).

[0111] Step S402: Input the high-dimensional feature embedding vector into the encoder of the trained deep spatiotemporal prediction model to generate the hidden state vector at each time step, and determine the hidden state vector at the last time step as the latent variable; wherein, the encoder is constructed based on a long short-term memory network; the latent variable is used to characterize the historical embrittlement state and the current embrittlement state of the wood.

[0112] The encoder of the deep spatiotemporal prediction model is a recurrent neural network built on Long Short-Term Memory (LSTM) network. LSTM is a special type of recurrent neural network (RNN) structure. Its core technology principle lies in the fact that it contains three gating mechanisms (input gate, forget gate, and output gate) and a cell state. These designs enable it to effectively learn long-term dependencies in long sequence data and alleviate the gradient vanishing or exploding problems that may occur in traditional RNNs during training. The encoder is usually composed of multiple stacked LSTM units, with each layer using the hidden state output of the previous layer as part of the input. RNN is a deep learning model that processes sequential data. Its characteristic is that it enables the network to have memory capabilities through recurrent connections, and can capture the dependencies of data in time or order.

[0113] For example, a sequence of high-dimensional feature embedding vectors arranged sequentially by time step is input into the encoder of a trained deep spatiotemporal prediction model. Starting from the first time step, the LSTM unit receives the embedding vector of the current time step. and the hidden state vector from the previous time step and cell state Update the current cell state through gating computation. Output the hidden state vector at the current time step. This process iterates until the end of the sequence. After processing the entire sequence of length... After the high-dimensional feature embedding vector sequence, the encoder's last time step (i.e., the first step) The hidden state vector output by step 1) This is considered to have encoded and condensed all the temporal patterns and historical information contained in the entire input sequence. This final hidden state vector... Determined as a latent variable ,Right now This fixed-dimensional vector It is a model of the past A highly abstract and generalized internal representation of how all complex environmental factors affect the state of timber within an hour.

[0114] Step S403: Based on latent variables, the decoder of the trained deep spatiotemporal prediction model is used to extrapolate future brittleness risks and generate a continuous risk probability curve for the target wooden component to undergo brittle fracture within a preset future time period.

[0115] In this context, the decoder of the deep spatiotemporal prediction model is typically an LSTM-based network. Its initial hidden states and cell states are initialized by a linear transformation of the latent variables to ensure that memories of historical states are passed on to the prediction stage. Future brittleness risk projection refers to the reverse deduction of the brittleness risk changes of wooden components at different time points and spatial locations based on trained fracture patterns, generating intuitive risk projection results. The preset future time period refers to a pre-defined time interval pointing to the future, such as […]. ].

[0116] For example, latent variables are input into the decoder of a pre-trained deep spatiotemporal prediction model for multi-step forward inference to generate future risk predictions. The decoder operates in an autoregressive manner, predicting the future step by step. Ultimately, a continuous risk probability curve is generated over a preset future time period. Multi-step forward inference is a way of thinking and method that uses the current state to progressively predict possible future development paths to assist in decision-making or analysis; autoregression is a modeling approach that predicts the future based on the historical information of the data itself, with the core logic of using the past to predict the present / future.

[0117] In this embodiment, the original heterogeneous feature matrix is ​​nonlinearly mapped to a high-dimensional, dense latent representation space through an embedding layer. An LSTM-based encoder performs deep bidirectional or unidirectional temporal dependency learning on the feature embedding sequence, gradually abstracting and ultimately condensing latent variables that carry all historical information. Starting from these latent variables, an autoregressive multi-step deduction is performed using an attention-based decoder to generate a continuous risk probability curve. This enables automated and intelligent conversion from raw environmental data to intuitive risk probabilities, providing far more refined decision support information than traditional threshold alarms for the preventative maintenance of wooden components.

[0118] In one embodiment, based on latent variables, the decoder of a trained deep spatiotemporal prediction model is used to extrapolate future brittleness risks, generating a continuous risk probability curve for brittle fracture of the target wooden component within a preset future time period, including:

[0119] Step S501: Based on the latent variables, initialize the hidden state and cell state of the decoder of the trained deep spatiotemporal prediction model; wherein, the decoder is constructed based on a long short-term memory network.

