Carbon emission data enhancement method based on conditional diffusion model

Abstract

The invention discloses a carbon emission data enhancement method based on a conditional diffusion model, relates to the field of data processing, and solves the technical problems of insufficient samples and low data quality in a traditional carbon emission data enhancement method. The method comprises the following steps: acquiring original carbon emission data, performing classification preprocessing on the original carbon emission data to generate preprocessed data, constructing a precoder, and training the precoder through the preprocessed data to generate a stable precoder; the method comprises the following steps: designing a diffusion model, optimizing the diffusion model in combination with a preset reconstruction error to establish a conditional diffusion model, and processing carbon emission original data through the conditional diffusion model and a stable precoder to generate a target carbon emission sample. According to the invention, high-quality carbon emission data enhancement is realized, and a more reliable and efficient solution is provided for digital construction of carbon emission data.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and more particularly, to a carbon emission data enhancement method based on a conditional diffusion model. Background Art

[0002] When enhancing carbon emission data, industry characteristics bring two major difficulties: (1) The sources of carbon emission data are diverse and heterogeneous, resulting in the frequent loss of specific features and uneven distribution of various types of data; (2) Different companies have different focuses on carbon emissions, so the data enhancement needs are also different.

[0003] To address the difficulty of modeling heterogeneous models, the following three reference documents are listed: Reference document 1, publication number CN119180420A, discloses a method and system for correcting carbon emission data errors. The method discloses obtaining environmental data by road section and time period based on traffic information, then dividing the road section into sub-segments based on this environmental data to obtain initial carbon emission factors. Adjustment coefficients are obtained based on the sub-segment environmental data and vehicle type. The initial factors are adjusted to obtain the final carbon emission factors, and the total carbon emissions are calculated to correct carbon emission data errors.

[0004] Comparative Document 2, Publication No. CN118282948A, discloses a method for network traffic generation data enhancement based on a diffusion model. This method converts raw network flow signals into an RGB image dataset, collects abnormal traffic, and converts these images into images for diffusion model training. After the model is trained, new abnormal traffic images are generated based on random noise. The generated data is then merged with the original dataset to form a balanced sample set. This method can address the scarcity of labeled data based on the diffusion model and achieve data enhancement.

[0005] Reference document 3, publication number CN118887117A, discloses a method and system for industrial image data enhancement and defect generation based on a diffusion model. The method involves first acquiring an original industrial image, then generating an image with simulated defects using a deep learning diffusion model. The diffusion model is then used to further adjust the data distribution, evaluate the generated quality, and optimize the model. The resulting model can then be used to perform image enhancement or defect generation on new data input.

[0006] However, existing methods (such as fitting adjustment coefficients) have limited modeling capabilities and are difficult to adapt to complex scenarios. In particular, these methods are often only able to enhance specific applications (such as transportation) and cannot effectively enhance carbon emission data from multiple sources.

[0007] Furthermore, current diffusion model designs for data augmentation generally ignore the issue of heterogeneity in demand and lack feedback-based iterative optimization mechanisms. Furthermore, such technologies have yet to be applied in the field of carbon emissions, necessitating the development of more advanced model design solutions. Furthermore, compared to other types of data, carbon emissions data often comes from diverse sources and is difficult to collect. Furthermore, the application requirements for carbon emissions are complex, and targeted optimization designs are lacking. Consequently, issues such as missing carbon emissions data and uneven distribution are widespread. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to address the shortcomings of the existing technology and provide a carbon emission data enhancement method based on the conditional diffusion model, which solves the technical problems of insufficient samples and low data quality in traditional carbon emission data enhancement methods.

[0009] The carbon emission data enhancement method based on the conditional diffusion model of the present invention comprises:

[0010] Acquiring raw carbon emission data, classifying and preprocessing the raw carbon emission data to generate preprocessed data, constructing a precoder, and training the precoder using the preprocessed data to generate a stable precoder;

[0011] A diffusion model is designed and optimized in combination with a preset reconstruction error to establish a conditional diffusion model. The carbon emission raw data is processed by the conditional diffusion model and a stable precoder f to generate a target carbon emission sample.

