WO2022116508A1 - 基于获取并识别噪声全景分布模型的信号分析方法及系统 - Google Patents
基于获取并识别噪声全景分布模型的信号分析方法及系统 Download PDFInfo
- Publication number
- WO2022116508A1 WO2022116508A1 PCT/CN2021/099384 CN2021099384W WO2022116508A1 WO 2022116508 A1 WO2022116508 A1 WO 2022116508A1 CN 2021099384 W CN2021099384 W CN 2021099384W WO 2022116508 A1 WO2022116508 A1 WO 2022116508A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- noise
- measurement
- sample
- test sample
- model
- Prior art date
Links
- 238000004458 analytical method Methods 0.000 title claims abstract description 52
- 238000005259 measurement Methods 0.000 claims abstract description 390
- 239000000523 sample Substances 0.000 claims abstract description 210
- 238000012360 testing method Methods 0.000 claims abstract description 109
- 239000013074 reference sample Substances 0.000 claims abstract description 97
- 238000012549 training Methods 0.000 claims abstract description 82
- 238000013473 artificial intelligence Methods 0.000 claims abstract description 69
- 238000000034 method Methods 0.000 claims abstract description 43
- 238000012545 processing Methods 0.000 claims abstract description 32
- 230000008569 process Effects 0.000 claims abstract description 28
- 230000007613 environmental effect Effects 0.000 claims description 17
- 238000010606 normalization Methods 0.000 claims description 15
- 230000008878 coupling Effects 0.000 claims description 12
- 238000010168 coupling process Methods 0.000 claims description 12
- 238000005859 coupling reaction Methods 0.000 claims description 12
- 230000002123 temporal effect Effects 0.000 claims description 10
- 230000007246 mechanism Effects 0.000 claims description 9
- 238000013528 artificial neural network Methods 0.000 claims description 6
- 230000005672 electromagnetic field Effects 0.000 claims description 6
- 239000012530 fluid Substances 0.000 claims description 6
- 238000007637 random forest analysis Methods 0.000 claims description 6
- 238000004904 shortening Methods 0.000 claims description 6
- 238000012706 support-vector machine Methods 0.000 claims description 6
- 230000001629 suppression Effects 0.000 claims description 6
- 238000006073 displacement reaction Methods 0.000 claims description 3
- 230000006698 induction Effects 0.000 claims description 3
- 238000003756 stirring Methods 0.000 claims 1
- 230000009467 reduction Effects 0.000 abstract description 5
- 238000001514 detection method Methods 0.000 description 10
- 238000010276 construction Methods 0.000 description 7
- 238000005516 engineering process Methods 0.000 description 6
- 230000006870 function Effects 0.000 description 6
- 238000013178 mathematical model Methods 0.000 description 6
- 238000005070 sampling Methods 0.000 description 6
- 230000008859 change Effects 0.000 description 4
- 238000013459 approach Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 3
- 238000007796 conventional method Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000000691 measurement method Methods 0.000 description 3
- 238000000605 extraction Methods 0.000 description 2
- 239000000126 substance Substances 0.000 description 2
- 238000013019 agitation Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000019771 cognition Effects 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000012067 mathematical method Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/10—Pre-processing; Data cleansing
- G06F18/15—Statistical pre-processing, e.g. techniques for normalisation or restoring missing data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
Definitions
- the invention belongs to the technical field of signal analysis, and in particular relates to a signal analysis method and system based on acquiring and identifying a noise panoramic distribution model.
- sample measurement methods exist in the prior art, such as ECG signals and EEG signals in the field of physiological detection, and spectral signals in the field of substance detection.
- ECG signals and EEG signals in the field of physiological detection
- spectral signals in the field of substance detection.
- the measurement result must be mixed data mixed with signal and noise.
- signal-to-noise ratio it is more difficult to extract the signal, which makes it difficult to analyze the signal comprehensively and effectively, which will directly affect the accurate cognition of the sample.
- the following two technical directions for dealing with noise are given in the prior art: 1.
- the noise is controlled or suppressed by measures such as improving the equipment accuracy and the measurement environment, so that the signal strength far exceeds 2.
- a mathematical method is used to build a mathematical model based on the pre-assumed noise statistical distribution, and the mathematical model is used to remove noise to further improve the overall signal-to-noise ratio of the measurement results. .
