Bone tissue bending resistance test data analysis method and system

Abstract

The present application provides a method and system for analyzing bone tissue anti-bending test data. When analyzing bone tissue anti-bending test data, the pressure response characterization vector of the target bone tissue anti-bending test data is extracted independently. Then, based on the pressure response characterization vector of the target bone tissue anti-bending test data, the disturbance data of the target bone tissue anti-bending test data and the reconstructed test data of the target bone tissue anti-bending test data at the previous level, when data reconstruction is performed, it can help the data complete the reconstruction of more subtle information and prevent excessive filtering caused by the de-disturbance process, thereby improving the data reconstruction effect.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular, to a method and system for analyzing bone tissue anti-bending test data. Background Art

[0002] In the field of biomedical engineering and materials science, accurate evaluation of the mechanical properties of bone tissue has always been one of the research hotspots. Bone tissue bending test is an important means to evaluate the mechanical properties of bone materials. The accuracy and reliability of its test data are directly related to the correctness of the evaluation results. However, in the actual test process, due to the influence of various factors such as experimental equipment, environmental conditions, and operating errors, the collected test data often inevitably contain disturbance components. These disturbance components not only reduce the signal-to-noise ratio of the test data, but also may mislead the subsequent mechanical properties evaluation, resulting in inaccurate evaluation results. In order to eliminate the disturbance components in the test data and improve the purity and accuracy of the data, researchers have developed a variety of data cleaning and denoising methods. However, traditional methods often have problems such as excessive filtering and loss of details, and it is difficult to retain the key feature information in the data while removing the disturbance. Especially in the analysis of bone tissue bending test data, due to the high complexity and nonlinear characteristics of the mechanical response of bone tissue, traditional methods often find it difficult to achieve ideal denoising effects. Summary of the invention

[0003] The object of the present invention is to provide a method and system for analyzing bone tissue bending resistance test data. The embodiment of the present application is implemented as follows:

[0004] In a first aspect, an embodiment of the present application provides a method for analyzing bone tissue anti-bending test data, comprising: obtaining target bone tissue anti-bending test data to be analyzed, wherein the target bone tissue anti-bending test data includes disturbance data; based on a representation information extraction network in a branch model at the jth level, extracting representation information from the target bone tissue anti-bending test data to obtain a pressure response representation vector of the target bone tissue anti-bending test data at the jth level, wherein the branch model at the jth level is a branch model at the jth level in an I-level branch model included in a target data disturbance cleaning model, wherein j>1; based on a disturbance information identification network in the branch model at the jth level, clearing disturbances from the disturbance data at the i-th level and the reconstructed test data at the i-th level of the target bone tissue anti-bending test data output by the branch model at the i-th level, Obtain the disturbance data of the target bone tissue anti-bending test data at the j-th level, the branch model of the i-th level is the previous level of the j-th level in the I levels, wherein i=j-1; based on the data reconstruction network in the branch model of the j-th level, perform data reconstruction operation on the pressure response characterization vector of the j-th level, the disturbance data of the j-th level and the reconstructed test data of the i-th level to obtain the reconstructed test data of the target bone tissue anti-bending test data at the j-th level; determine the disturbance data of the j-th level and the reconstructed test data of the j-th level as the input information of the branch model of the j+1-th level, iterate one by one, obtain the reconstructed test data of the target bone tissue anti-bending test data at the I-th level, as the reconstructed test data after the disturbance of the target bone tissue anti-bending test data is cleared.

[0005] In a second aspect, an embodiment of the present application provides a computer system, comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the above-mentioned method is implemented.

[0006] The beneficial effects of the present application include at least: the present application adopts the method of obtaining target bone tissue anti-bending test data including disturbance data, and performing disturbance clearing on the target bone tissue anti-bending test data based on a target data disturbance clearing model, so as to obtain reconstructed test data of the target bone tissue anti-bending test data after disturbance clearing, the target data disturbance clearing model includes a branch model of I levels, for the branch model of the j-th level, based on the representation information extraction network in the branch model of the j-th level, the representation information of the target bone tissue anti-bending test data is extracted, and the pressure response representation vector of the target bone tissue anti-bending test data at the j-th level is obtained; based on the disturbance information identification network in the branch model of the j-th level, the disturbance data at the i-th level and the pressure response representation vector of the target bone tissue anti-bending test data output by the branch model of the i-th level are extracted. The reconstructed test data at the i-th level are disturbed and cleared to obtain the disturbed data of the target bone tissue anti-bending test data at the j-th level; based on the data reconstruction network in the branch model at the j-th level, the pressure response characterization vector at the j-th level, the disturbed data at the j-th level and the reconstructed test data at the i-th level are reconstructed to obtain the reconstructed test data of the target bone tissue anti-bending test data at the j-th level, i=j-1; the disturbed data at the j-th level and the reconstructed test data at the j-th level are determined as the input of the branch model at the j+1-th level, and the operations are iterated one by one to finally obtain the reconstructed test data of the target bone tissue anti-bending test data at the I-th level, and then the reconstructed test data at the I-th level are determined as the reconstructed test data of the target bone tissue anti-bending test data after the disturbance is cleared. Based on this, when analyzing the bone tissue anti-bending test data, the present application independently extracts the pressure response characterization vector of the target bone tissue anti-bending test data, and then reconstructs the data based on the pressure response characterization vector of the target bone tissue anti-bending test data, the disturbance data of the target bone tissue anti-bending test data, and the reconstructed test data of the target bone tissue anti-bending test data at the previous level. This can help the data complete the reconstruction of more subtle information and prevent excessive filtering caused by the de-disturbance process, thereby improving the data reconstruction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 This is a flow chart of a bone tissue anti-bending test data analysis method provided in this application.

[0008] Figure 2 It is a schematic diagram of the composition of a computer system provided by this application. DETAILED DESCRIPTION

[0009] The execution subject of the bone tissue bending resistance test data analysis method in the embodiment of the present application is a computer system, including but not limited to a server, a personal computer, a laptop, a tablet computer, a smart phone, etc. The computer system includes a user device and a network device. Figure 1As shown, the method includes:

[0010] Step S100: Acquire the target bone tissue anti-bending test data to be analyzed, wherein the target bone tissue anti-bending test data includes disturbance data.

[0011] In the embodiment of the present application, it is assumed that a biomedical research institute is conducting an evaluation project on the mechanical properties of a new bone substitute material. In order to verify the performance of this material in practical applications, researchers have designed a series of bending resistance tests, including bending performance tests on target bone tissue (for example, samples for using new materials or control samples). During the test, it can be carried out on a professional mechanical testing machine, which can accurately measure and record various parameters of the sample when it is bent by external force, such as loading force, displacement, stress distribution, etc. The computer system is connected to the interface of the testing machine to obtain these original test data in real time or in batches afterwards. The data may include basic test parameters, dynamic response data, and disturbance data, etc. Among them, the basic test parameters include but are not limited to test time, sample number, loading rate, maximum loading force, etc. Dynamic response data, for example, are displacement-time curves, stress-strain curves, etc. of the sample during loading. These data are usually recorded in the form of time series, which can intuitively reflect the mechanical behavior of bone tissue samples during bending. Due to various factors such as experimental environment, equipment accuracy, and operating errors, some non-target signals, i.e., disturbance data, are often mixed into the actual test data. For example, the vibration of the testing machine itself, slight fluctuations in ambient temperature, electromagnetic interference, etc. may all affect the data. These disturbed data will interfere with the accurate analysis of the target signal (i.e., the true anti-bending performance of bone tissue).

[0012] The data received by the computer system is usually stored in a specific file format, such as CSV, Excel or a dedicated data exchange format. These files contain detailed records of all the test parameters and dynamic response data mentioned above. Before formal analysis, the computer system may need to perform some preprocessing on the raw data, such as removing outliers, smoothing to reduce noise, standardization or normalization, etc. Although these preprocessing steps are not directly part of step S100, they are important links to ensure the accuracy of subsequent analysis.

[0013] For example, suppose the computer system receives a CSV file from a test machine that contains the following sample data rows shown in Table 1:

[0014]

[0015] Table 1 Test data csv file example

[0016] In this example, SampleID represents the sample number, Time(s) represents the test time, Load(N) is the loading force (unit: Newton), Displacement(mm) is the displacement of the sample (unit: millimeter), Temperature(°C) records the ambient temperature, and NoiseLevel(dB) is an indicator that may represent the ambient noise level, which can also be regarded as part of the disturbance data.

[0017] Step S200: Based on the representation information extraction network in the j-th level branch model, the representation information of the target bone tissue anti-bending test data is extracted to obtain the pressure response representation vector of the target bone tissue anti-bending test data at the j-th level. The j-th level branch model is the j-th level branch model among the I-level branch models included in the target data disturbance cleaning model, where j>1.

[0018] In step S200, the computer system uses the j-th level branch model in the target data perturbation cleaning model, especially the characterization information extraction network within this level, to conduct an in-depth analysis of the target bone tissue anti-bending test data to extract the key pressure response characterization vector.

[0019] For example, suppose that in the process of analyzing the performance of a series of bone tissue samples in bending tests, the data contains complex mechanical response information, and is also mixed with various perturbations. In order to extract useful features from this data, the computer system uses a multi-level branching model structure, with each level focusing on extracting a specific level of information.

[0020] At the jth level (assuming j = 2 for simplicity), the representation information extraction network kicks in. This network is, for example, a complex deep learning model, such as a variant of a convolutional neural network (CNN) or a recurrent neural network (RNN), but is customized and optimized for the characteristics of bone tissue bending test data.

[0021] The computer system uses the original test data obtained in step S100 (after necessary preprocessing) as input. These data may include time series information such as loading force, displacement, and timestamp.

[0022] The network structure includes, for example:

[0023] Convolutional layer: used to automatically extract local features from the input data. In the bending test of bone tissue, these features may correspond to the stress distribution pattern under different loading stages.

[0024] Pooling layer: used to reduce the dimension of feature maps, reduce the amount of calculation, and retain important information. This helps the system ignore some unimportant details and focus on more macroscopic mechanical response characteristics.

[0025] Activation function: such as ReLU (Rectified Linear Unit), which is used to introduce nonlinear factors so that the network can learn more complex feature representations.

[0026] Fully connected layer: After the convolution and pooling layers, there may be several fully connected layers to integrate the previously extracted features and output the final pressure response representation vector.

[0027] After being processed by the representation information extraction network, a high-dimensional vector is output, namely the j-th level pressure response representation vector. This vector captures the key mechanical response characteristics of the target bone tissue in the bending test, such as the potential representation of elastic modulus, yield strength, fracture toughness and other indicators.

[0028] For example, assume that the representation information extraction network outputs a 128-dimensional pressure response representation vector, in which each dimension represents an abstract representation of the input data in a specific aspect. For example, the first dimension of the vector may reflect the change in stiffness of the sample during the initial loading stage, while the 64th dimension may be related to the stress concentration phenomenon of the sample when it approaches the fracture point. These feature vectors not only reduce the dimensionality of the original data, but also improve the efficiency and accuracy of subsequent analysis. In steps S300 and S400, the computer system will perform disturbance identification and data reconstruction based on these representation vectors to further clean up the data and restore a more realistic performance of the bone tissue's anti-bending performance.

[0029] Step S300: Based on the disturbance information identification network in the branch model of the j-th level, the disturbance data of the target bone tissue anti-bending test data at the i-th level output by the branch model of the i-th level and the reconstructed test data at the i-th level are cleared to obtain the disturbance data of the target bone tissue anti-bending test data at the j-th level, and the branch model of the i-th level is the previous level of the j-th level in the I levels, wherein i=j-1.

[0030] In step S300, the disturbance information recognition network in the branch model of the j-th level is used to analyze the output data from the previous level (i.e., the i-th level, where i=j-1), aiming to separate and remove the disturbance data therein to obtain purer target data features.

[0031] Test data is often disturbed by a variety of external factors, such as equipment vibration, ambient temperature fluctuations, electromagnetic interference, etc. These factors appear as unexpected fluctuations or noise in the data, i.e., disturbed data. In order to obtain accurate mechanical property evaluation, these disturbances must be effectively identified and removed.