[0012] As a further improvement, the method for classifying and preprocessing the raw carbon emission data to generate preprocessed data is as follows:

[0013] Set a carbon emission scenario, classify the carbon emission raw data according to the source type and environmental conditions of the carbon emission scenario to obtain classified data, and record the main feature set in the classified data as

[0014] Among them, d is the number of features, r i is the i-th feature; each feature r in the main feature set is analyzed by the word element analyzer i Transform to generate the corresponding feature vector t i , each feature r i and the corresponding eigenvector t i Splicing generates preprocessed data.

[0015] Furthermore, the carbon emission scenarios include industrial production emission scenarios, building use process emission scenarios and transportation emission scenarios.

[0016] Furthermore, the method of training the precoder by the preprocessed data to generate a stable precoder f is:

[0017] The attention module k is constructed by adding position encoding to the feature vector in the preprocessed data. p layer, for the attention module k p The pre-coded output a is obtained by repeating the layers in series; the pre-coded output a is input into the global encoder to obtain a continuous value e', and the absolute value error |e-e'| is obtained according to the continuous value e' and the carbon emission e in the pre-processed data, and the pre-coded output a is adjusted according to the absolute value error |e-e'| to obtain the target pre-coded The target precoding As the precoding of the stable precoder f output.

[0018] Furthermore, the diffusion model includes:

[0019] The carbon emission raw data is gradually added with noise through multiple steps T to obtain the initial noise sequence R (1) , R (2) , ..., R (T) , where the transformation of each step can be expressed as:

[0020] Among them, R (t) is a noisy sample, t is a noisy sample R (t) and the corresponding time step, ∈ is the sampling of the noise distribution of the original carbon emission data, β t is a preset noise variance coefficient;

[0021] Optimizing an initial noise sequence based on the stable precoder f to generate a target noise sequence;

[0022] Set up a reverse denoising model, and use the reverse denoising model to transform the noisy samples R in the target noise sequence into (t) Perform restoration;

[0023] Furthermore, when the main feature of the carbon emission raw data is a continuous feature, Gaussian noise is added to the carbon emission raw data to generate noisy samples; when the main feature of the carbon emission raw data is a discrete feature, Poisson noise is added to the carbon emission raw data to generate noisy samples.

[0024] Furthermore, a method for optimizing the diffusion model in combination with the reconstruction error to establish a conditional diffusion model is:

[0025] Construct prediction error ∈ θ (R (t), t), based on the stable precoder f prediction error ∈ θ (R (t) , t) package, obtain the error modeling function, add constraints c to the error modeling function to construct the reconstruction error ∈ θ (R (t) , t, c), according to the reconstruction error ∈ θ (R (t) , t, c) Reconstruct the reverse denoising model.

[0026] Furthermore, the expression of the error modeling function is:

[0027] ∈ θ (R (t) , t) = MLP([f(R (t) ), t]) = MLP([a (t) , t]);

[0028] Among them, MLP is a neural network composed of multiple layers, a (t) is the corresponding code obtained by the noisy sample through the precoder.

[0029] Furthermore, the expression of the reconstructed inverse denoising model is:

[0030]

[0031] Among them, R (t-1) is the denoised sample.

[0032] Furthermore, the method also includes an expert knowledge condition control and quality verification module for indicating that the uncertainty of generating the noisy sample is high and issuing an uncertainty alarm when the reconstruction error is greater than a preset reconstruction error threshold.

[0033] Beneficial effects

[0034] The advantages of the present invention are:

[0035] The present invention obtains raw carbon emission data, classifies and preprocesses the raw carbon emission data to generate preprocessed data, constructs a precoder, trains the precoder through the preprocessed data to generate a stable precoder f, designs a diffusion model, and optimizes the diffusion model in combination with a preset reconstruction error to establish a conditional diffusion model. The conditional diffusion model is used to process the precoding output of the stable precoder to generate a target carbon emission sample, thereby solving the problems of insufficient samples and low data quality in traditional carbon emission data enhancement methods, achieving high-quality carbon emission data enhancement, and providing a more reliable and efficient solution for the digital construction of carbon emission data. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of the carbon emission data enhancement method based on the conditional diffusion model of the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further described below in conjunction with the embodiments, but this does not constitute any limitation to the present invention. Any limited number of modifications made by anyone within the scope of the claims of the present invention are still within the scope of the claims of the present invention.