[0120] For example, setting a suitable initial state for an LSTM-based decoder involves the internal computation of the LSTM unit relying on two key state vectors: the hidden state and the cell state. The hidden state can be understood as the network's "short-term memory" or output at the current time step, while the cell state acts as a carrier of "long-term memory," responsible for transmitting information between sequence time steps. For example, the hidden variables are processed through two independent, learnable linear transformation layers (i.e., fully connected layers). These are respectively mapped to the initial hidden states of the decoder LSTM. and initial cell state This initialization process can be formally represented as: and ,in, It is a weight matrix. It is a bias vector. It is the hyperbolic tangent activation function. The essence of a fully connected layer is a linear transformation; it maps input data from the original feature space to a new feature space using a mathematical formula. Mapping is a core concept commonly used in mathematics and computer science; essentially, it describes the correspondence between two sets (which can be understood as a set of elements). The core idea is to find a unique matching element in another set for each element in one set.

[0121] Step S502: Based on the initialized hidden state and cell state, at each prediction time step of the decoder, calculate the attention weights between each element in the current decoder hidden state and the hidden state sequence of all time steps of the encoder.

[0122] For example, at each prediction time step of the decoder , Starting from 1, corresponding to the first prediction time in the future, based on the current hidden state of the decoder LSTM. (Initialized in the first step) The sequence of hidden states at all time steps generated by the encoder when processing the entire history sequence. The current hidden state of the decoder is calculated through an attention mechanism. and each hidden state in the encoder sequence The correlation score between them, for example: in dot product attention, the correlation score is calculated as... For all of the encoder sequences Location Perform this calculation on all of them to obtain the result containing The original scores are represented by a score vector. This score vector can be normalized using the Softmax function, transforming the original scores into a probability distribution, i.e., attention weights. The calculation formula is as follows:

[0123]

[0124] The resulting weight sequence satisfy And each weight Weight The size is used to visually reflect the generation of the first... When making future predictions, the model compares the historical data to the first future prediction. The attention level of information at each time step. The core principle of the attention mechanism is to simulate the selective focusing ability of humans when processing information, allowing the model to dynamically and selectively focus on different parts of the input sequence when generating the current output.

[0125] Step S503: Based on the attention weights, perform a weighted summation of each element in the hidden state sequence to generate the current prediction time step context vector of the decoder.

[0126] Weighted summation is a method for calculating the sum of multiple data points unequally. Its core is to assign different weights (i.e., importance coefficients) to different data points, and then obtain the final result by summing the products of "data points × weights". In essence, it allows more important data points to have a greater influence in the sum.

[0127] For example, the state of each encoder is hidden. and its corresponding attention weights Multiplying them together yields the vector of the state after weight modulation. For all The modulated vectors are summed to generate a single, comprehensive vector, called the context vector. The calculation formula is as follows:

[0128]

[0129] This context vector It is a summary of the essence of historical information that the decoder focuses on in the current prediction step, and its dimension is the same as that of the encoder's hidden state.

[0130] Step S504: Calculate the fragility risk probability of the current prediction time step based on the current prediction time step context vector and the current decoder hidden state.

[0131] For example, at the current prediction time step Two key vectors have been obtained: the decoder's hidden state, representing the decoder's current memory state. and the context vector representing historical information after focusing Vectors can be used as vectors. and The features are concatenated along the feature dimension to generate a joint vector with richer features. This joint vector is then input into a fully connected layer, which is essentially a linear transformation combined with a non-linear activation function. This fully connected layer maps the high-dimensional joint vector to a scalar value. This scalar value can be mapped to a scalar value using the sigmoid activation function. Within the interval, obtain the brittleness risk probability at the current prediction time step. This fragility risk probability is used to represent the model's prediction of the fragility risk in the future. Each time step (corresponding to absolute time) The instantaneous probability of brittle fracture of the target wooden component. The Sigmoid activation function is a commonly used nonlinear activation function in early deep learning, used to map any real-valued input to the (0,1) interval, achieving probabilistic output or nonlinear transformation. Its mathematical expression is: Concatenating features along the feature dimension is the core operation for integrating multi-source features in machine learning (especially deep learning). The core logic is to horizontally stack different feature dimensions to retain the complete information of each feature. Linear transformation is a mapping relationship in linear algebra that maintains linear operations between vector spaces. The core is to preserve the two key properties of addition and scalar multiplication. It can be understood as performing regular transformations on vectors without destroying their linear structure.