- the noise collected in the sample measurement process and mixed in the signal may be nonlinear, and may also have a rather complex form and content.
- the noise at different positions may exist. Differences; for measurements of audio samples, the noise may be different on different tracks, or even at different times on the same track.
- the conventional mathematical noise reduction method in the existing engineering technology is difficult to carry out directly, that is, it is difficult to design the denoising scheme through one or several commonly used mathematical models, resulting in that the signal-to-noise ratio of the measurement results cannot be improved to a sufficient level. the level of analysis;
- the measurement result may have the following properties: 1.
- the signal in the measurement result can be detected, that is, the signal strength is above the measurable lower limit of the measuring device; 2.
- the signal It is extremely weak, and its intensity is the same order of magnitude as the noise intensity or even lower; 3.
- the characteristics of the signal itself are very complex. In a measurement with the above properties, the signal is likely to be overwhelmed by noise.
- the main purpose of the present invention is to provide a signal analysis method based on acquiring and identifying a noise panoramic distribution model, aiming to solve the technical problem that the prior art is difficult to analyze complex signals under the condition of ultra-low signal-to-noise ratio.
- a signal analysis method based on acquiring and identifying a noise panoramic distribution model comprising the following steps:
- S1 In a rich conditional measurement environment, repeat the measurement of the reference sample and the test sample to obtain multiple measurement results; each measurement result includes a signal and different noise profiles; where the rich condition means that the The measurement conditions that do not involve noise suppression, natural, and include real complex noise factors for the purpose of maintaining the consistency of external conditions.
- S2 Process the measurement results of the reference sample and the test sample to form training data of the reference sample and the test sample respectively; wherein the training data includes a noise panorama or at least part of a noise panorama composed of multiple noise profiles;
- S4 Input the measurement result of the sample to be identified into the trained artificial intelligence model, and the output result of the artificial intelligence model is the specific type of the sample to be identified.
- step S1 before each measurement of the reference sample and the test sample, by introducing a slight disturbance, a rich-condition measurement environment is created, thereby increasing the noise observation dimension, so that the measurement results of each measurement contain different values. noise profile.
- the slight perturbation can be selected but not limited to spatial perturbation, time perturbation, physical perturbation and environmental perturbation; spatial perturbation includes but is not limited to: slight displacement of the measurement site, slight displacement of the measurement site Rotation; temporal perturbations include, but are not limited to: increasing the measurement duration, shortening the measurement duration, and changing the time interval between multiple measurements; physical perturbations include, but are not limited to: vibrating the measurement device or sample during measurement, performing measurements on fluid samples Agitation; environmental perturbations include, but are not limited to: changing the ambient temperature during measurement, changing the ambient humidity during measurement, changing the electromagnetic field during measurement, and changing air pressure during measurement.
- step S2 the measurement results of the reference sample and the test sample are processed, and the steps of forming the training data of the reference sample and the test sample respectively include:
- step S22 Based on the normalized result of step S21, establish a posterior probability model framework
- noise profiles constitute the noise panorama or at least part of the noise panorama.
- the overall measurement results of the two types of samples, and the signals in the measurement results will be It exhibits stable statistical characteristics; the statistical distribution pattern of noise will also tend to be stable with the construction of noise panorama.
- the artificial intelligence model can be selected but not limited to: artificial neural network, perceptron, support vector machine, Bayesian classifier, Bayesian network, random forest model or clustering model.
- step S3 during the training process of the artificial intelligence model, the model will, in an iterative manner, perform a large number of analysis on the features contained in the training data that can realize signal and noise identification, and the features that can realize the distinction between reference samples and test samples. Experiential learning, induction and convergence, and learned associations between features and preset labels.
- the features that can realize signal identification include statistical distribution patterns that are presented after processing multiple measurement results and conform to the true mathematical statistical laws of the signal; the features that can realize noise identification include those constructed by diverse noise profiles.
- the preset labels include output labels and input labels.
- the output label includes two labels representing the reference sample and the test sample respectively;
- the input label is two sets of coupling labels that respectively involve the training data of the reference sample and the test sample, and each coupling label is respectively related to the location where the sample is measured.
- a rich conditional measurement environment is associated.
- each coupling label in different groups respectively represents: in each independent measurement environment in the rich-condition measurement environment, the coupling between the measurement result of the reference sample or the test sample and the noise panorama; wherein the measurement result contains The noise profile of , is the noise profile obtained in this independent measurement environment.