[0038] See Figure 1 The carbon emission data enhancement method based on the conditional diffusion model of the present invention includes: obtaining carbon emission raw data, classifying and preprocessing the carbon emission raw data to generate preprocessed data, constructing a precoder, and training the precoder through the preprocessed data to generate a stable precoder f; designing a diffusion model, and optimizing the diffusion model in combination with a preset reconstruction error to establish a conditional diffusion model, and processing the carbon emission raw data through the conditional diffusion model and the stable precoder f to generate a target carbon emission sample.

[0039] For carbon emission data in different scenarios, this method first classifies and encodes the original data according to the source type and environmental conditions, and designs precoding and global coding to optimize the feature extraction capability. Based on the trained precoder, the conditional diffusion model is trained to support the batch generation of new samples. Subsequently, expert knowledge is introduced to optimize the conditional control, and the properties of the generated samples are precisely adjusted to enhance the physical interpretability of the generated data. Finally, based on the optimized conditional diffusion model, specific categories of data can be generated in a targeted manner to solve the problem of data imbalance, and poor quality data can also be cleaned by partial noise injection, ultimately achieving high-quality carbon emission data enhancement. The present invention innovatively applies the conditional diffusion model to the field of carbon emissions, providing a more reliable and efficient solution for the digital construction of carbon emission data.

[0040] Traditional methods typically apply to standardized carbon emissions data, using common data formats such as tables and dictionaries, which are relatively complete and uniform in format. However, the characteristics of carbon emissions data from various sources and types are not aligned. During integration, feature indexes do not overlap, making data collection difficult, leading to widespread data gaps and uneven distribution.

[0041] To solve this problem, this patent classifies and pre-processes the raw data according to source type and environmental conditions to prepare for subsequent conditional diffusion model training. A prominent difficulty is that the characteristics of carbon emissions from various sources vary greatly, but an encoding algorithm is required that has cross-type generalization. To this end, we designed a set of hierarchical encoders. The various features of carbon emission data of each category, except for the emissions, will be input into their respective pre-encoders, which obtain the input encoding based on a multi-layer Attention structure; the output encoding will be received by a global encoder-decoder, and the final result will be used to fit the carbon emissions to optimize the encoding effect. The specific processing steps are as follows:

[0042] The method for classifying and preprocessing raw carbon emission data to generate preprocessed data is to set carbon emission scenarios and classify the raw carbon emission data based on the source type and environmental conditions of the carbon emission scenarios to obtain preprocessed data. First, we preprocess the raw data. Classic carbon emission scenarios will be further classified based on the following principles:

[0043] Industrial production emission scenarios: classified according to different industries (steel, chemicals, building materials, etc.).

[0044] Emission scenarios during building use: classified by building type (residential, commercial, public buildings).

[0045] Transportation emission scenarios: Classified according to vehicle type, road section characteristics and time period.

[0046] The method of pre-encoding the pre-processed data to generate the pre-coder for carbon emission data is as follows: Taking the carbon emission of residential buildings as an example, the input R of the pre-processed data often contains the main features such as facade type, building height, and number of underground floors, which are recorded as Where d is the number of features, r i is the i-th feature, usually a (feature name - feature value) pair, such as (building height - 6 floors). i , we call the classic tokenizer to perform a preliminary conversion on the input to obtain the corresponding vector t i , and then concatenated with all features and input to the pre-encoder.

[0047] In the selection of the type of pre-encoder f, the attention module has become a common and widely verified framework. By adding position encoding to different features, the importance and special properties of different features can be captured. The vector t of each input feature i It will be regarded as a token with dimension n. The attention module maps all tokens to QKV, then matches K based on Q, determines the weight of each K through Softmax, and multiplies it with V to get the final output. The specific formula is:

[0048]

[0049] Where Q represents the query vector, K represents the key vector, V represents the value vector, dk represents the dimensions of the query vector and the key vector, QK' represents the matrix multiplication of the query vector and the key vector, and Softmax represents the normalization along the key dimension (i.e., each row of the matrix).

[0050] Repeatedly build such an attention module k p layer, and finally obtain the pre-coded output a, which is recorded as a=f(R).