[0132] Step S505: Based on the current predicted time step fragility risk probability, repeat the steps from calculating attention weights to calculating the current predicted time step fragility risk probability, and sequentially generate the fragility risk probability for each time step within the preset future time period.

[0133] For example, the current prediction time step fragility risk probability Combined with any other auxiliary information (if any), into an input vector. This combined vector LSTM as the decoder in the The "external input" or "output input of the previous time step" is used at the current time step. Simultaneously, the decoder LSTM will use its internal state (i.e., the current time step's output input) to input the "external input" or "output input of the previous time step". Hidden state updated after step calculation and cell state (This is passed to the next time step.) and As the new current state, As new input, the loop starting from step S502 is repeated: that is, first based on the new decoder hidden state. (Based on LSTM) (Calculated from the previous state) and the encoder hidden state sequence Calculate the new attention weights, and then generate a new context vector. Finally, the risk probability for the next future time step is calculated. This iterative process of "calculating attention weights - generating context vectors - calculating risk probabilities - feedback" is repeated until a vector of length [length missing] is generated. The future predicted sequence, i.e. ,in The total number of time steps included in the preset future time period.

[0134] Step S506: Connect all the fragility risk probabilities to generate a continuous risk probability curve.

[0135] For example, for each predicted probability Assign a corresponding future timestamp. Assume the current historical observation window ends at [time]. The prediction time interval is (For example, 1 hour), then the first The absolute future time corresponding to each prediction point is: Using these calculated future time points as the x-axis and the corresponding predicted probability values ​​as y-axis... As the ordinate, construct an ordered list of data points. A direct sequential connection method can be used to connect these data points on a graph or numerical sequence, forming a path from... arrive A continuous risk probability curve over a time frame. This curve is used to show the predicted trajectory of the risk of brittle fracture of wooden components over a predetermined future time period.

[0136] In this embodiment, latent variables are used to initialize the hidden state and cell state of the decoder of the trained deep spatiotemporal prediction model. At each prediction time step of the decoder, the attention weights between the current decoder hidden state and each element in the hidden state sequence of all time steps of the encoder are calculated. Based on the attention weights, the elements in the hidden state sequence are weighted and summed to generate the context vector of the current prediction time step of the decoder. Combined with the current decoder hidden state, the fragility risk probability of the current prediction time step is calculated. The steps of calculating the attention weights to calculating the fragility risk probability of the current prediction time step are repeated to generate the fragility risk probability of each time step within a preset future time period. All fragility risk probabilities are connected to generate a continuous risk probability curve. This method can capture the complete dynamic process of risk incubation, development, and aggravation, providing an unprecedented early warning capability with high spatiotemporal resolution for the safe maintenance of wooden components.

[0137] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0138] Based on the same inventive concept, this application also provides a deep learning-based device for predicting the low-temperature brittleness of wood components, which implements the aforementioned deep learning-based method for predicting the low-temperature brittleness of wood components. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more deep learning-based devices for predicting the low-temperature brittleness of wood components provided below can be found in the limitations of the deep learning-based method for predicting the low-temperature brittleness of wood components described above, and will not be repeated here.

[0139] In one exemplary embodiment, such as Figure 3 As shown, a deep learning-based low-temperature brittleness prediction device 300 for wooden components is provided, comprising:

[0140] The data acquisition module 301 is used to acquire the temperature and humidity time series data of the environment in which the target wooden component is located, perform standardized preprocessing on the temperature and humidity time series data, and generate a standardized environmental data sequence; wherein, the temperature and humidity time series data includes historical temperature and humidity time series data and real-time monitoring temperature and humidity time series data;

[0141] The feature calculation module 302 is used to calculate the temperature and humidity coupled derivative index based on the standardized environmental data sequence, and generate a multidimensional feature sequence based on the original time-series statistical features of the temperature and humidity coupled derivative index and the standardized environmental data sequence; wherein, the multidimensional feature sequence is used to characterize the joint driving force of the environment on wood embrittlement.