[0032] The disturbance information recognition network is, for example, a complex deep learning model, such as a combination of a convolutional neural network (CNN) and a long short-term memory network (LSTM), or a Transformer model based on a self-attention mechanism. These models learn to distinguish between target signals (i.e., the true anti-bending response of bone tissue) and disturbance signals (noise and interference) through training.

[0033] The network receives two main inputs - the perturbed data output by the i-th layer and the reconstructed test data. The perturbed data is a preliminary estimate of the noise component in the original data by the previous layer, while the reconstructed test data is a dataset that has been initially cleaned and reconstructed.

[0034] First, the network fuses the perturbation data and the reconstructed test data at the feature level to more accurately identify which parts belong to the perturbation and which parts belong to the real signal. For time series data (such as the change of loading force and displacement over time), the network uses sequence models such as LSTM or Transformer to capture long-term dependencies in the data to help distinguish periodic perturbations from real responses. By comparing the input data with the perturbation pattern learned by the model, the network outputs an updated estimate of the perturbation data at the current level. Finally, the network outputs the perturbation data of the jth level, which is more accurate than the previous level and reflects the effect of deeper data cleaning. For example, suppose that at the second level (j=2), the perturbation information recognition network receives the perturbation data and reconstructed test data output by the first level. Through its internal complex calculation logic, the network identifies the tiny vibration noise that still remains in the reconstructed test data, which may be difficult to detect in the original data, but cannot be ignored in high-precision analysis. For example, the network may calculate a new perturbation data estimate in the following way:

[0035] New perturbation data j = f (old perturbation data i, reconstructed test data i, θ);

[0036] Among them, f represents the mapping function of the perturbation information recognition network, θ is the parameter set of the network, including weights and biases, which are optimized through the training process. The old perturbation data i and the reconstructed test data i are the perturbation data and reconstructed test data output by the i-th level, respectively.

[0037] Step S400: Based on the data reconstruction network in the branch model of the j-th level, perform data reconstruction operations on the pressure response characterization vector of the j-th level, the disturbance data of the j-th level and the reconstructed test data of the ith level to obtain the reconstructed test data of the j-th level of the target bone tissue anti-bending test data.

[0038] In step S400, the computer system utilizes the data reconstruction network in the branch model of the jth level, integrates the pressure response characterization vector of the current level, the disturbance data, and the reconstructed test data of the previous level (the i-th level), and performs complex data reconstruction operations to generate a more accurate and pure version of the target bone tissue anti-bending test data.

[0039] The data reconstruction network is, for example, a complex model that integrates multiple deep learning techniques, including but not limited to autoencoders, variants of generative adversarial networks (GANs), or sequence-to-sequence (Seq2Seq) models combined with attention mechanisms. These models learn through training how to generate high-quality reconstructed test data based on input data (including pressure response representation vectors, perturbation data, and reconstructed test data of the previous layer).

[0040] The pressure response characterization vector of the jth level is a high-dimensional vector that captures the key mechanical response characteristics of the target bone tissue in the bending test. The perturbation data of the jth level is the perturbation estimate of the current level identified by the perturbation information recognition network, representing the noise and interference components in the data. The reconstructed test data of the i-th level is the data version processed by the previous level, which has been cleaned and reconstructed to a certain extent.

[0041] First, the network integrates the pressure response representation vector, perturbation data, and reconstructed test data of the previous level at the feature level. This may involve feature concatenation, weighted summation, or more complex feature interaction operations. The network takes the integrated features as input, compresses them into a low-dimensional representation through the encoder part, and reconstructs them back to the original data space through the decoder part, but in this process, the perturbation is reduced and the key features may be enhanced. If the GAN architecture is adopted, the generator will try to generate a reconstructed version that is as close to the real unperturbed data as possible, while the discriminator tries to distinguish the generated data from the real data. The two compete with each other during the training process and jointly improve the quality of the generated data. In order to control the reconstruction process more accurately, an attention mechanism can be introduced to enable the model to dynamically focus on the key parts of the input data, thereby giving more weight during reconstruction. Finally, the data reconstruction network outputs the reconstructed test data of the jth level. This data is purer than the previous level and better retains the real mechanical response characteristics of the target bone tissue. For example, suppose that at the third level (j=3), the data reconstruction network receives the pressure response representation vector of the third level (such as a 128-dimensional vector containing potential representations of indicators such as elastic modulus and yield strength), the disturbance data of the third level (a vector with the same dimension as the input data, identifying the location and intensity of noise and interference), and the reconstructed test data of the second level (time series data that has been preliminarily cleaned).

[0042] Through complex internal calculations, such as using the encoding-decoding framework of the autoencoder, the input data is compressed and reconstructed, and the weights of different features are dynamically adjusted using the attention mechanism during the reconstruction process. Finally, the network outputs the third level of reconstructed test data, which significantly reduces the disturbance component while retaining the key information of the original test data, providing a more reliable data basis for subsequent mechanical performance evaluation.

[0043] Step S500: Determine the disturbance data of the j-th level and the reconstructed test data of the j-th level as the input information of the branch model of the j+1-th level, and iterate one by one to obtain the reconstructed test data of the target bone tissue anti-bending test data at the I-th level as the reconstructed test data after the disturbance of the target bone tissue anti-bending test data is cleared.

[0044] During the bending test of bone tissue, the collected test data often contains certain disturbances due to various reasons such as experimental conditions, equipment accuracy, environmental factors, etc. In order to obtain accurate mechanical performance evaluation results, the computer system uses a branching model with multiple levels to process these data. Each level is responsible for further reducing the disturbance in the data and generating purer reconstructed test data.

[0045] In step S500, the computer system first takes the disturbance data and reconstructed test data of the jth level as input and passes them to the branch model of the j+1th level. This process is recursive, which means that each level will repeat a similar data processing flow: first obtain the pressure response representation vector through the representation information extraction network, then identify and remove the disturbance through the disturbance information identification network, and finally generate new reconstructed test data through the data reconstruction network.

[0046] For the j+1th level, its input includes two parts: one is the disturbance data identified by the jth level, which reflects the current level’s estimate of the disturbance; the other is the reconstructed test data generated by the jth level, which is the data version after cleaning and reconstructing by the current level.

[0047] As the number of levels j increases, each level will perform more in-depth data processing based on the output of the previous level. Each level of the branch model will use its own representation information extraction network, perturbation information recognition network and data reconstruction network to perform refined analysis and reconstruction of the input data.

[0048] After each level is processed, new perturbation data and reconstructed test data are output. These data are then passed to the next level (i.e., the j+2th level) as new input to continue the iterative processing process. The iterative process continues until the predetermined number of levels I is reached. At this point, the reconstructed test data of the Ith level is regarded as the result after the final perturbation removal, which is used for subsequent mechanical property evaluation and analysis. For example, assuming that the model has a total of 5 levels (I=5), after the processing of the 4th level is completed, the computer system will use the perturbation data and reconstructed test data of this level as input and pass it to the branch model of the 5th level. In the 5th level, the model will again perform the steps of characterization information extraction, perturbation information identification, and data reconstruction to further optimize the data quality. Finally, the reconstructed test data output from the 5th level will be the final result of the entire perturbation cleaning process. After multi-level iterative processing, the perturbation component of this data has been minimized to the greatest extent, while retaining the key mechanical response characteristics in the anti-bending test of bone tissue, providing reliable data support for subsequent mechanical property evaluation. Through this iterative process, the computer system can efficiently process complex bone tissue bending resistance test data to ensure the accuracy and reliability of the evaluation results.

[0049] As an implementation manner, the characterization information extraction network includes a characterization information extraction branch network and a characterization information refinement branch network. The step S200, based on the characterization information extraction network in the branch model at the j-th level, extracts the characterization information of the target bone tissue anti-bending test data to obtain the pressure response characterization vector of the target bone tissue anti-bending test data at the j-th level, which may specifically include:

[0050] Step S210: extracting characterization information from the target bone tissue anti-bending test data based on the characterization information extraction branch network to obtain a first pressure response characterization vector of the target bone tissue anti-bending test data;

[0051] Step S220: Based on the representation information extraction branch network, the representation information of the first pressure response representation vector is refined to obtain the j-th level pressure response representation vector.

[0052] In the bone tissue bending test, the data that the computer system needs to process usually include time series parameters such as loading force, displacement, and timestamp. These data not only reflect the dynamic response of bone tissue during the stress process, but are also mixed with various disturbance factors such as environmental noise and equipment vibration. In order to preliminarily extract the features related to the mechanical properties of the target bone tissue from these data, the computer system first uses the representation information extraction branch network for processing. The representation information extraction branch network is, for example, a deep neural network containing multiple convolutional layers, pooling layers, and activation functions. These network layers gradually extract high-level feature representations from the original test data through layer-by-layer abstraction.

[0053] For example, suppose the input data is a two-dimensional array containing time series loading force and displacement, the rows of the array represent different time steps, and the columns correspond to the values ​​of loading force and displacement respectively. The representation information extraction branch network first extracts local features through the convolution layer. For example, a rapid increase in loading force may indicate that the bone tissue begins to enter the elastic deformation stage. Subsequently, the pooling layer further simplifies these features, ignoring short-term fluctuations and retaining long-term trends. Finally, after the nonlinear transformation of the activation function, the system obtains a preliminary set of feature vectors, namely the first pressure response representation vector. This vector may contain multiple elements, each of which corresponds to a key feature dimension in the input data.

[0054] In step S210, the system has obtained the first pressure response characterization vector through the characterization information extraction branch network. However, this vector may still contain some redundant information or features that are not refined enough. In order to further improve the quality and purity of the characterization vector, the system needs to use the characterization information extraction branch network to further refine the features.

[0055] The representation information extraction branch network is, for example, a more refined neural network structure, which may contain deeper layers, more complex connections, and more nonlinear activation functions. The design goal of this network is to further extract the features most relevant to the bending resistance of the target bone tissue from the first pressure response representation vector. The representation information extraction branch network is, for example, a multi-layer fully connected neural network, each layer containing multiple neurons, and connected to the output of the previous layer through weights and biases. This structure allows the network to learn complex nonlinear mapping relationships, thereby extracting higher-level abstract representations from the initially extracted features.

[0056] In order to focus more on key features, the representation information extraction branch network can also integrate an attention mechanism. The attention mechanism allows the network to dynamically adjust the importance weights of different features, thereby giving more attention in subsequent processing. In the analysis of bone tissue bending test data, this can help the system better capture changes in key mechanical properties such as elastic modulus and yield strength. In order to prevent overfitting and improve the generalization ability of the model, the extraction branch network may also use regularization techniques (such as L1 / L2 regularization) and dropout strategies. These techniques reduce the model's excessive dependence on training data by constraining model complexity or randomly discarding some neuron connections.

[0057] For example, in step S220, the computer system passes the first pressure response characterization vector as input to the characterization information extraction branch network. This network first performs further feature transformation and combination through multiple fully connected layers, and each layer attempts to extract more meaningful representations from the input features. At the same time, the attention mechanism may play a role in this process, dynamically adjusting the importance weights of different features. For example, if a feature (such as a sudden change in loading force at a specific time point) is highly correlated with the yield behavior of the target bone tissue, then the attention mechanism will increase the weight of this feature so that it will receive more attention in subsequent processing.

[0058] After being processed by the representation information extraction branch network, the system finally outputs a more refined and pure pressure response representation vector, that is, the j-th level pressure response representation vector. This vector not only contains the most relevant feature information about the anti-bending performance of the target bone tissue, but also reduces the influence of redundancy and noise through attention mechanism and regularization technology. This representation vector will serve as an important input for subsequent steps (such as perturbation identification and data reconstruction), helping the computer system to more accurately analyze and evaluate the mechanical properties of bone tissue.

[0059] As an implementation manner, the characterization information extraction branch network includes a characterization information extraction component and a residual structure. The step S210, based on the characterization information extraction branch network, extracts the characterization information of the target bone tissue anti-bending test data to obtain a first pressure response characterization vector of the target bone tissue anti-bending test data, may specifically include:

[0060] Step S211: extracting a second pressure response characterization vector of the target bone tissue anti-bending test data based on the characterization information extraction component;

[0061] Step S212: merging the second pressure response characterization vector and the target bone tissue anti-bending test data in the channel dimension to obtain the first pressure response characterization vector.

