Time sequence sensing signal interpolation method based on consistency model view angle
By adopting the denoising diffusion probability model and optimal path search method with a consistent model perspective in time series sensing signal interpolation, the problem of slow existing interpolation methods is solved, and efficient real-time interpolation is achieved.
Patent Information
- Application Number
- CN202510304021.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-14
- Publication Date
- 2025-06-27
AI Technical Summary
The existing time series sensing signal interpolation method based on diffusion model requires dozens or even hundreds of model verification steps during the sample generation process, resulting in a slow interpolation process and cannot meet the needs of real-time application scenarios.
The denoising diffusion probability model based on the perspective of the consistency model is adopted, and the sampling path is optimized through the optimal path search method to reduce the steps and time required for sample generation.
The time series sensing signal interpolation model is accelerated, efficiency is improved, the needs of real-time scenarios can be met, and the stability of interpolation performance is maintained.
Smart Images
Figure CN120216893A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of sensing, and specifically relates to an interpolation method for time series sensing signals from the perspective of a consistency model. Background Art
[0002] Time series data is ubiquitous in life. For example, in air quality detection, a series of indicators recorded over time constitute a time series. These data play an important role in people's lives and can effectively guide decision-making. However, due to reasons such as sensor failure and information transmission errors, many time series data have natural missing values, that is, there is no real data, which will reduce the reliability of subsequent applications based on this data and cause unnecessary losses. Many technologies have tried to solve this problem. For example, statistical-based methods use some simple statistical models to achieve interpolation, and machine learning-based methods use more complex parametric models to achieve more accurate interpolation, etc. Among these methods, the interpolation method based on the diffusion model stands out and achieves a more powerful interpolation accuracy compared to other methods.
[0003] Although the interpolation method based on the diffusion model has a high interpolation accuracy, its sample generation process requires dozens or even hundreds of model verification steps, and the number of samples required for a single interpolation is also large. Therefore, the interpolation process is quite slow. It may take several minutes to obtain the processing result of a time series data segment, which is unacceptable for some real-time application scenarios.
[0004] To solve this problem, some technologies use non-Markov processes to construct sampling methods and use deterministic sampling processes to achieve skip sampling, but this brings certain performance losses; other technologies attempt to achieve skip steps while retaining accuracy through high-order approximations, but their performance on some datasets is even worse than that of low-order methods. In addition, they often do not consider the sampling path planning during skip sampling and only use artificially specified sampling plans or uniform plans, which also limits their sampling performance to a certain extent. Summary of the Invention
[0005] To solve the problems in the background art, the present invention proposes an interpolation method for time series sensing signals from the perspective of a consistency model.
[0006] The technical solution adopted by the interpolation method of the present invention includes the following steps:
[0007] S1. Obtain time series sensing signals through sensors installed on industrial equipment.
[0008] S2. Set the interval of the required diffusion step size according to the obtained time series sensing signal and obtain the signal-to-noise ratio at each diffusion step size. Divide the interval of the diffusion step size into three stages: early, middle, and late according to the preset two signal-to-noise ratio breakpoints, and preset the number of sampling points in each stage. Then, perform division processing and sampling point processing on the interval of the diffusion step size in turn to obtain a restricted solution space. Use an optimal path search method to process the obtained restricted solution space to obtain the sampled path after searching.
[0009] S3. Use the denoising diffusion probability model accelerated from the perspective of the consistency model to perform signal interpolation processing on the time series sensing signal according to the obtained sampled path, and obtain the time series sensing signal after signal interpolation.
[0010] The obtained time series sensing signal after signal interpolation is used as relevant reference data for the maintenance of industrial equipment, thereby further reducing the production line downtime.
[0011] The specific steps of step S2 include the following steps:
[0012] S21. Determine the total required diffusion step size according to the obtained time series sensing signal, thereby obtaining the interval of the diffusion step size.
[0013] S22. Obtain the signal-to-noise ratio at each diffusion step size. Divide the interval of the diffusion step size into three stages: early, middle, and late according to the preset two signal-to-noise ratio breakpoints, and preset the number of sampling points in each stage.
[0014] S23. The set of all sampling paths composed of the sampling points in all stages forms a restricted solution space.
[0015] S24. Use an optimal path search method to process the obtained restricted solution space to obtain the sampled path after searching.
[0016] The optimal path search method in step S24 uses a genetic algorithm.
[0017] The two signal-to-noise ratio breakpoints in step S22 are respectively the signal-to-noise ratio equal to 10 -2 and the signal-to-noise ratio equal to 1.
[0018] The denoising diffusion probability model accelerated from the perspective of the consistency model is a model that only accelerates the reverse process of the denoising diffusion probability model.