[0142] The matrix generation module 303 is used to extract a multi-dimensional feature sequence according to a first preset sliding time window, generate multiple feature matrices, and assign a label to each feature matrix to obtain a labeled feature matrix; wherein the label is the fragile risk state corresponding to the future time.

[0143] The sample generation module 304 is used to arrange all labeled feature matrices in chronological order to generate deep spatiotemporal sequence samples.

[0144] The prediction curve generation module 305 is used to input deep spatiotemporal sequence samples into a trained deep spatiotemporal prediction model to generate a continuous risk probability curve of brittle fracture of the target wooden component within a preset future time period; wherein, the encoder of the trained deep spatiotemporal prediction model includes an embedding layer, an encoder, and a decoder.

[0145] In one embodiment, the feature calculation module 302 is further configured to:

[0146] Based on the temperature and relative humidity values ​​in the standardized environmental data sequence, the absolute humidity estimate for each time point is calculated, and all absolute humidity estimates are arranged in chronological order to generate the first time-series feature sequence.

[0147] The temperature and relative humidity values ​​are multiplied point by point to generate a temperature and humidity interaction term for each time point. The temperature and humidity interaction term for the current time point is combined with the temperature and humidity interaction term for the preset historical lag time to obtain the temperature and humidity interaction feature group for the current time point.

[0148] Arrange all temperature and humidity interaction feature groups in chronological order to generate a second time-series feature sequence;

[0149] Based on temperature and relative humidity values, the embrittlement driving force index at each time point was calculated; the embrittlement driving force index is used to reflect the degree of synergistic stress of low temperature and high humidity.

[0150] Arrange all embrittlement driving force indices in chronological order to generate an embrittlement driving force index sequence;

[0151] Within the second preset sliding time window, the statistical characteristics of the standardized environmental data sequence, the first time-series feature sequence, the second time-series feature sequence, and the embrittlement driving force index sequence are calculated respectively; wherein, the statistical characteristics include the mean, standard deviation, maximum value, and minimum value;

[0152] The statistical features within the second preset sliding time window are combined to generate a high-dimensional feature vector corresponding to the second preset sliding time window.

[0153] Arrange all high-dimensional feature vectors in chronological order to generate a multi-dimensional feature sequence.

[0154] In one embodiment, the feature calculation module 302 is further configured to:

[0155] Based on temperature and relative humidity values, the embrittlement driving force index at each time point was calculated using the embrittlement driving force index calculation formula; the embrittlement driving force index is used to reflect the degree of synergistic stress from low temperature and high humidity.

[0156]

[0157] in, The embrittlement driving force index, The preset reference temperature for the onset of embrittlement effect. This is an adjustable scaling factor. This is the relative humidity value. This is the temperature value.

[0158] In one embodiment, the prediction curve generation module 305 is further configured to:

[0159] The feature matrix from the deep spatiotemporal sequence sample is input into the embedding layer of the trained deep spatiotemporal prediction model to generate a high-dimensional feature embedding vector, which is arranged continuously according to the time step.

[0160] The high-dimensional feature embedding vector is input into the encoder of the trained deep spatiotemporal prediction model to generate the hidden state vector at each time step, and the hidden state vector at the last time step is determined as the latent variable; where the encoder is constructed based on a long short-term memory network; the latent variable is used to characterize the historical embrittlement state and the current embrittlement state of the wood.

[0161] Based on latent variables, the decoder of a trained deep spatiotemporal prediction model is used to extrapolate future brittleness risks and generate a continuous risk probability curve for the brittle fracture of the target wooden component within a preset future time period.

[0162] In one embodiment, the prediction curve generation module 305 is further configured to:

[0163] Based on latent variables, the hidden states and cell states of the decoder of the trained deep spatiotemporal prediction model are initialized; the decoder is constructed based on a long short-term memory network.