- the present invention also provides a signal analysis system based on acquiring and identifying the noise panoramic distribution model, comprising a measurement module, a processing module, a training module and an analysis module;
- the measurement module performs repeated measurements on the reference sample and the test sample, and obtains multiple measurement results respectively; wherein each measurement result includes a signal and different noise profiles; wherein, the rich condition means that the The measurement conditions that do not involve noise suppression, natural, and include real complex noise factors for the purpose of maintaining the consistency of external conditions.
- the processing module processes the measurement results of the reference sample and the test sample to form training data of the reference sample and the test sample respectively; wherein the training data includes a noise panorama or at least part of a noise panorama composed of multiple noise profiles;
- the training module takes the observability of noise as the convergence target, and trains the artificial intelligence model, so that the model can identify the signal and noise from the measurement results, and distinguish the reference sample and the test sample;
- the analysis module inputs it into the trained artificial intelligence model, and the output result of the artificial intelligence model is the specific type of the sample to be identified.
- the measurement module includes a perturbation mechanism, and before the measurement module performs each measurement on the reference sample and the test sample, the perturbation mechanism introduces a slight perturbation to create a condition-rich measurement environment, thereby increasing the noise observation dimension of the sample measurement, Include a different noise profile in the measurement results for each sample measurement.
- the slight perturbation introduced by the perturbation mechanism can be selected from, but not limited to, spatial perturbation, temporal perturbation, physical perturbation and environmental perturbation.
- Spatial perturbations include, but are not limited to: slightly displacing the measurement site, slightly rotating the measurement site; temporal perturbations include, but are not limited to: increasing the measurement duration, shortening the measurement duration, and changing the time interval between multiple measurements ; Physical perturbations include but are not limited to: vibrating the measurement equipment or sample, agitating the fluid sample during measurement; environmental perturbations include but are not limited to: changing the ambient temperature during measurement, changing the ambient humidity during measurement, and changing the temperature during measurement. Electromagnetic fields, and changing the air pressure during measurement.
- the processing module includes a normalization module and a posterior probability module
- the normalization module normalizes the measurement results of the reference sample and the test sample, and outputs the normalized results respectively;
- the posterior probability module establishes the posterior probability model framework based on the normalized results, and forms the required .
- the training data of reference samples and test samples are used for subsequent artificial intelligence model training.
- different noise profiles constitute the noise panorama or at least part of the noise panorama.
- the overall measurement results of the two types of samples and the measurement results The signals of , respectively, will show stable statistical characteristics; the statistical distribution pattern of noise will also tend to be stable with the construction of the noise panorama.
- the artificial intelligence model can be selected but not limited to: artificial neural network, perceptron, support vector machine, Bayesian classifier, Bayesian network, random forest model or clustering model.
- the model will iteratively conduct a large number of empirical learning and induction on the features contained in the training data that can realize signal and noise identification, as well as the features that can realize the distinction between reference samples and test samples. and convergence, and learn the association between features and preset labels.
- the features that can realize signal identification include statistical distribution patterns that are presented after processing multiple measurement results and conform to the true mathematical statistical laws of the signal; the features that can realize noise identification include those constructed by diverse noise profiles.
- the preset labels include output labels and input labels.
- the output label includes two labels representing the reference sample and the test sample respectively;
- the input label is two sets of coupling labels that respectively involve the training data of the reference sample and the test sample, and each coupling label is respectively related to the sample measurement time, The rich condition of the measurement environment is associated.
- each coupling label in different groups respectively represents: in each independent measurement environment in the rich-condition measurement environment, the coupling between the measurement result of the reference sample or the test sample and the noise panorama; wherein the measurement result contains The noise profile of , is the noise profile obtained in this independent measurement environment.
- the present invention is different from the noise processing scheme in the prior art, from a completely different technical angle, provides a kind of signal analysis method based on acquiring and identifying the noise panorama distribution model, to solve the noise reduction problem that the prior art is difficult to handle .
- the field of signal detection often involves signals that have analytical value but are buried in noise because of their very weak strength and/or their complex characteristics. In this case, the distribution model of noise cannot be reasonably assumed, which makes it difficult to effectively practice the existing noise reduction methods for mathematically modeling noise signals.