[0051] However, it should be noted that the above is only the input construction method. Our ultimate goal is to obtain a good encoder f. In order to iteratively optimize the encoding effect and capture the global features related to carbon emissions, we need to introduce a global encoder for further processing and optimization. In the model selection, we continue to use the attention module to ensure the effectiveness and consistency of the encoding. Specifically, the input a passes through the attention module k p The output after the attention module is further input to a fully connected layer to obtain a continuous value e' as the output. By minimizing the absolute error between the output and the carbon emission e of the sample (note: absolute error = |e-e'|), we can optimize the pre-encoding module and the global encoding module. During the entire data management phase, all carbon emission samples will be serially input into the two modules in a random order. The parameters are iteratively updated through the gradient descent method to obtain a stable pre-encoder. Based on this model encoder f, we can batch generate the final target pre-encoder Where n is the total number of samples of carbon emission data.

[0052] Neural networks can be viewed as constructing input values ​​(i.e., r i ) to the output value (i.e., e'). The backpropagation algorithm automatically optimizes the model parameters, or the neural network, by minimizing the loss function associated with the output e' (i.e., |e - e'|). This solution is based on the concept of an auto encoder-decoder, which views the neural network computational process from input to encoding and then from encoding to output. While the encoding is not directly optimized, it can be co-optimized with the output optimization process to compress and model key properties of the input. The element a in the above content corresponds to this element, which will be adjusted and optimized during the backpropagation algorithm's optimization of |e - e'|.

[0053] Diffusion model training and optimization

[0054] In order to enhance carbon emission data, we design a corresponding diffusion model, the basic structure of which is as follows:

[0055] 1. In the forward diffusion process, the original data R is gradually added with noise through multiple steps T to obtain the noise sequence R (1) , R (2) , ..., R (T) , where the transformation of each step can be expressed as:

[0056] Where ∈ is the sampling of the noise distribution, β t is a predetermined noise variance coefficient.

[0057] 2. In the reverse denoising process, it is necessary to train the neural network to predict the noise ∈, gradually from R (t) Restore to R. The denoising process can be formulated as an inverse denoising model:

[0058]

[0059] In the above formula, R (t) is a noisy sample, ∈ is a sampling of noise distribution, β t is the preset noise variance coefficient.

[0060] In view of the complex and diverse nature of carbon emission data, we designed the following noise addition strategy:

[0061] 1. Design an appropriate sequence of noise diffusion steps based on the characteristic distribution of different data types. When the raw carbon emission data is continuous, Gaussian noise is added to the data; when the raw carbon emission data is discrete, Poisson noise is added to the data. That is, for continuous features, Gaussian noise is used; for discrete features, Poisson noise is used.

[0062] 2. When adding noise, the physical characteristics and boundary constraints of the data need to be considered, such as many eigenvalues ​​cannot be negative.

[0063] The optimization of the entire diffusion model is based on the noise addition strategy L = E[||∈-∈ θ (R (t) , t)|| 2 ]. Note that in the above denoising process, it is necessary to construct the prediction error ∈ θ (R (t) , t), whose input is the noisy sample R at step t of the diffusion process and the corresponding time step t. This model needs to be able to independently model different types of inputs and have sufficient generalization to integrate them, which is the purpose of training the precoder f in the first step. Specifically, the error modeling function will be encapsulated as follows:

[0064] ∈ θ (R (t), t) = MLP([f(R (t) ), t]) = MLP([a (t) , t]);

[0065] Among them, MLP is a neural network composed of multiple layers of fully connected layers, with noise samples R (t) First, get the corresponding code a through the first stage trained f (t) , while freezing the model parameters f, the parameters of the MLP are trained to minimize the loss error. The trained model can generate carbon emission samples based on random noise.

[0066] On this basis, we designed a conditional diffusion model to enhance the model's generation performance and facilitate users in generating specific types of carbon emission data. For example, in building carbon emissions, there are relatively few samples with three underground floors. To balance the sample distribution, additional samples need to be generated. The number of underground floors can be used as the input condition c, and the conditional distribution of the modeling error, i.e., the reconstruction error, is:

[0067] ∈ θ (R (t) , t, c) = MLP([a (t) ,t,c]);

[0068] The design of this module has a considerable degree of flexibility. For any single or multiple feature requirements, corresponding restrictions can be added to the MLP training in the second stage.