[0019] The acceleration processing of the reverse process of the denoising diffusion probability model is set according to the following formula:
[0020]
[0021] where is the t at the i-th step i sampling point according to Estimates made on real data For step i at time t i Noisy data at the sampling point For step i at time t i Coefficients at the sampling point For step i at time t i Prediction of noise at the sampling point For step i - 1 at time t i-1 Noisy data at the sampling point For step i - 1 at time t i-1 Coefficients at the sampling point Is the standard normal distribution, ∈ is Gaussian noise, and i is the i-th sampling step.
[0022] The beneficial effects of the present invention are as follows:
[0023] 1. The present invention realizes the acceleration of the diffusion interpolation model for any time series sensing signal, enabling a significant improvement in efficiency while maintaining performance, and can meet timely scenarios.
[0024] 2. The optimal path search method under the restricted solution space proposed by the present invention can solve the sampling path planning problem, find a path superior to uniform sampling, and does not require manual intervention. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 Is a comparison diagram of the restricted solution space and the standard solution space in Embodiment 1 of the present invention;
[0026] Figure 2 Is a reverse process diagram of the accelerated diffusion interpolation model in Embodiment 1 of the present invention;
[0027] Figure 3 Is a schematic diagram of the influence of the sampling quantity on the DDPM sampling performance in Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0028] The following further describes the present invention with reference to the drawings. It should be understood that the embodiments described herein are only used to illustrate and explain the present invention, and do not limit the protection scope of the claims of this application.
[0029] Embodiment 1
[0030] This embodiment is implemented by the following steps:
[0031] S1. Obtain time series sensing signals through sensors installed on industrial equipment.
[0032] S2. Set the interval of the required diffusion step size according to the obtained time series sensing signal and obtain the signal-to-noise ratio at each diffusion step size. Divide the interval of the diffusion step size and process the sampling points successively according to the preset two signal-to-noise ratio demarcation points to obtain a restricted solution space. Use the optimal path search method to process the obtained restricted solution space to obtain the sampled path after searching.
[0033] The diffusion step size is the diffusion step size that needs to be determined for all diffusion models used for interpolating the time series sensing signal.
[0034] S21. Determine the total required diffusion step size according to the obtained time series sensing signal, so as to obtain the interval of the diffusion step size.
[0035] S22. Obtain the signal-to-noise ratio at each diffusion step size. Divide the interval of the diffusion step size into three stages: early, middle, and late according to the two signal-to-noise ratio demarcation points, and preset the number of sampling points in each stage.
[0036] The sampling point is a specific diffusion step selected within the diffusion step interval.
[0037] S23. The set of all sampling paths composed of a certain number of sampling points in all stages forms a restricted solution space.
[0038] S24. Use the optimal path search method to process the obtained restricted solution space to obtain the optimal sampled path after searching.
[0039] The two signal-to-noise ratio demarcation points in step S22 are the signal-to-noise ratio equal to 10 -2 and the signal-to-noise ratio equal to 1.
[0040] As Figure 1 shown, determine that the total diffusion step size is T, and obtain the interval of the diffusion step size as [0, T - 1]. The upper part is the conventional method without demarcating the interval of the diffusion step size, and directly sample j sampling points within the interval [1, T - 1] of the diffusion step size. The set of all sampling paths composed of these j sampling points forms a standard solution space. In specific implementation, the sampling path includes the endpoints with the diffusion step size of T.
[0041] The standard solution space is very large. For example, when T = 50, if we want to search for a sampling path with a total extended step size of 10 steps, after excluding the endpoint 49, we need to randomly select 9 sampling points from 1 - 48. This solution space is quite large and may even cause the optimal path search method to fail in the evolutionary search (under specific parameter settings).
[0042] In this embodiment, after determining that the total diffusion step size is T and obtaining the interval of the diffusion step size as [0, T - 1], then obtain the signal-to-noise ratio at each diffusion step size, and according to the two signal-to-noise ratio demarcation points 10-2 The interval [0, T - 1] of the diffusion step size is divided into three stages: early, middle, and late (excluding the two endpoints). The early stage is [τ1, T - 1], the middle stage is [τ2, τ1 - 1], and the late stage is [1, τ2 - 1]. n1 sampling points are sampled in the early stage, n2 sampling points are sampled in the middle stage, and n3 sampling points are sampled in the late stage, and n1 + n2 + n3 = j. The set of all sampling paths composed of n1 sampling points, n2 sampling points, and n3 sampling points forms a restricted solution space. In a specific implementation, the sampling paths all include the endpoints with a diffusion step size of T.