- the signal analysis method provided by the present invention is based on the principle of mathematical statistics, and does not directly separate the signal and the noise, but can still effectively distinguish the noise and the signal, and realize the successful identification of multiple independent signals based on different measurement samples, thereby developing the sample Practical applications such as detection and substance classification.
- the present invention utilizes artificial intelligence technology to perform hybrid modeling of noise and signals submerged therein. Even if the noise has no mathematical assumptions, the trained artificial intelligence model can deeply mine the hidden mathematical statistical laws in the measurement results, and accurately obtain the mathematical distribution model of the signal and noise.
- the present invention does not set consistent measurement conditions, but creates multiple noise profiles that differ due to changes in measurement conditions by creating diverse measurement conditions, and combines multiple noise profiles.
- the profiling combination constitutes a noise panorama or at least part of a noise panorama, which in turn identifies a mathematical distribution model of the noise. This operation method will not lead to loss and mistaken deletion of the signal, and avoid the influence of the denoising step commonly used in the prior art on the signal itself.
- the relatively stable and signal data distribution form can be more clearly presented, thereby improving the visibility of the signals in the measurement results, which is beneficial for subsequent signal extraction and analysis.
- the perturbation environment provided by the diversity measurement conditions provides different noise observation dimensions for each sample measurement, ensuring the randomness of the noise samples.
- the noise panorama can be obtained through a large number of repeated measurements, that is, a large number of noise "samples" can cover almost all the possibilities of the noise itself.
- the distribution model of noise will also tend to its true distribution form.
- the invention can find the mathematical statistical law of noise from the mixed data distribution form of the sample measurement results, and realize the distinction between noise and signal and the identification of different types of signals from the perspective of the data distribution model. Based on this technical idea, in the sample measurement results, the noise and the signal will show their true mathematical statistical laws respectively. Compared with the direct noise removal and signal extraction in the existing engineering technology, the present invention deeply excavates the mathematical statistical law of noise and signal, which can avoid the false elimination of the signal by the denoising operation and ensure the validity of the data. Therefore, noise does not interfere with signal analysis, nor does it affect the identification and classification of independent signals. It can be seen that the present invention provides an effective solution to the problem of removing the noise itself from the mixed sample measurement results, or extracting the signal itself, which cannot be solved by the prior art.
- the measurement results obtained by each sample measurement may be infinitely close to the real signal, but always only It can be "statistically stable” in the vicinity of the signal, and the “statistically stable” change of the measurement result and the noise in it is unpredictable, that is, it is impossible to make assumptions about the exact value of the next sample measurement result.
- the artificial intelligence model is used to deeply mine the noise data distribution model, which can obtain highly empirical and more accurate analysis results.
- the noise profile obtained can construct at least part of the noise panorama, and the data distribution model of noise has tended to reflect the theoretical real distribution model of noise.
- the present invention uses artificial intelligence technology to explore the distribution model of noise.
- the sample measurement stage involves multiple sample measurements under various perturbation conditions, thereby obtaining a large number of measurement results mixed with signal and noise.
- the above operations are beneficial to obtain the noise panorama, and the huge measurement value also provides a sufficient data basis for the training of the artificial intelligence model.
- the trained artificial intelligence model can find the real features in the high background noise data that meet the analysis needs of the experimenter or the purpose of their analysis, can provide more efficient mathematical operations, and output highly empirical and more accurate in real time. Analyze the results.
- the present invention creates a condition-rich measurement environment through different perturbation introduction methods, so as to increase the noise observation dimension in the sample measurement, so that the measurement results under a large number of repeated measurement conditions can show a complete noise panorama, or can be used for subsequent signals.
- the analysis provides at least a partial noise panorama with sufficient precision.
- the practical difficulty of different perturbation introduction methods is different, and it may also lead to different effects in increasing the dimension of noise observation.
- the experimenter can completely introduce the method from the perturbation provided by the present invention according to actual needs. to choose from.
- the diversified perturbation introduction methods disclosed in the present invention provide experimenters with a wide range of choices, and also reduce the difficulty of application of the present invention to a certain extent, making the technical solution more valuable for popularization and application.
- FIG. 1 is a schematic flowchart of a signal analysis method based on acquiring and identifying a noise panoramic distribution model provided by the present invention.
- FIG. 2 is a schematic flowchart of step 2 in the signal analysis method shown in FIG. 1 of the specification.