[0069] The method of optimizing the diffusion model in combination with the reconstruction error to establish the conditional diffusion model is:

[0070] The precoding is used as the input of the conditional diffusion model, and the noise addition strategy is triggered according to the precoding to generate noisy samples. The noisy samples are reversely denoised according to the reconstruction error to generate denoised samples, and the denoised samples are used as the output of the conditional diffusion model.

[0071] The expression for generating denoised samples by reverse denoising the noisy samples according to the reconstruction error is:

[0072]

[0073] Among them, R (t) is a noisy sample, ∈ θ (R (t) , t, c) is the reconstruction error, β t is the preset noise variance coefficient, t is the time step in the process of adding noise, R (t-1) is the denoised sample, and c is the constraint condition.

[0074] The expression of reconstruction error is:

[0075] ∈θ (R (t) , t, c) = MLP([f(R (t) ), t, c]) = MLP([a (t) ,t,c]);

[0076] Among them, MLP is a neural network composed of multiple layers, R (t) is a noisy sample, a (t) is the corresponding code obtained by the noisy sample through the precoder, and c is the constraint condition.

[0077] Condition control and quality verification modules based on expert knowledge:

[0078] To maximize the use of the conditional diffusion model, this paper introduces the concept of human-in-the-loop (HIL) to maximize the model's generation effect under limited expert guidance. This approach systematically integrates expert knowledge into the model's training and validation processes, ensuring the quality and reliability of the generated data.

[0079] In practical applications, the integration of expert knowledge follows the following process:

[0080] 1. During the preprocessing phase, carbon emission experts will define the boundary conditions and constraints for data generation, focusing on setting a reasonable range of characteristic values ​​and variation intervals, and provide easily optimized evaluation criteria for the quality of constructed data. For example, if buildings with higher floors have higher carbon emissions, that is, the carbon emissions of buildings with higher floors are greater than the preset building carbon emission threshold, when the opposite generation result occurs, the preprocessing method of the model needs to be optimized and adjusted to make the carbon emissions of buildings with higher floors less than or equal to the building carbon emission threshold.

[0081] 2. During the generation process, the noise evaluation module can be used in reverse to evaluate the noise of the generated results. When the prediction error ∈ θ (R (t) When the mean of the evaluation results output by t) is greater than the preset prediction error threshold, it can be regarded that the uncertainty of the generated sample is high, and an uncertainty alarm is issued, which requires expert intervention, so as to achieve cost-effective dynamic adjustment of model parameters and noise distribution.

[0082] Therefore, based on the generalized design, it can be ensured that the quality verification link meets the three-level control mechanism:

[0083] 1. Automated Verification: Perform numerical rationality checks based on the aforementioned rules, analyze the original samples, and calculate consistency verification of statistical features.

[0084] 2. Expert review: Introduce manual evaluation under the above sample evaluation mechanism to generate a qualitative evaluation of the effect.

[0085] 3. System optimization. Continuously improve the model based on the verification results, such as designing specific pre-encoders for features with poor modeling; optimize the conditional control strategy to meet the dynamic needs of users.

[0086] Based on the trained pre-encoder f and MLP, users can batch enhance the dataset under the conditional constraints c defined by experts. The specific steps are as follows:

[0087] 1. Randomly generate noise R (t) , whose size is aligned with the dimension of the input feature R.

[0088] 2. Based on the reverse denoising process, continuously strip the noise term by predicting the noise term ∈ θ (R (t) , t, c).

[0089] 3. Generate the final sample R (0) .

[0090] On the other hand, by adding a small amount of noise to samples with missing values or outliers until the time step T’ < T, repeat the above steps for T’ steps until the repaired samples are generated.

[0091] Through this systematic method, we can significantly improve the efficiency and reliability of data augmentation while ensuring data quality. The effective integration of expert knowledge not only improves the interpretability of the model but also provides a strong guarantee for the practical application of the generated data.

[0092] The above are only the preferred embodiments of the present invention. It should be noted that for those skilled in the art, without departing from the structure of the present invention, several modifications and improvements can be made, which will not affect the implementation effect of the present invention and the practicality of the patent.