[0062] In the bone tissue bending test, the test data usually contains information in multiple dimensions, such as loading force, displacement, time stamp, etc. These data together describe the dynamic response of bone tissue during the stress process. In order to extract features closely related to the mechanical properties of the target bone tissue from these data, the computer system first uses the characterization information extraction component to perform preliminary processing on the original test data.

[0063] The representation information extraction component is, for example, a deep learning module containing multiple network layers, which are responsible for automatically learning and extracting useful features from raw data. Common network layers include convolutional layers, pooling layers, etc., which capture local and global patterns in the data through specific operations.

[0064] As the core part of the characterization information extraction component, the convolution layer applies multiple convolution kernels to the input data in a sliding window manner to extract local features. In the context of bone tissue bending test data, the convolution kernel may be designed to be sensitive to the specific shapes of the loading force curve and displacement curve, thereby capturing key mechanical response features such as elastic deformation and plastic yield.

[0065] The convolutional layer can be followed by one or more pooling layers to reduce the dimensionality and computational complexity of the feature map while retaining important information. Pooling operations such as max pooling or average pooling can help the system ignore detail noise and focus on more macroscopic feature patterns. In order to improve the nonlinear expression ability of the network, activation functions (such as ReLU, sigmoid, etc.) are applied to the output of the convolutional layer. In the ReLU function, all negative values ​​are set to zero and positive values ​​remain unchanged, which helps introduce nonlinear factors and enhance the feature learning ability of the network.

[0066] After multiple layers of convolution and possible pooling operations, the characterization information extraction component outputs a two-dimensional array, namely the second pressure response characterization vector. This vector contains the key feature information extracted from the original test data, providing rich material for subsequent data processing.

[0067] In step S211, the computer system has obtained the second pressure response characterization vector through the characterization information extraction component. However, although this vector contains rich feature information, it may still lose some important details in the original test data. In order to make up for this deficiency, step S212 proposes a strategy to merge the characterization vector with the original test data in the channel dimension.

[0068] The merging operation may involve simple stacking or more complex feature transformations. In simple cases, the system directly adds the characterization vector as a new channel to the original data. In more complex cases, the system may first perform certain transformations (such as scaling, normalization, etc.) on the characterization vector to ensure that it is consistent with the original data in terms of numerical range and distribution before merging. For example, assume that the original test data is a three-dimensional array with a shape of [number of samples, time step, number of features], where the feature number includes two dimensions: load force and displacement. The second pressure response characterization vector is a two-dimensional array with a shape of [number of samples, time step, feature dimension], where the feature dimension depends on the output size of the characterization information extraction component. In the merging operation, the computer system first expands a new channel on the third dimension (number of features) of the target bone tissue bending test data, and then fills the second pressure response characterization vector into this new channel. In this way, the shape of the original data becomes [number of samples, time step, number of features + 1], where the newly added channel contains the feature information obtained by the characterization information extraction component. The merged data not only retains all the information of the original test data, but also incorporates the key features automatically extracted from the data, providing a richer and more comprehensive input for subsequent data processing. This feature fusion method helps the computer system better understand the dynamic response characteristics of bone tissue in bending resistance testing, thereby more accurately evaluating its mechanical properties.

[0069] As an implementation mode, the disturbance information identification network includes a first deep network layer, and the step S300, based on the disturbance information identification network in the branch model of the jth level, performs disturbance removal on the disturbance data of the target bone tissue anti-bending test data at the i-th level output by the branch model of the i-th level and the reconstructed test data at the i-th level, to obtain the disturbance data of the target bone tissue anti-bending test data at the j-th level, which may specifically include:

[0070] Step S310: performing disturbance information calculation according to the disturbance data of the i-th level, the reconstructed test data of the i-th level and the target bone tissue anti-bending test data to obtain first disturbance data of the target bone tissue anti-bending test data;

[0071] Step S320: Based on the first deep network layer, clear the disturbance of the first disturbance data to obtain the disturbance data of the j-th level.

[0072] In the jth level of the target data perturbation cleaning model, the computer system uses a perturbation information recognition network to conduct in-depth analysis of the perturbed data and reconstructed test data from the i-th level, aiming to further identify and remove the perturbation components in the data. The perturbation information recognition network usually contains a complex deep learning structure, such as the first deep network layer composed of multiple residual components, which work together through convolution, normalization and activation operations to extract and eliminate noise in the data.

[0073] During the anti-bending test of bone tissue, due to the influence of various factors such as experimental equipment and environmental conditions, the collected test data often contains a certain degree of disturbance. These disturbances may appear as abnormal fluctuations in the loading force curve, random jumps in displacement data, etc., which will interfere with the analysis of the true mechanical properties of bone tissue. In order to accurately evaluate the anti-bending properties of bone tissue, the computer system needs to first identify and quantify these disturbances. In step S310, the computer system performs a series of complex operations to generate the first disturbance data based on the disturbance data of the i-th level, the reconstructed test data, and the original target bone tissue anti-bending test data. This process may involve a variety of mathematical and statistical methods, such as differentiation, correlation analysis, filtering, etc., to extract disturbance features in the data.

[0074] For example, suppose the original test data is a time series array that records the change of loading force over time. The computer system first calculates the difference between adjacent data points to reveal the trend of the change of loading force. Then, it compares the difference between the difference results and the expected trend (such as theoretical predictions based on the mechanical properties of the material) to identify abnormal fluctuations. At the same time, the system also uses correlation analysis to compare the reconstructed test data with the original data to determine which time periods have significant data differences. Finally, the system applies a low-pass filter to smooth the raw data, filter out high-frequency noise, and use the difference between the data before and after filtering as the first disturbance data.

[0075] In step S310, the computer system has obtained the first disturbance data of the target bone tissue anti-bending test data. However, these data may still contain certain redundant information and noise, and need to be further processed to obtain a purer disturbance estimate.

[0076] The first deep network layer can be composed of multiple residual components, each of which contains a convolutional layer, a normalization layer, and an activation layer. This structure enables the network to automatically learn complex feature representations of the data and retain key parts of the input information through residual connections, thereby improving the performance and stability of the network.

[0077] The convolution layer applies multiple convolution kernels to the input data (i.e., the first perturbation data) in a sliding window manner to extract local features. The normalization layer accelerates the network training process and improves the generalization ability of the model by standardizing the input features. The activation layer (such as the ReLU layer) increases the expressive power of the network by introducing nonlinear factors. The residual connection allows the network to skip certain layers and directly pass information to subsequent layers, which helps alleviate the problem of gradient disappearance or explosion in deep networks.

[0078] For example, in step S320, the computer system passes the first disturbance data as input to the first deep network layer. The network layer first extracts disturbance features through the convolution layer, such as the spatial distribution of random noise, the frequency characteristics of periodic disturbances, etc. Then, the normalization layer normalizes the features to reduce the impact of data distribution differences. Next, the activation layer enhances the network's expressive power and distinguishes different types of disturbance components through nonlinear transformations. Finally, through the iterative processing of multiple residual components and the auxiliary role of residual connections, the system gradually removes the disturbance components in the data and obtains the disturbance data estimate of the jth level.

[0079] As an implementation manner, the step S310, performing disturbance information calculation according to the disturbance data of the i-th level, the reconstructed test data of the i-th level and the target bone tissue anti-bending test data to obtain first disturbance data of the target bone tissue anti-bending test data, may specifically include:

[0080] Step S311: multiplying the difference between the target bone tissue anti-bending test data and the reconstructed test data of the i-th level by the first model parameter update rate to obtain a first disturbance value;

[0081] Step S312: performing product of a first variable and the disturbance data of the i-th level to obtain a second disturbance value, wherein the first variable is a difference between a set value and an update rate of the first model parameter;

[0082] Step S313: taking the sum of the first disturbance value and the second disturbance value as the first disturbance data.

[0083] In the bending resistance test of bone tissue, due to the influence of various factors such as experimental conditions, equipment accuracy and external environment, the collected test data often contains a certain degree of disturbance. These disturbances may appear as small fluctuations in the loading force curve, random deviations in displacement data, etc., which will interfere with the analysis of the true mechanical properties of bone tissue. In order to accurately identify these disturbances, the computer system first needs to calculate the difference between the original test data and the reconstructed test data, and generate a preliminary estimate of the first disturbance value based on this difference.

[0084] In step S311, the computer system performs the following operations:

[0085] First, the system calculates the difference between the target bone tissue bending resistance test data and the reconstructed test data of the i-th level. This difference reflects the part of the reconstructed data that was not fully captured when restoring the original data, that is, the part that may contain disturbance information. Then, the system multiplies this difference with the first model parameter update rate (a preset value). The first model parameter update rate is a parameter used to adjust the sensitivity of the disturbance value calculation, which is set according to the needs of the actual application scenario. By multiplying by this rate, the system can amplify or reduce the difference to more accurately reflect the size and importance of the disturbance.

[0086] For example, suppose the load force of the target bone tissue bending resistance test data at a certain moment is 100 N, and the load force of the reconstructed test data at the i-th level at the corresponding moment is 98 N. The difference between the two is 2 N. If the update rate of the first model parameter is set to 0.5, the first disturbance value is calculated as follows:

[0087] The first perturbation value = (100 − 98) × 0.5 = 1N;

[0088] This calculation result shows that at this moment, the difference between the original test data and the reconstructed test data (i.e., the possible disturbance) is quantified as a loading force deviation of 1N.

[0089] In step S311, the computer system has calculated the first disturbance value based on the difference between the original test data and the reconstructed test data. However, this value may not be sufficient to fully reflect the disturbance in the data. In order to obtain more complete disturbance information, the system also needs to consider the disturbance data of the i-th level itself and incorporate it into the calculation of the disturbance value through certain operation rules.

[0090] In step S312, the computer system performs the following operations:

[0091] First, the system defines a first variable, whose value is the difference between the set value (such as 1) and the update rate of the first model parameter. This difference is actually an adjustment factor for the update rate of the first model parameter, which is used to introduce a "balance" mechanism in subsequent calculations to prevent a single parameter from having too much influence on the final result. Then, the system performs a product operation on this first variable and the disturbance data of the i-th level. The purpose of this step is to merge the magnitude of the existing disturbance data with the first disturbance value obtained by the difference calculation to form a more comprehensive estimate of the disturbance information.

[0092] For example, continuing the above example, assume that the first variable is set to 1-0.5=0.5 (where 0.5 is the update rate of the first model parameter), and the loading force deviation of the disturbance data of the i-th level at this moment is 0.8 N. Then the second disturbance value is calculated as follows: second disturbance value = 0.5 × 0.8 = 0.4 N; this calculation result shows that the disturbance data of the i-th level itself also contributes a certain amount of disturbance (0.4 N), which will be combined with the first disturbance value to form the final estimate of the first disturbance data.

[0093] After the calculations in steps S311 and S312, the computer system has obtained the first disturbance value and the second disturbance value, respectively. These two values ​​reflect the difference between the original test data and the reconstructed test data and the information of the i-th level disturbance data itself. In order to obtain a more accurate disturbance estimate, the system needs to merge these two values.

[0094] In step S313, the computer system simply adds the first disturbance value and the second disturbance value to obtain the first disturbance data. This addition operation is based on a basic assumption: the disturbance component in the data is the result of the combined effect of multiple factors, so the overall disturbance can be approximated by a linear combination. For example, continuing the previous example, the first disturbance value is 1N and the second disturbance value is 0.4N. Adding these two values ​​​​gets the final estimate of the first disturbance data: first disturbance data = 1N + 0.4N = 1.4N; this result combines the difference between the original test data and the reconstructed test data and the information of the i-th level disturbance data itself, providing a more accurate disturbance estimate for the subsequent disturbance removal step.

[0095] As an implementation manner, the data reconstruction network includes a second deep network layer, and the step S400, based on the data reconstruction network in the branch model of the j-th level, performs a data reconstruction operation on the pressure response characterization vector of the j-th level, the disturbance data of the j-th level, and the reconstructed test data of the i-th level to obtain the reconstructed test data of the target bone tissue anti-bending test data at the j-th level, including:

[0096] Step S410: reconstructing data according to the disturbance data of the j-th level and the reconstructed test data of the i-th level to obtain first reconstructed test data of the target bone tissue anti-bending test data;

[0097] Step S420: integrating the representation vectors according to the first reconstructed test data and the pressure response representation vector of the j-th level to obtain an integrated representation vector of the target bone tissue anti-bending test data;

[0098] Step S430: performing a data reconstruction operation on the integrated representation vector based on the second deep network layer to obtain the reconstructed test data of the j-th level.