- FIG. 3 is a schematic structural diagram of a signal analysis system based on acquiring and identifying a noise panoramic distribution model provided by the present invention.
- FIG. 4 is a schematic diagram showing the principle of forming a noise panorama or at least part of a noise panorama by multiple noise profiles.
- the purpose of signal processing is to extract useful information from sample measurements, such as content of research value, or differential characteristics that distinguish it from other signals. Limited by many uncertain factors in the sample measurement process, these "research-worthy content” and “differential characteristics” are often not represented by independent values, but are reflected by the overall statistical distribution of the signal.
- the measurement results obtained during the actual sample measurement process must be mixed with noise in addition to the signals reflecting the real characteristics.
- the influence of noise is either eliminated from the measurement result, or the signal is extracted from the measurement result.
- the technical purpose of the present invention is to find out the "unknown" distribution model of the signal and the noise from the measurement result based on the principle of data statistics, so as to realize the effective distinction between the signal and the noise.
- the accurate identification of the sample type can be realized by finding a differentiated data distribution model.
- each sample measurement is equivalent to a random sampling in the sample population, and the measurement results corresponding to random sampling cannot reflect the real characteristics.
- the signal and noise have specific data statistics laws and conform to a specific data distribution model, when the number of measurements increases and the sampling range expands to approach the sample population, a large number of test results reflect the overall data. Statistical laws tend to reflect the real situation.
- the present invention realizes sample measurement in a perturbed environment by creating diverse measurement conditions, and generates noise profiles from multiple observation dimensions.
- the omnidirectional observation of noise can be reflected, that is, a noise panorama that can show a complete statistical distribution model of data is constructed, and this statistical distribution model of data will infinitely approach the true distribution of noise. It can be seen that from a statistical point of view, it is completely theoretically feasible to show the statistical distribution model of noise through the noise panorama construction in the perturbation environment.
- the present invention uses artificial intelligence technology to deeply explore the statistical law of noise.
- Artificial intelligence technology is an effective means for various data analysis and solving empirical data processing.
- the artificial intelligence deep learning model can simulate the human learning process and quickly summarize the human empirical data processing methods, so as to realize signal recognition and judgment behavior.
- the accuracy of the empirical analysis results output by the artificial intelligence model trained on big data can be guaranteed, so that the mathematical distribution model of noise can be effectively identified, and subsequent noise separation, signal classification, etc. can be carried out accordingly. specific analysis work.
- a signal analysis method based on acquiring and identifying a noise panorama distribution model a schematic flowchart of which is shown in Figure 1 of the description, and the method includes the following steps:
- S1 In a rich conditional measurement environment, repeat the measurement of the reference sample and the test sample to obtain multiple measurement results; each measurement result includes a signal and a different noise profile;
- the signals in the measurement results of the reference sample and the test sample will always remain statistically unchanged, but the noise will vary due to changes in the environment, that is, changes in the environment. will increase the observational dimension of the noise.
- the rich noise profile is the basis for constructing the noise panorama and identifying the noise based on the statistical laws of the data in the subsequent steps.
- S2 Process the measurement results of the reference sample and the test sample to form training data of the reference sample and the test sample respectively; wherein the training data includes a noise panorama or at least part of a noise panorama composed of multiple noise profiles;
- the noise panorama refers to that the noise distribution model can fully reflect its theoretical real distribution model;
- the partial noise panorama refers to that the noise distribution model cannot fully reflect its theoretical real distribution model, But the distribution model already has a level of precision that can be used for subsequent signal analysis.
- the training data of the reference sample and the test sample will be randomly allocated to the learning data and the detection data in a preset ratio, respectively.
- Use the learning data to train the artificial intelligence model input the detection data into the trained artificial intelligence model, and calculate the signal recognition result. If the signal recognition accuracy rate is lower than the preset threshold, use the learning data to continue training. If the recognition accuracy rate is higher than the preset threshold, it is considered that the artificial intelligence model has completed the training.
- S4 Input the measurement result of the sample to be identified into the trained artificial intelligence model, and the output result of the artificial intelligence model is the specific type of the sample to be identified.
- the reference sample and the test sample are used as two kinds of known samples, and the training data formed respectively after the multiple measurement results of the two are processed, and the trained artificial intelligence model can realize the effectiveness of the two known samples. distinguish.
- the artificial intelligence model can accurately identify the specific type of the sample to be identified.