[0043] Executing the optimal path search method under the restricted solution space has higher efficiency and is more likely to converge to the optimal sampling path under the same population parameter settings.
[0044] In this embodiment, the sampled path after search is [t j , t j-1 ,..., t1], and t j = T, t1 = 0.
[0045] The optimal path search method in step S24 adopts a genetic algorithm.
[0046] In this embodiment, a single sampling path that meets the requirements of the restricted solution space is used as an individual, and a number of (50) random individuals are used as the initial population, and are gradually optimized according to the genetic algorithm to generate subsequent populations.
[0047] The following is a basic explanation of the genetic algorithm:
[0048] The genetic algorithm can gradually obtain the "optimal" individuals in a huge solution space. Compared with greedy algorithms, brute-force searches, etc., the genetic algorithm has better execution efficiency and is not easily trapped in local optimal solutions. Specifically:
[0049] Individuals and populations: An individual represents a single possible element in the search space, while a population is a group composed of a specific number of individuals. The first-generation population is initialized with 50 individuals, and each individual includes fixed diffusion step sizes 0 and T. It should be noted that in order to better define the crossover and mutation operations, we represent the individual as a binary list L of length T - 1, where "0" represents the selected step and "1" represents the unselected step. This list corresponds one-to-one with the sampling path.
[0050] Fitness function and selection: The mean absolute error (MAE) or mean squared error (MSE) on the validation set is used as the fitness function to evaluate the performance of each individual. Using the tournament selection strategy, 50 individuals with higher fitness are selected from the population as the parents of the next generation.
[0051] Mutation and Crossover: The selected individuals are randomly paired for crossover, and there is a 50% probability for each pair to perform the crossover operation. During the crossover process, the values at each position are exchanged with a probability of 0.5. Next, each individual in the new generation has a 20% chance of mutating, that is, randomly selecting a value at a position to be flipped. Finally, the newly generated 50 individuals will completely replace the original population.
[0052] Repair Operation: To meet the constraint conditions of the search space, after crossover and mutation, a repair operation is performed on each individual to ensure that the number of steps in each stage meets the specified conditions.
[0053] S3. According to the obtained sampling path, use the denoising diffusion probabilistic model (DDPM) accelerated from the perspective of the consistency model to perform signal interpolation processing on the time series sensing signal, and obtain the time series sensing signal after signal interpolation.
[0054] The obtained time series sensing signal after signal interpolation is used as relevant reference data for the maintenance of industrial equipment, thereby further reducing the production line downtime.
[0055] In specific implementation, the time series sensing signal after signal interpolation can be the interpolated vibration, temperature, pressure, etc. These time series sensing signals are used for subsequent optimization and maintenance of industrial equipment, thereby reducing the production line downtime.
[0056] The denoising diffusion probabilistic model accelerated from the perspective of the consistency model is a model that only accelerates the reverse process of the denoising diffusion probabilistic model (DDPM).
[0057] The perspective of the consistency model considers mapping the intermediate points of any diffusion trajectory to the initial points, and can achieve a trade-off between generation quality and speed through further adding noise and denoising, such as Figure 2 shown, the acceleration processing of the reverse process of the denoising diffusion probabilistic model is set according to the following formula:
[0058]
[0059] where, is the estimate of the true data according to i at the sampling point of the i-th step t , is the data with noise at the sampling point of the i-th step t i , is the coefficient at the sampling point of the i-th step t i , is the prediction of the noise at the sampling point of the i-th step t i , is the data with noise at the sampling point of the (i - 1)-th step t i-1 is the coefficient at the sampling point for the (i - 1)-th step t i-1 and obeys the standard normal distribution, ∈ is Gaussian noise, i is the sampling at the i-th step and i ~ (0, j), where j is the number of sampling points.
[0060] In specific implementation, for the diffusion step with a signal-to-noise ratio greater than or equal to 1, the calculation steps in formula (2) can be skipped, thereby reducing the uncertainty during the generation process and further improving the sample quality.
[0061] In specific implementation, when i = j, let
[0062] DDPM is a probabilistic generation method that can generate samples with inherent uncertainty. As Figure 3 shown in the schematic diagram, reducing the sampling quantity will lead to a significant decline in the sampling performance of DDPM. Due to this uncertainty, if we attempt to perform interpolation based on samples generated by a limited number of models, it usually results in a decline in interpolation performance. However, by adopting the reversed process after accelerating DDPM, the number of noise injections is reduced, thereby achieving more stable sample generation. In fact, within the range of the number of samples nsamples ∈ [10, 100], the average variance between different samples obtained by the accelerated diffusion interpolation model based on the perspective of the consistency model is several times lower than that of DDPM sampling. This stability enables satisfactory interpolation results to be achieved with a reduced total number of samples (around 10), that is, the beneficial effect of reducing the number of samples during the interpolation process is realized.