[0099] In the jth level of the target data perturbation cleaning model, the data reconstruction network plays a vital role. It performs a series of refined data reconstruction operations based on the pressure response representation vector of the current level, the perturbation data, and the reconstructed test data of the previous level to generate a purer and more accurate reconstructed version of the target bone tissue anti-bending test data. This process requires not only a powerful data processing capability of the computer system, but also precise control of the weight and contribution of different input data in the reconstruction process.

[0100] During the bone tissue bending test, the collected test data often contains disturbance components due to the influence of various factors such as experimental conditions and equipment accuracy. Although the system has identified and cleared the disturbance to a certain extent through the disturbance information recognition network in the previous steps, the remaining disturbance may still affect the data analysis results. Therefore, in the initial stage of data reconstruction, the computer system needs to comprehensively consider the disturbance data at the current level and the reconstructed test data at the previous level to generate preliminary reconstructed test data.

[0101] In step S410, the computer system performs the following operations: First, the system obtains the disturbance data of the jth level. These data reflect the current level's estimate of the disturbance components in the original data. The system determines the mode and degree of its influence in the reconstruction process by analyzing the characteristics of the disturbance data (such as amplitude, frequency, etc.). Next, the system uses the reconstructed test data of the i-th level as an important reference. This data set has been preliminarily cleaned and reconstructed, reducing some disturbance components and retaining the key features of the original data as much as possible. The system determines which parts need to be further optimized by comparing and analyzing the differences between the reconstructed test data of the previous level and the original data. Based on a comprehensive consideration of the disturbance data and the reconstructed test data of the previous level, the system performs a preliminary reconstruction operation of the data. This operation may involve a variety of mathematical and statistical methods, such as interpolation, filtering, smoothing, etc., to generate the first reconstructed test data of the target bone tissue anti-bending test data. The goal of the preliminary reconstruction is to reduce the influence of the remaining disturbance components while keeping the overall structure and key features of the data unchanged. For example, suppose that in the jth level, the disturbance data appears as a small fluctuation on the loading force curve, while the reconstructed test data of the previous level has cleaned up most of the disturbances. In step S410, the computer system first analyzes the frequency and amplitude characteristics of the disturbance data and finds that it is mainly concentrated in the high-frequency band. Then, the system uses a low-pass filter to process the reconstructed test data of the previous level to filter out the high-frequency noise components (i.e., the main part of the disturbance data). At the same time, in order to maintain the continuity and smoothness of the data, the system uses the cubic spline interpolation method to interpolate the filtered data. The first reconstructed test data finally generated significantly reduces the disturbance components while retaining the key features of the original data.

[0102] In step S410, the computer system has generated preliminary reconstructed test data. However, these data may still lack certain key features or information details, which may affect the accuracy of subsequent analysis. In order to make up for this deficiency, the system needs to integrate the first reconstructed test data with the pressure response characterization vector of the current level to generate an integrated characterization vector containing more complete information.

[0103] In step S420, the computer system performs the following operations: first, it is determined that the pressure response characterization vector of the jth level is a set of key features automatically extracted from the original data through the characterization information extraction network. These features not only reflect the dynamic response characteristics of bone tissue in the bending resistance test, but also contain important information required to identify disturbances and reconstruct data. In order to effectively fuse the first reconstructed test data with the pressure response characterization vector, the system needs to design a reasonable feature fusion strategy. This strategy may involve a variety of methods such as feature splicing, weighted summation, feature mapping, etc. The specific choice depends on the characteristics of the data and the analysis requirements. After determining the feature fusion strategy, the system integrates the first reconstructed test data with the pressure response characterization vector. This process may involve multiple mathematical transformations and matrix operation steps to ensure that the integrated vector retains the structural information of the original data and incorporates the key feature representation.

[0104] For example, assume that the pressure response characterization vector is a vector set containing multiple key mechanical parameters (such as elastic modulus, yield strength, etc.), and the first reconstructed test data is a time series array. In step S420, the computer system first maps each parameter value in the characterization vector to the time axis of the first reconstructed test data to form a new feature channel. Then, these extended features are spliced ​​with the original time series data in the channel dimension to generate an integrated characterization vector. For example, if the characterization vector contains two parameters, elastic modulus and yield strength, the integrated characterization vector will have two more channels than the original time series data, corresponding to the extended features of these two parameters.

[0105] After processing steps S410 and S420, the computer system has obtained an integrated representation vector containing preliminary reconstructed data and key feature information. However, this vector may still have some noise or slight disturbances that need to be further cleaned and optimized to obtain a higher quality data reconstruction version. Therefore, step S430 introduces a second deep network layer to perform a data fine reconstruction operation.

[0106] In step S430, the second deep network layer is, for example, a complex neural network structure including multiple hidden layers, such as a convolutional neural network (CNN), a recurrent neural network (RNN), or a variant thereof (such as a long short-term memory network LSTM, a gated recurrent unit GRU, etc.). These network layers gradually optimize the data distribution and feature representation in the integrated representation vector through layer-by-layer abstraction and feature transformation, and finally generate the reconstructed test data of the jth level. The design of the second deep network layer needs to fully consider the characteristics of the data and the reconstruction target. For example, for time series data (such as loading force curves, displacement curves, etc.), CNN can be used to extract local features, while RNN or LSTM is more suitable for capturing long-term dependencies. In addition, in order to enhance the generalization ability and robustness of the network, residual connections, batch normalization, dropout and other strategies may also be introduced. After the integrated representation vector is input into the second deep network layer, the network gradually optimizes the data distribution and feature representation through layer-by-layer processing. This process may involve multiple forward propagation and back propagation iteration steps to minimize the reconstruction error (such as mean square error MSE, mean absolute error MAE, etc.). As the number of iterations increases, the network gradually learns how to recover reconstructed test data that is closer to the actual mechanical response from the integrated representation vector. Finally, the second deep network layer outputs the reconstructed test data of the jth level. These data significantly reduce the disturbance components while retaining the key features of the original test data, improving the accuracy and reliability of the data. In order to evaluate the reconstruction effect, the computer system can also compare the reconstructed test data with the unperturbed data (or gold standard data) to calculate performance indicators such as reconstruction error to guide subsequent data processing and analysis.

[0107] For example, assume that the second deep network layer is a hybrid neural network structure composed of multiple convolutional layers and LSTM layers. In step S430, the integrated representation vector is first input into the network for processing. The convolutional layer first extracts local features such as the waveform characteristics of the loading force curve, the changing trend of the displacement curve, etc.; then the LSTM layer captures the long-term dependencies of these features in the time series; finally, the fully connected layer maps the features back to the spatial dimensions of the reconstructed test data to generate the final reconstructed version. Throughout the process, the system uses MSE as the evaluation indicator of the reconstruction error and continuously optimizes the network parameters through the back propagation algorithm to minimize the reconstruction error. After multiple iterations, the system outputs the reconstructed test data of the jth level. These data significantly reduce the disturbance components while retaining the key features of the original data, providing more accurate and reliable data support for subsequent mechanical property evaluation.

[0108] As an implementation manner, the step S410, performing data reconstruction according to the disturbance data of the j-th level and the reconstructed test data of the i-th level to obtain the first reconstructed test data of the target bone tissue anti-bending test data, may specifically include:

[0109] Step S411: determining a difference result between the target bone tissue anti-bending test data and the j-th level disturbance data, and determining a multiplication result between the shielding test data corresponding to the target bone tissue anti-bending test data and the second model parameter update rate;

[0110] Step S412: performing inner product calculation on the difference result and the multiplication result to obtain first reconstructed data;

[0111] Step S413: multiplying the second variable and the reconstructed test data of the i-th level to obtain second reconstructed data, wherein the second variable is a difference result between a set value and the multiplication result;

[0112] Step S414: taking the sum of the first reconstructed data and the second reconstructed data as the first reconstructed test data.

[0113] In the analysis of bone tissue bending test data, the original test data often contains disturbance components, which may come from experimental equipment, environmental conditions or random errors in the data acquisition process. In order to separate the disturbance components from the original data, the computer system first needs to quantify these disturbances in some way. At the same time, in order to reconstruct the data more accurately, the system also needs to consider specific information that may exist in the original data (such as shielded test data), which can help the system identify which parts of the data are more susceptible to disturbances.

[0114] In step S411, the computer system performs the following operations: First, the system calculates the difference between the target bone tissue bending resistance test data and the j-th level disturbance data. This difference reflects the change in the original data under the influence of the disturbance, that is, the actual impact of the disturbance on the data.

[0115] Difference = target bone tissue bending resistance test data − disturbance data of the jth level.

[0116] Next, the system obtains the shielded test data (i.e., mask data) corresponding to the target bone tissue bending resistance test data. The shielded test data is usually an array or matrix with the same shape as the original data, in which the value of each element indicates whether the data at the corresponding position may be severely affected by the disturbance (for example, 1 indicates that it may be severely affected, and 0 indicates that the impact is small). Then, the system multiplies the shielded test data with a second model parameter update rate (a preset value). This multiplication result is used in subsequent calculations to adjust the weights of different data components in the reconstruction process.

[0117] Multiplication result = shielded test data × second model parameter update rate

[0118] For example, suppose the load force of the target bone tissue bending test data at a certain time point is 100N, and the load force deviation of the j-th level perturbation data at this time point is -2N. Then the result of the difference calculation is: difference = 100N-(-2N) = 102N; at the same time, suppose the value of the shielded test data at this time point is 0.8 (indicating that the data at this position may be affected by a certain degree of perturbation), and the update rate of the second model parameter is 0.5. Then the result of the multiplication calculation is: multiplication result = 0.8×0.5=0.4; this multiplication result will be used in the subsequent steps to adjust the weights of different data components in the reconstruction process.

[0119] In step S411, the computer system has calculated the difference and the multiplication result. Next, the system needs to integrate these results in some way to generate a preliminary estimate of the first reconstructed data. Inner product solution is a commonly used data integration method, which can generate a new scalar value or vector by calculating the dot product of two vectors. In step S412, the system uses inner product solution to generate a preliminary estimate of the first reconstructed data.

[0120] In step S412, the computer system performs the following operations: the system first solves the inner product of the difference calculated in step S411 and the multiplication result. However, it should be noted that the inner product solution here is not a vector dot product operation in the traditional sense, because the difference is, for example, a multidimensional array (such as time series data), and the multiplication result is a scalar value. In practical applications, the system may use the multiplication result as a weight factor to perform weighted summation on the difference (if the difference is multidimensional), or expand the multiplication result to the same shape as the difference before performing element-wise multiplication (i.e., Hadamard product).

[0121] The first preliminary estimate of the reconstructed data = difference × multiplication result;

[0122] In order to be closer to the original meaning of the inner product solution, we can assume that the difference is converted into a vector form (if it is not a vector itself) and introduce a full-1 vector with the same shape as the difference as "another vector" to perform the inner product operation (although this is not common in practice and is a bit far-fetched). But a more reasonable explanation is that the system uses a simplified weighting strategy to approximate the effect of the inner product solution.

[0123] For simplicity of explanation, it is assumed that the system actually directly applies the multiplication result as a weight factor to the difference (this is, for example, an approximate weighted summation method) to generate a preliminary estimate of the first reconstructed data. That is:

[0124] The first preliminary estimate of the reconstructed data (simplified representation) ≈ the difference (after some form of weighting)

[0125] The specific method of weighted processing may depend on the multiplication result and other factors, but it will not be further elaborated here. In step S412, the computer system has generated a preliminary estimate of the first reconstructed data. However, this estimate may not be accurate enough because it only considers the impact of the disturbance data on the original data and does not fully utilize the useful information contained in the reconstructed test data of the previous level. In order to further improve the accuracy of the reconstructed data, the system needs to make further adjustments and optimizations based on the preliminary estimate of the first reconstructed data combined with the reconstructed test data of the previous level.