[0164] Based on the initialized hidden state and cell state, at each prediction time step of the decoder, the attention weights between each element in the current decoder hidden state and the hidden state sequence of all time steps of the encoder are calculated.

[0165] Based on the attention weights, the elements in the hidden state sequence are summed in a weighted manner to generate the current prediction time step context vector of the decoder.

[0166] Based on the context vector of the current prediction time step and the current decoder hidden state, the fragility risk probability of the current prediction time step is calculated.

[0167] Based on the current predicted time step fragility risk probability, repeat the steps from calculating attention weights to calculating the current predicted time step fragility risk probability, and sequentially generate the fragility risk probability for each time step within the preset future time period.

[0168] Connect all the fragile risk probabilities to generate a continuous risk probability curve.

[0169] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the deep learning-based method for predicting the low-temperature brittleness of wood components as described above.

[0170] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0171] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0172] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for predicting the low-temperature brittleness of wooden components based on deep learning, characterized in that, The method includes: The temperature and humidity time series data of the environment in which the target wooden component is located are obtained, and the temperature and humidity time series data are standardized and preprocessed to generate a standardized environmental data sequence; wherein, the temperature and humidity time series data includes historical temperature and humidity time series data and real-time monitoring temperature and humidity time series data. Based on the standardized environmental data sequence, a temperature and humidity coupled derivative index is calculated, and a multidimensional feature sequence is generated based on the original time-series statistical characteristics of the temperature and humidity coupled derivative index and the standardized environmental data sequence; wherein, the multidimensional feature sequence is used to characterize the joint driving force of the environment on wood embrittlement. The multidimensional feature sequence is extracted according to a first preset sliding time window to generate multiple feature matrices, and a label is assigned to each feature matrix to obtain a labeled feature matrix; wherein, the label is the fragile risk state corresponding to a future time. Arrange all the labeled feature matrices in chronological order to generate deep spatiotemporal sequence samples; The deep spatiotemporal sequence samples are input into the trained deep spatiotemporal prediction model to generate a continuous risk probability curve of brittle fracture of the target wooden component within a preset future time period; wherein, the encoder of the trained deep spatiotemporal prediction model includes an embedding layer, an encoder, and a decoder.

2. The method according to claim 1, characterized in that, Based on the standardized environmental data sequence, a temperature and humidity coupled derivative index is calculated, and a multidimensional feature sequence is generated based on the original time-series statistical characteristics of the temperature and humidity coupled derivative index and the standardized environmental data sequence; wherein, the multidimensional feature sequence is used to characterize the joint driving force of the environment on wood embrittlement, including: Based on the temperature and relative humidity values ​​in the standardized environmental data sequence, the absolute humidity estimate for each time point is calculated, and all the absolute humidity estimates are arranged in chronological order to generate a first time-series feature sequence. The temperature value and the relative humidity value are multiplied point by point to generate a temperature and humidity interaction item for each time point. The temperature and humidity interaction item at the current time point is combined with the temperature and humidity interaction item at a preset historical lag time to obtain the temperature and humidity interaction feature group at the current time point. Arrange all the temperature and humidity interaction feature groups in chronological order to generate a second time-series feature sequence; Based on the temperature and relative humidity values, the embrittlement driving force index at each time point is calculated; wherein, the embrittlement driving force index is used to reflect the degree of synergistic stress of low temperature and high humidity; Arrange all the embrittlement driving force indices in chronological order to generate an embrittlement driving force index sequence; Within a second preset sliding time window, the statistical characteristics of the standardized environmental data sequence, the first time-series feature sequence, the second time-series feature sequence, and the embrittlement driving force index sequence are calculated respectively; wherein, the statistical characteristics include mean, standard deviation, maximum value, and minimum value; The statistical features within the second preset sliding time window are combined to generate a high-dimensional feature vector corresponding to the second preset sliding time window; All the high-dimensional feature vectors are arranged in chronological order to generate a multi-dimensional feature sequence.