[0063] Example 2:
[0064] Implement by accelerating DDPM that originally required 100 steps to 6 steps, and the implementation steps are as follows.
[0065] D1. First, divide the stages based on the signal-to-noise ratio at each length. Here, assume that the early, middle, and late stages are 20, 40, and 40 steps respectively (in sequence). In addition to the determined endpoint 99, we also need to select a total of 5 diffusion step lengths from 1 - 98. We set n1 = 1, traverse the cases of n2 and n3, search on the validation set, and obtain the optimal sampling path, such as [99, 82, 65, 53, 32, 15, 0].
[0066] D2. Then, use the optimal sampling path searched above to perform a single generation of samples. First, we sample a pure Gaussian noise as the starting point of sampling. Based on this starting point, we first predict x0 once to obtain At this time, the most preliminary sample is obtained, but this result is still relatively rough. To further improve its performance, re-inject noise to obtain x82 ; Based on x 82 it is possible to perform the consistency prediction of x0 again to obtain This is a more refined result compared to ; Repeat the above process of adding and removing noise until when sampling reaches the late stage, it is no longer necessary to and re-add noise, but directly perform prediction based on them, and finally we obtain a sampling sample.
[0067] D3. To achieve the final imputation, repeat the sampling 10 times in step D2 to obtain 10 sampling samples. Then, based on these 10 samples, use the median aggregation method. Specifically, for the values at each position, calculate the median of these values as the representative value at that position. Obtain the final imputation result
[0068] The imputation process of this embodiment is accelerated by a total of (100 * 100) / (6 * 10) times, approximately 166.7 times.
[0069] The following further elaborates on the present invention in conjunction with the accompanying drawings and embodiments. However, the present invention is not limited thereto. For those of ordinary skill in the art of this technology, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also considered within the protection scope of the present invention. The content not described in detail in this specification belongs to the prior art well-known to those of ordinary skill in the art.
Claims
1. A method for interpolating time series sensor signals based on a consistency model perspective, characterized in that: The following steps are involved: S1. Obtaining time series sensor signals through sensors installed on industrial equipment; S2. According to the obtained time series sensor signal, the required diffusion step interval is set and the signal-to-noise ratio under each diffusion step is obtained. According to the preset two signal-to-noise ratio dividing points, the diffusion step interval is divided and sampled to obtain a restricted solution space. The obtained restricted solution space is processed by the optimal path search method to obtain the sampling path after the search. S3. According to the obtained sampling path, a denoising diffusion probability model based on consistency model perspective acceleration is used to perform signal interpolation processing on the time series sensor signal to obtain the time series sensor signal after signal interpolation.
2. The interpolation method of time series sensor signals based on the consistency model perspective according to claim 1 is characterized in that: The step S2 specifically includes the following steps: S21, determining the required total diffusion step length according to the obtained time series sensor signal, thereby obtaining the diffusion step length interval; S22, obtaining the signal-to-noise ratio under each diffusion step length, dividing the diffusion step length interval into three stages: early, middle and late according to two preset signal-to-noise ratio dividing points, and presetting the number of sampling points in each stage; S23, the set of all sampling paths formed by the sampling points of all stages forms a restricted solution space; S24. Using the optimal path search method to process the obtained restricted solution space, and obtain the sampled path after the search.
3. The interpolation method of time series sensor signals based on the consistency model perspective according to claim 2 is characterized in that: The optimal path search method in step S24 adopts a genetic algorithm.
4. The interpolation method of time series sensor signals based on the consistency model perspective according to claim 2 is characterized in that: The two signal-to-noise ratio demarcation points in step S22 are respectively -2 and the signal-to-noise ratio is equal to 1.
5. The interpolation method of time series sensor signals based on the consistency model perspective according to claim 1 is characterized in that: The denoising diffusion probability model based on consistency model perspective acceleration is a model that only accelerates the reverse process of the denoising diffusion probability model.
6. The interpolation method of time series sensor signals based on the consistency model perspective according to claim 1 is characterized in that: The acceleration process of the reverse process of the denoising diffusion probability model is set according to the following formula: in, For step t i Sampling point according to Estimates of real data. For step t i The data contains noise at the sampling point. For step t i The coefficients at the sampling point, For step t i The prediction of noise at the sampling point, For step i-1 t i-1 The data contains noise at the sampling point. For step i-1 t i-1 The coefficients at the sampling point, is the standard normal distribution, ∈ is Gaussian noise, and i is the sampling of the i-th step.