[0126] In step S413, the computer system performs the following operations: the system first defines a second variable, which is the difference between a certain set value (such as 1) and the multiplication result calculated in step S411. Then the system performs a product operation on the second variable and the reconstructed test data of the i-th level to generate the second reconstructed data. The purpose of this step is to adjust and optimize the preliminary estimate of the first reconstructed data according to the reconstruction result of the previous level. Second variable = set value - multiplication result; second reconstructed data = second variable × reconstructed test data of the i-th level.

[0127] For example, continuing the previous example, assume that the set value is 1 and the multiplication result is 0.4 (as calculated in step S411). Then the second variable is calculated as: second variable = 1-0.4 = 0.6; assume that the loading force of the reconstructed test data of the i-th level at this time point is 99N (it has been preliminarily cleaned and reconstructed but may still contain certain errors). Then the second reconstructed data is calculated as: second reconstructed data = 0.6×99N = 59.4N; this second reconstructed data reflects the contribution of the reconstructed test data of the previous level to the preliminary estimate of the first reconstructed data after certain adjustments.

[0128] After processing steps S412 and S413, the computer system has generated a preliminary estimate of the first reconstruction data and a second reconstruction data, respectively. The two data represent useful information extracted from the disturbance data and the previous level reconstruction test data, respectively. In order to generate the final first reconstruction test data, the system needs to integrate the two data.

[0129] In step S414, the computer system performs the following operations: the system adds the first reconstructed data preliminary estimate generated in step S412 and the second reconstructed data generated in step S413 to generate first reconstructed test data. This step actually fuses the information extracted from the disturbance data with the information extracted from the previous level of reconstructed test data to generate a more accurate and reliable reconstructed data set.

[0130] First reconstructed test data=first reconstructed data preliminary estimation+second reconstructed data.

[0131] As an embodiment, the step S420, performs characterization vector integration based on the first reconstructed test data and the pressure response characterization vector of the j-th level to obtain an integrated characterization vector of the target bone tissue anti-bending test data, including: merging the first reconstructed test data and the pressure response characterization vector of the j-th level in the channel dimension to obtain the integrated characterization vector.

[0132] During the anti-bending test of bone tissue, the collected original test data contains the dynamic response information of bone tissue during the stress process. However, these data are often affected by a variety of disturbance factors, resulting in deviations in the analysis results. In order to more accurately evaluate the mechanical properties of bone tissue, the computer system processes the data layer by layer through a multi-level branching model, gradually reducing disturbances and reconstructing the test data. In step S420, the system has obtained the first reconstructed test data and the j-th level pressure response characterization vector. The former is a data set that has been preliminarily cleaned and reconstructed, and the latter is a set of key features automatically extracted from the original data.

[0133] The core task of step S420 is to integrate the first reconstructed test data with the j-th level pressure response representation vector to generate an integrated representation vector containing more dimensional information. This process is mainly achieved by merging data in the channel dimension.

[0134] In step S420, the computer system performs the following operations to merge the data: First, ensure that the first reconstructed test data and the j-th level pressure response characterization vector are prepared before merging. The first reconstructed test data is, for example, a multidimensional array (such as time series data) containing test parameters (such as load force, displacement, etc.) that have been preliminarily cleaned and reconstructed. The pressure response characterization vector is a vector or low-dimensional array containing key mechanical parameters (such as elastic modulus, yield strength, etc.). In order to merge the two, the system needs to expand the channel dimension of one of the data sets (usually the pressure response characterization vector). This means copying or expanding the characterization vector to the same number of channels as the first reconstructed test data (even if the characterization vector itself has only one "channel", it needs to be treated as part of a data set with multiple channels in some way). However, in this particular scenario, since the characterization vector is usually of low dimensionality and contains highly generalized feature information, it is more common to add it as a new independent channel to the first reconstructed test data. After the channel dimension is expanded, the system simply stacks or splices each channel of the first reconstructed test data with the characterization vector (now as a new channel) on the same time step or sample. Thus, each time step or sample now contains an additional feature dimension - the value of the pressure response representation vector.

[0135] For example, assume that the first reconstructed test data is a three-dimensional array of shape [number of samples, time steps, number of features], where the number of features includes two dimensions: load and displacement. The pressure response representation vector of the jth level is a one-dimensional array containing two key parameters: elastic modulus and yield strength. To merge these data, the system first treats the representation vector as a new channel with two elements (corresponding to the two parameters) and expands it to the same number of samples and time steps as the first reconstructed test data (although in fact this expansion is conceptual because the value of the representation vector is the same for all samples and time steps). Then, the system adds this new channel to the channel dimension of the first reconstructed test data to generate an integrated representation vector of shape [number of samples, time steps, number of features + 1]. Now, each time step of the integrated representation vector contains information in four dimensions: load, displacement, elastic modulus, and yield strength.

[0136] Through the implementation of step S420, the computer system successfully integrates the first reconstructed test data with the j-th level pressure response characterization vector to generate an integrated characterization vector containing richer information. This process not only retains the key dynamic response information of the original test data, but also incorporates the key mechanical features automatically extracted from the data, providing a more comprehensive and accurate data basis for subsequent data analysis and mechanical performance evaluation.

[0137] As an implementation mode, before performing disturbance removal on the target bone tissue anti-bending test data based on the target data disturbance removal model, the method may further include: preprocessing the target bone tissue anti-bending test data. Then, performing disturbance removal on the target bone tissue anti-bending test data based on the target data disturbance removal model includes: performing disturbance removal on the preprocessed target bone tissue anti-bending test data based on the target data disturbance removal model.

[0138] The preprocessing step aims to reduce the noise and outliers in the data, and adjust the format and range of the data to make it more suitable for subsequent machine learning model processing. For example, during preprocessing, the computer system can clean the target bone tissue bending test data, with the aim of removing or correcting errors, missing values ​​and outliers in the data. For example, if the data at certain time points are lost due to equipment failure during the test, the system may use interpolation methods (such as linear interpolation, Lagrange interpolation, etc.) to fill in these missing values. In addition, if obvious outliers are detected (such as a sudden surge in loading force to an unreasonable level), the system may mark these values ​​as invalid and replace them with the mean or median within a reasonable range. Next, the computer system can normalize the cleaned data. Since data from different test samples or under different test conditions may have different dimensions and ranges, directly inputting these data into the machine learning model may cause the model to be difficult to converge or the training effect is poor. Therefore, the system will normalize or standardize the data, scaling all features to the same numerical range (such as [0, 1] or [-1, 1]) or adjusting them to have the same variance. In the preprocessing stage, the computer system can also perform feature selection and engineering operations. This step aims to extract the most useful features for subsequent analysis from the raw data, and may generate new and more representative features through feature transformation. For example, the system may calculate the ratio between loading force and displacement as a new feature to reflect the stiffness characteristics of the material.

[0139] After completing the above preprocessing steps, the computer system will perform disturbance removal on the preprocessed target bone tissue bending resistance test data based on the target data disturbance removal model. This step uses deep learning models or traditional signal processing techniques to identify and separate the disturbance components in the data, thereby restoring a data set that is closer to the actual mechanical response. During the disturbance removal process, the system can input the preprocessed data into the target data disturbance removal model. The model may contain multiple levels of branching structures, each of which is responsible for further reducing the disturbance in the data and optimizing the quality of the reconstructed data. At each level, the system extracts key features through the representation information extraction network, identifies the disturbance components through the disturbance information identification network, and finally generates new reconstructed test data through the data reconstruction network. These steps are iterated until the preset number of levels is reached or specific convergence conditions are met.

[0140] As an implementation manner, the method further includes a debugging step of the target data disturbance cleaning model, specifically including:

[0141] Step S10: acquiring bone tissue anti-bending test sample data, wherein the bone tissue anti-bending test sample data includes disturbance data;

[0142] Step S20: extracting the characterization information of the bone tissue anti-bending test sample data based on the characterization information extraction network in the j-th level branch model, and obtaining the pressure response characterization vector of the bone tissue anti-bending test sample data at the j-th level;

[0143] Step S30: Based on the disturbance information identification network in the j-th level branch model, the disturbance data of the bone tissue anti-bending test sample data at the i-th level output by the i-th level branch model and the reconstructed test data at the i-th level are cleared to obtain the disturbance data of the bone tissue anti-bending test sample data at the j-th level;

[0144] Step S40: Based on the data reconstruction network in the branch model of the j-th level, a data reconstruction operation is performed on the pressure response characterization vector of the j-th level, the disturbance data of the j-th level and the reconstructed test data of the i-th level to obtain the reconstructed test data of the j-th level of the bone tissue anti-bending test sample data;

[0145] Step S50: Determine the disturbance data of the j-th level and the reconstructed test data of the j-th level as the input information of the branch model of the j+1-th level, iterate one by one, and obtain the reconstructed test data of the bone tissue anti-bending test sample data at each level in the I levels;

[0146] Step S60: performing error determination according to the reconstructed test data of the bone tissue anti-bending test sample data at each level and the unperturbed data corresponding to the bone tissue anti-bending test sample data, to obtain the error of the target data disturbance cleaning model;

[0147] Step S70: updating the model parameters of the target data disturbance cleaning model according to the error;

[0148] The representation information extraction network includes a pre-debugged representation information extraction component, and the target data disturbance cleaning model does not debug the representation information extraction component during the debugging phase.

[0149] At the beginning of the training process, the computer system needs to obtain a large amount of bone tissue bending test sample data. These data should include various parameters collected during the actual test, such as load force, displacement, timestamp, etc., and these data should naturally include various disturbance factors, such as equipment vibration, ambient temperature fluctuations, electromagnetic interference, etc. These disturbance data are an indispensable part of the model training process because they will help the model learn how to accurately identify and remove disturbances in complex environments. For example, the training sample data may be stored in a format such as CSV or HDF5, and each sample contains multiple fields, such as "SampleID" (sample number), "Time" (timestamp), "Load" (load force), "Displacement" (displacement), etc.

[0150] After obtaining the training sample data, the computer system will process the data based on the representation information extraction network in the branch model of the jth level to extract the key pressure response representation vector. This step is a key link in the model training process because it directly affects the effect of subsequent disturbance identification and data reconstruction. The representation information extraction network is, for example, a deep neural network containing multiple convolutional layers, pooling layers, and activation functions. These network layers extract high-level feature representations from the raw data in a layer-by-layer abstraction manner. In the scenario of bone tissue bending test data analysis, these features may include the changing trend of the loading force, the specific shape of the displacement curve, etc.

[0151] Since the representation information extraction network may have been pre-debugged (i.e. pre-trained) before training, its internal parameters are usually not adjusted significantly when training the target data perturbation cleaning model. The benefit of this is that it can speed up training and avoid the risk of overfitting.

[0152] For example, assume that the representation information extraction network receives the sample data obtained in step S10 as input. The network first extracts local features of the input data through multiple convolutional layers. Each convolutional layer contains multiple learnable convolution kernels that slide on the input data and calculate the convolution results to generate feature maps. Then, the pooling layer downsamples the feature map to reduce the amount of calculation and avoid overfitting. Finally, an activation function (such as ReLU) is applied to the output of the convolutional layer to introduce nonlinear factors and enhance the expression ability of the network.

[0153] After multiple layers of convolution, pooling, and activation operations, the representation information extraction network outputs a high-dimensional vector as the pressure response representation vector. This vector contains the key feature information automatically extracted from the original data, providing an important basis for disturbance identification and data reconstruction in subsequent steps.

[0154] After obtaining the pressure response characterization vector, the computer system will use the disturbance information identification network in the branch model of the jth level to further process the disturbance data and reconstructed test data output by the branch model of the i-th level to identify the disturbance components in the current level.

[0155] The disturbance information recognition network is, for example, a complex neural network containing multiple fully connected layers, attention mechanisms, and residual connections. These network layers extract disturbance features from the input data through nonlinear transformation and feature fusion, and combine them with the pressure response representation vector to generate a more accurate disturbance estimate.

[0156] During the training process, the perturbation information recognition network learns how to distinguish the difference between the true mechanical response and the perturbation component. This usually involves learning from a large number of training samples with perturbations and continuously optimizing the network parameters to minimize the reconstruction error.