3. The method according to claim 2, characterized in that, The embrittlement driving force index is calculated based on the temperature and relative humidity values ​​at each time point; wherein, the embrittlement driving force index is used to reflect the degree of synergistic stress of low temperature and high humidity, including: Based on the temperature and relative humidity values, the embrittlement driving force index at each time point is calculated using the following formula: in, The embrittlement driving force index, The preset reference temperature for the onset of embrittlement effect. This is an adjustable scaling factor. This is the relative humidity value. This is the temperature value.

4. The method according to claim 1, characterized in that, The deep spatiotemporal sequence samples are input into a trained deep spatiotemporal prediction model to generate a continuous risk probability curve for brittle fracture of the target wooden component within a preset future time period; wherein, the encoder of the trained deep spatiotemporal prediction model includes an embedding layer, an encoder, and a decoder, comprising: The feature matrix in the deep spatiotemporal sequence sample is input into the embedding layer of the trained deep spatiotemporal prediction model to generate a high-dimensional feature embedding vector, which is arranged continuously according to time steps. The high-dimensional feature embedding vector is input into the encoder of the trained deep spatiotemporal prediction model to generate a hidden state vector at each time step, and the hidden state vector at the last time step is determined as a latent variable; wherein, the encoder is constructed based on a long short-term memory network; the latent variable is used to characterize the historical brittleness state and the current brittleness state of the wood; Based on the latent variables, the decoder of the trained deep spatiotemporal prediction model is used to extrapolate future brittleness risks and generate a continuous risk probability curve for the target wooden component to undergo brittle fracture within a preset future time period.

5. The method according to claim 4, characterized in that, The step of extrapolating future brittleness risk based on the latent variables and using the decoder of the trained deep spatiotemporal prediction model to generate a continuous risk probability curve for the target wooden component to undergo brittle fracture within a preset future time period includes: Based on the latent variables, the hidden state and cell state of the decoder of the trained deep spatiotemporal prediction model are initialized; wherein, the decoder is constructed based on a long short-term memory network. Based on the initialized hidden state and the cell state, at each prediction time step of the decoder, the attention weight between the current decoder hidden state and each element in the hidden state sequence of all time steps of the encoder is calculated. Based on the attention weights, the elements in the hidden state sequence are weighted and summed to generate the current prediction time step context vector of the decoder. Based on the current prediction time step context vector and the current decoder hidden state, the fragility risk probability of the current prediction time step is calculated; Based on the current predicted time step fragility risk probability, the steps of calculating the attention weight and calculating the current predicted time step fragility risk probability are repeated to generate the fragility risk probability of each time step in the preset future time period in sequence. Connect all the aforementioned fragility risk probabilities to generate a continuous risk probability curve.

6. A deep learning-based device for predicting the low-temperature brittleness of wooden components, characterized in that, The device includes: The data acquisition module is used to acquire the temperature and humidity time series data of the environment in which the target wooden component is located, and to perform standardized preprocessing on the temperature and humidity time series data to generate a standardized environmental data sequence; wherein, the temperature and humidity time series data includes historical temperature and humidity time series data and real-time monitoring temperature and humidity time series data; The feature calculation module is used to calculate the temperature and humidity coupled derivative index based on the standardized environmental data sequence, and generate a multidimensional feature sequence based on the temperature and humidity coupled derivative index and the original time-series statistical features of the standardized environmental data sequence; wherein, the multidimensional feature sequence is used to characterize the joint driving force of the environment on wood embrittlement. The matrix generation module is used to extract the multidimensional feature sequence according to a first preset sliding time window, generate multiple feature matrices, and assign a label to each feature matrix to obtain a labeled feature matrix; wherein the label is the fragile risk state corresponding to a future time. The sample generation module is used to arrange all the labeled feature matrices in chronological order to generate deep spatiotemporal sequence samples. The prediction curve generation module is used to input the deep spatiotemporal sequence samples into the trained deep spatiotemporal prediction model to generate a continuous risk probability curve of the target wooden component undergoing brittle fracture within a preset future time period; wherein, the encoder of the trained deep spatiotemporal prediction model includes an embedding layer, an encoder, and a decoder.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images