Claims

1. A carbon emission data enhancement method based on a conditional diffusion model, characterized in that: The method includes: Acquire raw carbon emission data, perform classification preprocessing on the raw carbon emission data to generate preprocessed data, construct a precoder, and train the precoder using the preprocessed data to generate a stable precoder f; A diffusion model is designed and optimized in combination with a preset reconstruction error to establish a conditional diffusion model. The carbon emission raw data is processed by the conditional diffusion model and a stable precoder f to generate a target carbon emission sample.

2. The carbon emission data enhancement method based on the conditional diffusion model according to claim 1 is characterized in that: The method for classifying and preprocessing the raw carbon emission data to generate preprocessed data is as follows: Set a carbon emission scenario, classify the carbon emission raw data according to the source type and environmental conditions of the carbon emission scenario to obtain classified data, and record the main feature set in the classified data as Among them, d is the number of features, r i is the i-th feature; each feature r in the main feature set is analyzed by the word element analyzer i Transform to generate the corresponding feature vector t i , each feature r i and the corresponding eigenvector t i Splicing generates preprocessed data.

3. The carbon emission data enhancement method based on the conditional diffusion model according to claim 2 is characterized in that: The carbon emission scenarios include industrial production emission scenarios, building use process emission scenarios and transportation emission scenarios.

4. The carbon emission data enhancement method based on the conditional diffusion model according to claim 1 is characterized in that: The method for training the precoder by the preprocessed data to generate a stable precoder f is: The attention module k is constructed by adding position encoding to the feature vector in the preprocessed data. p layer, for the attention module k p The pre-coded output a is obtained by repeating the layers in series; the pre-coded output a is input into the global encoder to obtain a continuous value e', and the absolute value error |e-e'| is obtained according to the continuous value e' and the carbon emission e in the pre-processed data, and the pre-coded output a is adjusted according to the absolute value error |e-e'| to obtain the target pre-coded The target precoding As the precoding of the stable precoder f output.

5. The carbon emission data enhancement method based on the conditional diffusion model according to claim 1 is characterized in that: The diffusion model includes: The carbon emission raw data is gradually added with noise through multiple steps T to obtain the initial noise sequence R (1) , R (2) , ..., R (T) , where the transformation of each step can be expressed as: Among them, R (t) is the noisy sample of the initial noise sequence, t is the noisy sample R (t) and the corresponding time step, ∈ is the sampling of the noise distribution of the original carbon emission data, β t is a preset noise variance coefficient; Optimizing an initial noise sequence based on the stable precoder f to generate a target noise sequence; A reverse denoising model is set, and noisy samples of the target noise sequence are restored by the reverse denoising model.

6. The carbon emission data enhancement method based on the conditional diffusion model according to claim 5 is characterized in that: When the main feature of the carbon emission raw data is a continuous feature, Gaussian noise is added to the carbon emission raw data to generate noisy samples; when the main feature of the carbon emission raw data is a discrete feature, Poisson noise is added to the carbon emission raw data to generate noisy samples.

7. The carbon emission data enhancement method based on the conditional diffusion model according to claim 5 is characterized in that: The method for optimizing the diffusion model in combination with the reconstruction error to establish a conditional diffusion model is: Construct prediction error ∈ θ (R (t) , t), based on the stable precoder f prediction error ∈ θ (R (t) , t) is encapsulated to obtain an error modeling function, and a constraint condition c is added to the error modeling function to construct a reconstruction error ∈ θ (R (t) , t, c), according to the reconstruction error ∈ θ (R (t) , t, c) Reconstruct the reverse denoising model.

8. The carbon emission data enhancement method based on the conditional diffusion model according to claim 7 is characterized in that: The expression of the error modeling function is: ∈ θ (R (t) ,t)=MLP([f(R (t) ),t])=MLP([a (t) ,t]); Among them, MLP is a neural network composed of multiple layers, a (t) is the noisy sample R of the initial noise sequence (t) The corresponding coding is obtained after optimization of the stable precoder f.

9. The carbon emission data enhancement method based on the conditional diffusion model according to claim 7 is characterized in that: The expression of the reconstructed inverse denoising model is: Among them, R (t-1) is the denoised sample.

10. The carbon emission data enhancement method based on the conditional diffusion model according to claim 7, characterized in that: The method also includes an expert knowledge condition control and quality verification module for indicating that the uncertainty of generating the noisy sample is high and issuing an uncertainty alarm when the reconstruction error is greater than a preset reconstruction error threshold.

Images