[0157] For example, suppose that in the i-th level, the branch model has output a set of perturbation data and reconstructed test data. These data are passed to the perturbation information recognition network of the j-th level for further processing. The network first performs a nonlinear transformation on the input data through a fully connected layer and extracts high-level features. Then the attention mechanism is used to dynamically adjust the importance weights of different features to highlight the perturbation features. Finally, the perturbation features are combined with the pressure response representation vector through a residual connection to generate a perturbation estimate for the current level. This process may involve multiple forward propagation and backpropagation iterations to optimize network parameters and reduce reconstruction errors. In each iteration, the system calculates the difference between the perturbation data and the true perturbation data at the current level (i.e., the perturbation error), and updates the network parameters based on the error until the convergence conditions are met.

[0158] After obtaining the disturbance estimate of the current level, the computer system will use the data reconstruction network in the branch model of the jth level to comprehensively process the pressure response characterization vector, disturbance data and the reconstructed test data of the previous level to generate new reconstructed test data.

[0159] The data reconstruction network is, for example, a neural network that includes a decoder structure and an upsampling operation. The decoder structure restores the low-dimensional feature map to the dimension and resolution of the original data through layer-by-layer upsampling and convolution operations. At the same time, the upsampling operation fills in the missing pixel values ​​through interpolation or learned upsampling kernels to ensure the integrity and continuity of the reconstructed data.

[0160] During the training process, the data reconstruction network learns how to recover the key features of the original data from the pressure response representation vector and the perturbation data and generate a reconstructed version that is as close to the real undisturbed data as possible. To achieve this goal, the network needs to continuously optimize its parameters to minimize the reconstruction error and improve the quality of the reconstructed data. For example, assume that in the jth level, the system has obtained the pressure response representation vector, the perturbation data, and the reconstructed test data of the previous level. These data are passed to the data reconstruction network for comprehensive processing. The network first uses the decoder structure to upsample and convolve the pressure response representation vector to generate a preliminary reconstructed data framework. The network then adds the perturbation data as a noise component to the framework to simulate the perturbation in the real test data. Finally, the network generates the final reconstructed test data by fine-tuning and optimizing the parameters.

[0161] During the training process, the system calculates the difference between the reconstructed test data and the real undisturbed data (i.e., reconstruction error) and updates the parameters of the data reconstruction network based on the error. Through multiple iterations of optimization, the network can gradually learn how to recover the real mechanical response from the perturbed data and generate high-quality reconstructed test data.

[0162] After obtaining the reconstructed test data of the current level, the computer system uses it as the input data of the next level to continue iterative processing. This process will be carried out one by one until all the branch models of the I level have been traversed. Through the layer-by-layer iterative processing system can gradually reduce the disturbance components in the data and improve the quality of the reconstructed data.

[0163] In each iteration, the system will pass the reconstructed test data and perturbation data of the current level to the branch model of the next level for processing. The branch model of the next level will repeat the operations in steps S20 to S40 to generate new reconstructed test data and perturbation estimates. This process will be repeated until the preset number of levels I is reached or a specific convergence condition is met.

[0164] Through layer-by-layer iterative processing, the system can gradually extract deeper and more abstract feature representations and generate higher quality reconstructed test data. At the same time, as the number of layers increases, the system's ability to identify and remove disturbance components will gradually increase, thereby further improving the accuracy and reliability of the reconstructed data. After completing all levels of iterative processing, the computer system determines the performance of the target data disturbance cleaning model based on the error between the reconstructed test data and the true unperturbed data. This process involves the calculation and analysis of the reconstruction error and the evaluation of the model's generalization ability.

[0165] In the error determination phase, the system first calculates the difference between the reconstructed test data and the true unperturbed data at each level (i.e., reconstruction error). These errors can be measured by a variety of metrics such as mean square error (MSE), mean absolute error (MAE), etc. The system then aggregates and analyzes these errors to evaluate the performance of the model at different levels.

[0166] In order to more comprehensively evaluate the model performance, the system can also introduce other evaluation indicators such as correlation coefficient (R^2), determination coefficient (R-squared), etc. to measure the similarity and goodness of fit between the reconstructed test data and the real data.

[0167] For example, suppose that at each level the system calculates the MSE error between the reconstructed test data and the true unperturbed data. The system then aggregates these errors and plots them into an error curve to visually observe the error trend as the level increases. By analyzing the error curve, the system can find that as the level increases, the MSE error gradually decreases, indicating that the model's ability to identify and remove disturbance components gradually increases.

[0168] As an implementation mode, the step S60, performing error determination according to the reconstructed test data of the bone tissue anti-bending test sample data at each level and the unperturbed data corresponding to the bone tissue anti-bending test sample data, to obtain the error of the target data disturbance cleaning model, includes:

[0169] Step S61: determining a first error between the reconstructed test data of each level and the undisturbed data in a shielded interval, wherein the shielded interval is a shielded interval in the shielded test data corresponding to the bone tissue anti-bending test sample data;

[0170] Step S62: determining a second error between the reconstructed test data of each level and the undisturbed data in the masking interval;

[0171] Step S63: determining a third error between the reconstructed test data of each level and the undisturbed data in the masking interval;

[0172] Step S64: performing error determination according to the first error, the second error and the third error to obtain the error of the target data disturbance cleaning model.

[0173] In step S61, the goal of the computer system is to determine the first error between the reconstructed test data and the unperturbed data at each level in a specific masked interval (i.e., the position of the mask). The masked interval here usually refers to those parts of the data that are considered to be particularly susceptible to perturbations, or specific time periods or data points that researchers are particularly concerned about. By calculating the error only in these intervals, the performance of the model when processing key data can be more accurately evaluated.

[0174] During the bending test of bone tissue, certain time periods or loading stages may be particularly sensitive to data quality, such as the moment when the material begins to yield or fracture. The data at these stages are crucial for evaluating the mechanical properties of the material, but they are also more susceptible to interference from external factors such as changes in experimental conditions and equipment vibration. Therefore, researchers may mark these critical time periods as shielded intervals to give special attention during model evaluation.

[0175] Before training begins, researchers generate masked test data based on prior knowledge or preliminary data analysis results. These data are usually in the form of binary masks, where 1 indicates that the corresponding data point is within the mask interval and needs special attention; 0 indicates that the data point is not within the mask interval.

[0176] For each level of reconstructed test data, the computer system first identifies the masked intervals corresponding to the unperturbed data. Then, the difference between the reconstructed test data and the unperturbed data is calculated on these intervals to obtain the first error. This difference can be measured by a variety of error metrics, such as mean square error (MSE), mean absolute error (MAE), etc.

[0177] For example, suppose that in the bone tissue bending test data, the researcher marked the shielding interval from 1 second before the load reaches the maximum value to 0.5 seconds after. During this time period, the material may be undergoing a transition from elastic deformation to plastic deformation, and data quality is particularly important. For a certain level of reconstructed test data, the computer system calculates the MSE error between the reconstructed load force and the undisturbed load force point by point in the shielding interval. For example, at the tth time point, if the reconstructed load force is Frecon,t and the undisturbed load force is Ftrue,t, the MSE error contribution at this time point is (Frecon,t−Ftrue,t)2. After summing and averaging the error contributions in all shielding intervals, the first error of this level can be obtained.

[0178] Step S62 further refines the error analysis process by introducing different error metrics to comprehensively evaluate the model performance. In this step, the computer system calculates the second error between the reconstructed test data and the undisturbed data on the masked interval, which may focus on reflecting the overall trend or structural differences of the data.

[0179] Unlike the first error, the calculation of the second error may use more complex error metrics, such as the structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), or domain-specific error functions. These criteria can capture nonlinear relationships or high-level feature differences between data, thereby providing a more comprehensive performance evaluation. Similar to the calculation process of the first error, the computer system first determines the data points in the masked interval and then applies the selected error metric to calculate the difference between the reconstructed test data and the unperturbed data. This difference may not only consider the degree of numerical proximity but also the similarity in data distribution, texture, structure, etc. For example, taking SSIM as an example, this indicator evaluates the similarity of two images (or time series data) by comparing local brightness, contrast and structural information. In the analysis of bone tissue bending test data, the computer system can treat the loading force curve or displacement curve as a one-dimensional image and calculate the SSIM value of the reconstructed test curve and the unperturbed curve in the masked interval. A lower SSIM value indicates that there is a large structural difference between the two curves, which may mean that the reconstruction performance of the model in this interval is poor.

[0180] Step S63 evaluates the accuracy of the model in disturbance identification by calculating the difference between the disturbance components identified in the reconstructed test data and the real disturbance components. This step helps to understand whether the model introduces new errors or misses some disturbances when removing disturbances.

[0181] The real disturbance component is usually not directly observable but can be indirectly estimated by comparing the unperturbed data with the original test data containing the disturbance. For example, a filtering or denoising algorithm can be used to try to remove the disturbance component from the original data to obtain an approximate estimate of the real disturbance component. After obtaining the estimate of the reconstructed disturbance component and the real disturbance component, the computer system calculates the difference between them as the third error. This difference may be measured by MSE, MAE or other suitable error metrics. For example, assume that the model outputs a set of reconstructed disturbance data at a certain level to represent the noise or interference that may exist in the original test data. In order to evaluate the accuracy of this set of data, the computer system first uses a denoising algorithm (such as wavelet threshold denoising) to try to remove the noise component from the original test data to obtain an approximate estimate of the real disturbance component. Then the MSE error between the reconstructed disturbance data and the approximate real disturbance data is calculated as the third error. This error value reflects the accuracy and robustness of the model in disturbance identification.

[0182] In step S64, the computer system comprehensively considers the first error, the second error, and the third error calculated previously to obtain a comprehensive evaluation of the performance of the target data disturbance cleaning model. This process may involve weighted summation of different types of errors or the use of other forms of fusion strategies. Since different types of errors may reflect different aspects of model performance (such as numerical accuracy, structural similarity, disturbance recognition ability, etc.), the computer system may assign different weights to each error according to actual needs. These weights can be optimized by cross-validation, grid search, etc. to ensure the accuracy and effectiveness of model evaluation. After obtaining the weighted error value, the computer system may use a simple summation operation or a more complex fusion algorithm to generate the final error index. This index will serve as the basis for model performance evaluation and guide subsequent model optimization work.

[0183] As an implementation manner, the step S61 of determining the first error between the reconstructed test data of each level and the undisturbed data in the masking interval may specifically include:

[0184] Step S611: for the i-th level in the I levels, determining the mean square error between the reconstructed test data of the i-th level and the undisturbed data, 1≤i≤I;

[0185] Step S612: performing inner product calculation on the mean square error corresponding to the i-th level and the shielded test data corresponding to the bone tissue anti-bending test sample data to obtain a first multiplication result;

[0186] Step S613: Multiply the first multiplication result and the compromise model parameter of the i-th level to obtain a first error corresponding to the i-th level.

[0187] During the bending test of bone tissue, the collected test data often contains a certain degree of disturbance due to the influence of various factors such as experimental conditions and equipment accuracy. In order to accurately evaluate the mechanical properties of the material, these disturbances need to be cleaned up and reconstructed test data that is as close to the actual mechanical response as possible needs to be generated. During the training process of the target data perturbation cleaning model, calculating the difference between the reconstructed test data and the unperturbed data is one of the key steps in evaluating the model performance.

[0188] The core task of step S611 is to calculate the mean square error (MSE) between the reconstructed test data and the undisturbed data at the i-th level. MSE is a commonly used error metric that reflects the average value of the sum of squares of the differences between a data point and its true value. In the analysis of bone tissue bending test data, MSE can intuitively show how close the reconstructed test data is to the undisturbed data in terms of numerical value. The specific calculation process is as follows:

[0189] Assume that the reconstructed test data of the i-th level is X recon,i , the unperturbed data is X true , and both are time series data with a length of N. Then the MSE calculation formula is:

[0190] ;

[0191] Among them, x recon,i,t represents the value of the reconstructed test data at the i-th level at time point t, x true,t represents the value of the unperturbed data at time point t.

[0192] In step S611, the MSE values ​​between the reconstructed test data and the undisturbed data at each level have been obtained. However, these MSE values ​​reflect the average error over the entire time series and may not fully capture the subtle changes in data quality within the masked interval. In order to more accurately evaluate the performance of the model at key data points, it is necessary to further analyze the error distribution on the masked interval.

[0193] Step S612 introduces shielded test data (i.e., mask data) to guide error analysis. Shielded test data is usually a binary array or matrix with the same shape as the original test data, where 1 indicates that the corresponding data point is within the shielded interval (i.e., the key data point), and 0 indicates that the data point is not within the shielded interval. By solving the inner product with the shielded test data, the MSE error can be concentrated on the shielded interval for evaluation. The specific calculation process is as follows: Assume that the shielded test data is M, where m t =1 means that the time point t is within the shielding interval, m t =0 means that the time point t is not in the shielding interval. Then the inner product solution process can be expressed as:

[0194] Masked_MSE i =MSE i M;

[0195] Continuing with the previous example, assume that the masked test data marks the interval from time point T1 to T2 as the masked interval. Recalculate the MSE value in this interval (considering only the data points in the masked interval):

[0196] ;

[0197] Where |Mask Interval| represents the length of the mask interval (i.e., the number of time points). However, please note that this formula is not actually the result of solving the inner product with the masked test data, but a simplified "masked interval MSE" calculation formula. In order to be consistent with the subsequent steps, it can be regarded as a special case of the first multiplication result, that is, all errors within the mask interval are taken into account and errors outside the mask interval are ignored.

[0198] In actual operation, the computer system can select data points in the masked interval through logical indexing or masking operations and directly apply the MSE calculation formula to these data points to obtain the weighted MSE value (ie, the first multiplication result).

[0199] In step S612, a weighted MSE value has been obtained, which reflects the difference between the reconstructed test data and the undisturbed data in the masked interval. However, in order to more comprehensively evaluate the model performance and take into account the comparability of errors between different levels, it may be necessary to further adjust this error value. Step S613 achieves this goal by introducing a compromise model parameter.

[0200] The compromise model parameter is a coefficient used to adjust the error value. It can be set according to actual needs to balance the weights of errors between different levels or take into account other factors that affect error evaluation. In step S613, the computer system performs a product operation on the compromise model parameter and the first multiplication result (i.e., the "weighted" MSE value) to obtain the final first error value. The specific calculation process is as follows: Assuming that the compromise model parameter of the i-th level is λi, the calculation formula for the first error is:

[0201] First Errori=λi·Masked_MSEi;

[0202] By adjusting the value of λi, the computer system can flexibly control the weight of errors at different levels in the overall error assessment. For example, if the data quality of a certain level is particularly important or the reconstruction process of this level is particularly complex, then a larger λi value can be set to highlight the error situation at this level. For example, continuing the previous example, assume that the compromise model parameter λi of the i-th level is 1.2 (this is just an example value and the actual value may depend on the specific situation). Then the calculation process of the first error is as follows: First Errori=1.2·Masked_MSEi;

[0203] Wherein Masked_MSEi is the weighted MSE value calculated in step S612. By multiplying this value with the compromise model parameter, the computer system obtains the final first error value, which will serve as an important basis for evaluating the quality of the i-th level reconstructed test data.

[0204] In summary, step S61 successfully determines the first error between the reconstructed test data and the undisturbed data at each level in the shielded interval by calculating the mean square error, solving the inner product with the shielded test data, and performing the product with the compromise model parameters. This process not only takes into account the key parts of the data, but also ensures the accuracy and effectiveness of the evaluation results through reasonable error metrics and calculation methods. In the training process of the target data disturbance cleaning model, these first error values ​​will serve as important feedback signals to guide the optimization of model parameters and the improvement of performance.

[0205] As an implementation manner, the step S62 of determining the second error between the reconstructed test data of each level and the undisturbed data in the masking interval may specifically include:

[0206] Step S621: for the i-th level among the I levels, determining the first absolute error between the reconstructed test data of the i-th level and the undisturbed data, 1≤i≤I;

[0207] Step S622: performing inner product calculation on the first absolute error corresponding to the i-th level and the shielded test data corresponding to the bone tissue anti-bending test sample data to obtain a second multiplication result;

[0208] Step S623: Multiply the second multiplication result by the compromise model parameter of the i-th level to obtain a second error corresponding to the i-th level.

[0209] In step S621, the first task of the computer system is to calculate the first absolute error between the reconstructed test data and the unperturbed data at the i-th level. Unlike the mean square error (MSE), the absolute error directly measures the absolute value of the difference between the data point and its true value, without considering the square of the error. This error measurement method may be more intuitive and easier to understand in some cases, especially when it is necessary to directly compare the deviation of data points.

[0210] Assume that the reconstructed test data of the i-th level is X recon,i， The unperturbed data is X true , both are time series data with a length of N. Then the calculation formula for the first absolute error is:

[0211] Absolute Errori,t=|x recon,i,t −x true,t ∣;

[0212] Among them, x recon,i,t and x true,tRespectively represent the values ​​of the i-th level reconstructed test data and the undisturbed data at time point t. In order to obtain the first absolute error of the entire time series, the absolute errors at each time point need to be summarized or further processed. However, in step S62, we are more concerned about the error in the masked interval, so the next steps will focus on analyzing this part of the data.

[0213] In step S622, the computer system performs an inner product on the first absolute error of the i-th level and the shielded test data corresponding to the bone tissue bending resistance test sample data. However, it should be noted that the inner product in mathematics usually refers to the operation of multiplying the corresponding elements of two vectors and then summing them, but when processing time series data and shielded test data, it is actually a weighted summation of the absolute errors in the shielded interval (although the weight here is binary 1 or 0).

[0214] To describe this process more accurately, it can be considered as a weighted sum operation instead of the traditional inner product calculation. The masked test data is used as a weight to select which absolute errors at which time points should be included in the calculation of the second error.

[0215] Assume that the shielded test data is M, where m t =1 means that the time point t is within the shielding interval, m t =0 means that the time point t is not in the shielding interval. Then the weighted summation (i.e. "inner product solution") process can be expressed as: ;

[0216] The summation here is actually done while traversing all time points, only when m t =1 (i.e., time point t is within the masking interval), the corresponding absolute error is included in the sum. In order to be consistent with the subsequent steps, this result is called the second multiplication result, although it is not obtained by the inner product calculation in the traditional sense. For example, continuing with the previous example, assume that the masked test data marks the interval from time point T1 to T2 as the masking interval. In this interval, the weighted sum of the first absolute errors is calculated as the second multiplication result:

[0217] ;

[0218] Wherein, Mask Interval represents the time range of the masking interval. By traversing all time points within this time range and adding their corresponding first absolute errors, a second multiplication result is obtained. In step S623, the computer system performs a product operation on the second multiplication result and the compromise model parameter of the i-th level to obtain the second error corresponding to the level. The compromise model parameter is a coefficient used to adjust the size of the error value. It can be set according to actual needs to balance the weights of errors between different levels or take into account other factors that affect error evaluation.

[0219] Assume that the parameter of the trade-off model at the i-th level is λ i , then the calculation formula of the second error is:

[0220] ;

[0221] By adjusting λ i The computer system can flexibly control the contribution of errors at different levels in the overall error assessment.

[0222] is the weighted absolute error sum (i.e., the second multiplication result) calculated in step S622. By multiplying this result with the compromise model parameter, the computer system obtains the final second error value, which will serve as an important basis for evaluating the error of the reconstructed test data of the i-th level in the shielding interval. In summary, step S62 successfully determines the second error between the reconstructed test data and the undisturbed data at each level in the shielding interval through a series of operations such as calculating the first absolute error, weighted summing with the shielded test data, and multiplying with the compromise model parameter. This process not only takes into account the absolute value of the error, but also realizes the detailed analysis and flexible adjustment of the error on the key data points through the introduction of the shielded test data and the compromise model parameter. In the training process of the target data disturbance cleaning model, these second error values ​​will provide important guidance for the optimization of model parameters and the improvement of performance.

[0223] As an implementation manner, the step S63 of determining the third error between the reconstructed test data of each level and the undisturbed data in the masked interval may specifically include:

[0224] Step S631: for the i-th level in the I levels, actual disturbance data is obtained according to the bone tissue anti-bending test sample data and the non-disturbance data, and a second absolute error between the disturbance data of the i-th level and the actual disturbance data is determined, 1≤i≤I;

[0225] Step S632: performing inner product calculation on the second absolute error corresponding to the i-th level and the shielded test data corresponding to the bone tissue anti-bending test sample data to obtain a third multiplication result;

[0226] Step S633: Multiply the third multiplication result and the compromise model parameter of the i-th level to obtain a third error corresponding to the i-th level.

[0227] In the training process of the target data disturbance cleaning model, step S63 focuses on evaluating the accuracy of the model in perturbation identification. Specifically, it aims to quantify the difference between the perturbation components in the reconstructed test data generated by the model and the perturbation components in the actual test data.

[0228] In step S631, the primary task of the computer system is to determine the actual disturbance data and calculate the difference between the disturbance data generated by the model and the actual disturbance data. The actual disturbance data cannot usually be observed directly, but can be estimated indirectly by comparing the original test data containing the disturbance and the non-disturbance data. In the bone tissue bending test, due to the influence of external factors such as equipment noise and ambient temperature changes, the collected test data inevitably contains disturbance components. In order to evaluate the ability of the model in disturbance identification, it is necessary to know the degree of closeness between the disturbance data generated by the model and the actual disturbance data. However, since the actual disturbance data is unknown, some filtering or denoising algorithm is usually used to try to remove the disturbance component from the original test data, so as to obtain an approximate estimate of the actual disturbance component. The computer system first applies some denoising algorithm (such as wavelet threshold denoising, median filtering, etc.) to the original test data to try to remove the disturbance component therein. The denoised data is regarded as an approximate representation of the non-disturbance data. Then, by subtracting this approximate non-disturbance data from the original data, an approximate estimate of the actual disturbance data is obtained. Assume that the original test data is X raw , the denoised data is X denoised , then the approximate estimate of the actual disturbance data is:

[0229] ;

[0230] Next, the computer system calculates the perturbation data X of the i-th level perturbation,i The second absolute error between the approximate estimate (generated by the model) and the actual perturbed data. The second absolute error measures the magnitude of the direct deviation from the model's perturbation-discriminating ability.

[0231] In step S632, the computer system performs an inner product solution (actually a weighted sum operation) on the second absolute error and the masked test data to focus the error analysis on the masked interval.

[0232] Assume that the shielded test data is M, where mt =1 means that the time point t is within the shielding interval, m t =0 means that the time point t is not within the shielding interval. Then the calculation formula for the third multiplication result (i.e., the weighted absolute error sum) is:

[0233] ;

[0234] Here, N is the total length of the time series. By summing the absolute errors only within the masked interval, the model’s ability to identify perturbations at key data points can be more accurately assessed.

[0235] Continuing with the previous example, assume that the shielded test data marks the interval from time point T1 to T2 as the shielded interval. In this interval, the weighted sum of the second absolute error is calculated as the third multiplication result. In step S633, the computer system performs a product operation on the third multiplication result and the compromise model parameter of the i-th level to obtain the third error corresponding to the level. The compromise model parameter is used to adjust the size of the error value so as to more flexibly control the weights of errors at different levels during the model training process. In summary, step S63 successfully determines the third error between the reconstructed test data and the unperturbed data at each level in the shielded interval through a series of fine operations. This process not only takes into account the accuracy of the model in perturbation identification, but also realizes the fine analysis and flexible adjustment of the errors on key data points through the introduction of shielded test data and compromise model parameters. These third error values ​​will provide important guidance for the optimization of model parameters and the improvement of performance during the training process of the target data perturbation cleaning model.

[0236] As an implementation manner, the step S64, performing error determination according to the first error, the second error and the third error to obtain the error of the target data disturbance cleaning model, may specifically include:

[0237] Step S641: summing the first errors corresponding to each level in the I levels to obtain a first value;

[0238] Step S642: summing the second errors corresponding to each level in the I levels, and multiplying the summation result by a preset first importance coefficient to obtain a second value;

[0239] Step S643: summing the third errors corresponding to each level in the I levels, and multiplying the summation result by a preset second importance coefficient to obtain a third value;

[0240] Step S644: sum the first value, the second value and the third value to obtain the error of the target data disturbance cleaning model.

[0241] In step S641, the primary task of the computer system is to sum up the first errors of all levels (a total of I levels) to obtain a comprehensive first numerical indicator (first value). The purpose of this step is to summarize the mean square error performance of the model on the masked interval for the reconstructed test data at each level, so as to evaluate the performance of the model in terms of overall numerical accuracy.

[0242] In step S642, the computer system sums the second errors of each level and multiplies the summation result with the preset first importance coefficient to obtain a weighted second numerical indicator (second value). The purpose of this step is to consider the relative importance of different levels in the model performance evaluation and to comprehensively evaluate the second errors.

[0243] Step S643 is similar to step S642, but processes the third error. In this step, the computer system sums the third errors of each level and multiplies the summation result with the preset second importance coefficient to obtain a weighted third numerical indicator (third value). This step is intended to evaluate the accuracy of the model in disturbance identification and reflect the importance differences of different levels in disturbance identification through weighted processing.

[0244] In step S644, the computer system sums the first value, the weighted second value, and the weighted third value to obtain the final error of the target data perturbation cleaning model. This step comprehensively evaluates the performance of the model in three aspects: numerical accuracy, structural similarity, and perturbation recognition accuracy. The first error, the second error, and the third error are comprehensively evaluated by summing and weighted summing, and the final error of the target data perturbation cleaning model is obtained. This process not only takes into account the performance of the model in various aspects, but also flexibly adjusts the errors at different levels through preset importance coefficients. In the scenario of bone tissue bending resistance test data analysis, this error evaluation method can comprehensively and accurately reflect the performance level of the model and provide strong support for subsequent data processing and analysis.

[0245] The present application also provides a computer system, such as Figure 2 As shown, the computer system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the computer system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation on the embodiments of the present application.

[0246] An embodiment of the present application provides a computer system, and the computer system in the embodiment of the present application includes: one or more processors; a memory; one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more programs are executed by the processor, the above-mentioned bone tissue anti-bending test data analysis method is implemented.

Claims

1. A method for analyzing bone tissue bending resistance test data, characterized in that: include: Obtaining the target bone tissue bending resistance test data to be analyzed, wherein the target bone tissue bending resistance test data includes basic test parameters, dynamic response data and disturbance data, wherein the basic test parameters include test time, sample number, loading rate, and maximum loading force; the dynamic response data is the displacement-time curve and stress-strain curve of the sample during the loading process; Based on the representation information extraction network in the j-th level branch model, the representation information of the target bone tissue anti-bending test data is extracted to obtain the pressure response representation vector of the target bone tissue anti-bending test data at the j-th level, wherein the j-th level branch model is the j-th level branch model among the I-level branch models included in the target data disturbance cleaning model, wherein j>1; Based on the disturbance information identification network in the branch model of the j-th level, the disturbance data of the target bone tissue anti-bending test data at the i-th level and the reconstructed test data at the i-th level output by the branch model of the i-th level are cleared to obtain the disturbance data of the target bone tissue anti-bending test data at the j-th level, and the branch model of the i-th level is the previous level of the j-th level in the I levels, wherein i=j-1; Based on the data reconstruction network in the branch model of the j-th level, a data reconstruction operation is performed on the pressure response characterization vector of the j-th level, the disturbance data of the j-th level and the reconstructed test data of the i-th level to obtain the reconstructed test data of the j-th level of the target bone tissue anti-bending test data; The disturbance data of the j-th level and the reconstructed test data of the j-th level are determined as the input information of the branch model of the j+1-th level, and are iterated one by one to obtain the reconstructed test data of the target bone tissue anti-bending test data at the I-th level as the reconstructed test data after the disturbance of the target bone tissue anti-bending test data is cleared.

2. The method according to claim 1, characterized in that The characterization information extraction network includes a characterization information extraction branch network and a characterization information refinement branch network. The characterization information extraction network in the branch model based on the j-th level extracts characterization information from the target bone tissue anti-bending test data to obtain a pressure response characterization vector of the target bone tissue anti-bending test data at the j-th level, including: Based on the characterization information extraction branch network, characterization information is extracted from the target bone tissue anti-bending test data to obtain a first pressure response characterization vector of the target bone tissue anti-bending test data; Based on the representation information extraction branch network, the representation information of the first pressure response representation vector is refined to obtain the j-th level pressure response representation vector.

3. The method according to claim 2, characterized in that The characterization information extraction branch network includes a characterization information extraction component and a residual structure. Based on the characterization information extraction branch network, characterization information is extracted from the target bone tissue anti-bending test data to obtain a first pressure response characterization vector of the target bone tissue anti-bending test data, including: Extracting a second pressure response characterization vector of the target bone tissue anti-bending test data based on the characterization information extraction component; The second pressure response characterization vector and the target bone tissue anti-bending test data are combined in the channel dimension to obtain the first pressure response characterization vector.

4. The method according to any one of claims 1 to 3, characterized in that The disturbance information identification network includes a first deep network layer, and the disturbance information identification network based on the branch model at the jth level performs disturbance removal on the disturbance data at the ith level and the reconstructed test data at the ith level of the target bone tissue anti-bending test data output by the branch model at the ith level, to obtain the disturbance data of the target bone tissue anti-bending test data at the jth level, including: Performing disturbance information calculation according to the disturbance data of the i-th level, the reconstruction test data of the i-th level and the target bone tissue anti-bending test data to obtain first disturbance data of the target bone tissue anti-bending test data; Based on the first deep network layer, clear the first disturbance data to obtain the disturbance data of the j-th level; The data reconstruction network includes a second deep network layer, and the data reconstruction network in the branch model based on the j-th level performs a data reconstruction operation on the pressure response characterization vector of the j-th level, the disturbance data of the j-th level, and the reconstructed test data of the i-th level to obtain the reconstructed test data of the target bone tissue anti-bending test data at the j-th level, including: Reconstruct data according to the disturbance data of the j-th level and the reconstructed test data of the i-th level to obtain first reconstructed test data of the target bone tissue anti-bending test data; Performing characterization vector integration according to the first reconstructed test data and the j-th level pressure response characterization vector to obtain an integrated characterization vector of the target bone tissue anti-bending test data; Based on the second deep network layer, a data reconstruction operation is performed on the integrated representation vector to obtain the reconstructed test data of the j-th level.

5. The method according to claim 4, characterized in that The performing disturbance information calculation according to the disturbance data of the i-th level, the reconstructed test data of the i-th level and the target bone tissue anti-bending test data to obtain first disturbance data of the target bone tissue anti-bending test data includes: The first disturbance value is obtained by multiplying the difference between the target bone tissue anti-bending test data and the reconstructed test data of the i-th level with the first model parameter update rate; A first variable is multiplied by the disturbance data of the i-th level to obtain a second disturbance value, wherein the first variable is a difference between a set value and an update rate of the first model parameter; taking the sum of the first disturbance value and the second disturbance value as the first disturbance data; The data reconstruction is performed according to the disturbance data of the j-th level and the reconstructed test data of the i-th level to obtain first reconstructed test data of the target bone tissue anti-bending test data, including: Determine the difference between the target bone tissue anti-bending test data and the j-th level disturbance data, and determine the multiplication result of the shielding test data corresponding to the target bone tissue anti-bending test data and the second model parameter update rate; Solving the inner product of the difference result and the multiplication result to obtain first reconstructed data; A second variable is multiplied by the reconstructed test data of the i-th level to obtain second reconstructed data, wherein the second variable is a difference between a set value and the multiplication result; taking the sum of the first reconstructed data and the second reconstructed data as the first reconstructed test data; The step of integrating the representation vectors according to the first reconstructed test data and the pressure response representation vector of the j-th level to obtain the integrated representation vector of the target bone tissue anti-bending test data includes: The first reconstructed test data and the j-th level pressure response representation vector are merged in the channel dimension to obtain the integrated representation vector.

6. The method according to any one of claims 1 to 3, characterized in that: The method also includes debugging steps of the target data disturbance cleaning model, including: Acquiring bone tissue anti-bending test sample data, wherein the bone tissue anti-bending test sample data includes disturbance data; Based on the representation information extraction network in the branch model at the j-th level, the representation information of the bone tissue anti-bending test sample data is extracted to obtain a pressure response representation vector of the bone tissue anti-bending test sample data at the j-th level; Based on the disturbance information recognition network in the j-th level branch model, the disturbance data of the bone tissue anti-bending test sample data at the i-th level output by the i-th level branch model and the reconstructed test data at the i-th level are cleared to obtain the disturbance data of the bone tissue anti-bending test sample data at the j-th level; Based on the data reconstruction network in the branch model of the j-th level, a data reconstruction operation is performed on the pressure response characterization vector of the j-th level, the disturbance data of the j-th level and the reconstructed test data of the i-th level to obtain the reconstructed test data of the j-th level of the bone tissue anti-bending test sample data; Determine the disturbance data of the j-th level and the reconstructed test data of the j-th level as the input information of the branch model of the j+1-th level, iterate one by one, and obtain the reconstructed test data of the bone tissue anti-bending test sample data at each level in the I levels; Determine the error based on the reconstructed test data of the bone tissue anti-bending test sample data at each level and the unperturbed data corresponding to the bone tissue anti-bending test sample data to obtain the error of the target data disturbance cleaning model; According to the error, updating the model parameter variables of the target data disturbance cleaning model; The representation information extraction network includes a pre-debugged representation information extraction component, and the target data disturbance cleaning model does not debug the representation information extraction component during the debugging phase.

7. The method according to claim 6, characterized in that The error determination is performed based on the reconstructed test data of the bone tissue anti-bending test sample data at each level and the unperturbed data corresponding to the bone tissue anti-bending test sample data to obtain the error of the target data disturbance cleaning model, including: Determine a first error between the reconstructed test data of each level and the undisturbed data in a shielded interval, wherein the shielded interval is a shielded interval in the shielded test data corresponding to the bone tissue anti-bending test sample data; Determine a second error between the reconstructed test data of each level and the undisturbed data in the masked interval; Determine a third error between the reconstructed test data of each level and the undisturbed data in the masked interval; An error is determined according to the first error, the second error and the third error to obtain an error of the target data disturbance cleaning model.

8. The method according to claim 7, characterized in that The determining of a first error between the reconstructed test data of each level and the undisturbed data in a masked interval includes: For the i-th level among the I levels, determine the mean square error between the reconstructed test data of the i-th level and the unperturbed data, 1≤i≤I; Solve the inner product of the mean square error corresponding to the i-th level and the shielded test data corresponding to the bone tissue anti-bending test sample data to obtain a first multiplication result; Multiplying the first multiplication result by the compromise model parameter of the i-th level to obtain a first error corresponding to the i-th level; The determining of a second error between the reconstructed test data of each level and the undisturbed data in the masking interval includes: For the i-th level among the I levels, determining a first absolute error between the reconstructed test data of the i-th level and the undisturbed data, 1≤i≤I; Solve the inner product of the first absolute error corresponding to the i-th level and the shielded test data corresponding to the bone tissue anti-bending test sample data to obtain a second multiplication result; Multiplying the second multiplication result by the compromise model parameter of the i-th level to obtain a second error corresponding to the i-th level; The determining of a third error between the reconstructed test data of each level and the undisturbed data in the masking interval comprises: For the i-th level among the I levels, actual disturbance data is obtained according to the bone tissue anti-bending test sample data and the non-disturbance data, and a second absolute error between the disturbance data of the i-th level and the actual disturbance data is determined, 1≤i≤I; Solve the inner product of the second absolute error corresponding to the i-th level and the shielded test data corresponding to the bone tissue anti-bending test sample data to obtain a third multiplication result; The third multiplication result is multiplied by the compromise model parameter of the i-th level to obtain a third error corresponding to the i-th level.

9. The method according to claim 7, characterized in that: The performing error determination according to the first error, the second error and the third error to obtain the error of the target data disturbance cleaning model includes: Sum the first errors corresponding to each level in the I levels to obtain a first value; Sum the second errors corresponding to each level in the I levels, and multiply the summation result by a preset first importance coefficient to obtain a second value; summing the third errors corresponding to each level in the I levels, and multiplying the summation result by a preset second importance coefficient to obtain a third value; The first value, the second value and the third value are summed to obtain an error of the target data disturbance cleaning model.

10. A computer system, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 9 is implemented.

Images