- step S1 before each measurement of the reference sample and the test sample, by introducing a slight disturbance, a rich-condition measurement environment is created, thereby increasing the noise observation dimension, so that the measurement results of each measurement contain different values. noise profile.
- the slight perturbation introduced before each measurement of the reference sample and the test sample can be selected but not limited to spatial perturbation, temporal perturbation, physical perturbation and environmental perturbation.
- Spatial perturbations include, but are not limited to: slightly displacing the measurement site, slightly rotating the measurement site; temporal perturbations include, but are not limited to: increasing the measurement duration, shortening the measurement duration, and changing the time interval between multiple measurements ; Physical perturbations include but are not limited to: vibrating the measurement equipment or sample, agitating the fluid sample during measurement; environmental perturbations include but are not limited to: changing the ambient temperature during measurement, changing the ambient humidity during measurement, and changing the temperature during measurement. Electromagnetic fields, and changing the air pressure during measurement.
- step S2 the measurement results of the reference sample and the test sample are processed, and the steps of respectively forming the training data of the reference sample and the test sample include:
- step S22 Based on the normalized result of step S21, establish a posterior probability model framework.
- the measurement results of the reference sample and the test sample are regarded as measurement values obtained by measuring a measurement target composed of a complex system.
- Equation (3) is the normalization condition.
- step S21 is adopted to normalize the measurement results of the reference sample and the test sample.
- Equation (2) is rewritten in the form of an ensemble:
- ⁇ S is the measured information entropy
- ⁇ P is the measured environmental change.
- Equation (8) is the posterior probability condition.
- step S22 is adopted to establish a posterior probability model framework based on the normalized results obtained in step S21.
- ⁇ n * argmax ⁇ n P(H ⁇ n>
- the measurement result processed in steps S21-S22 can satisfy the normalization condition of formula (3) and the posterior probability condition of formula (8).
- the measurement results satisfying the above two conditions can be used to realize the estimation of the statistical fluctuation of complex systems in Eq. (9).
- the measurement results that meet the above two conditions will be used as training data in the subsequent artificial intelligence model training steps.
- noise profiles constitute the noise panorama or at least part of the noise panorama.
- the overall measurement results of the two types of samples, and the signals in the measurement results will be It exhibits stable statistical characteristics; the statistical distribution pattern of noise will also tend to be stable with the construction of noise panorama.
- the artificial intelligence model can be selected but not limited to: artificial neural network, perceptron, support vector machine, Bayesian classifier, Bayesian network, random forest model or clustering model.
- the estimation of the statistical fluctuation of the complex system in the above formula (9) will be realized by an artificial intelligence model.
- a signal analysis system based on acquiring and identifying a noise panoramic distribution model the schematic diagram of which is shown in Figure 3 of the description, the system includes a measurement module 1, a processing module 2, a training module 3 and an analysis module 4;
- the measurement module 1 performs repeated measurements on the reference sample and the test sample, and obtains multiple measurement results respectively; wherein, each measurement result includes a signal and a different noise profile;
- the measurement module 1 performs repeated measurements of the reference sample and the test sample, which will form a rich noise profile.
- the rich noise profile is the basis for the subsequent construction of noise panorama and the identification of noise based on statistical laws of data.
- the processing module 2 processes the measurement results of the reference sample and the test sample to form training data of the reference sample and the test sample respectively; wherein the training data includes a noise panorama or at least part of a noise panorama composed of multiple noise profiles;
- the noise panorama refers to that the noise distribution model can fully reflect its theoretical real distribution model;
- the partial noise panorama refers to that the noise distribution model cannot fully reflect its theoretical real distribution model, But the distribution model already has a level of precision that can be used for subsequent signal analysis.
- the training module 3 takes the observability of noise as the convergence target, and trains the artificial intelligence model, so that the model can identify the signal and noise from the measurement results, and distinguish the reference sample and the test sample. ;
- the training data of the reference sample and the test sample will be randomly allocated to the learning data and the detection data in a preset ratio, respectively.
- the training module 3 uses the learning data to train the artificial intelligence model, and inputs the detection data into the trained artificial intelligence model, and calculates the signal identification result. If the signal identification accuracy rate is lower than the preset threshold, the learning data is used to continue the training. , if the signal recognition accuracy rate is higher than the preset threshold, the artificial intelligence model is deemed to have completed the training.
- the analysis module 4 inputs it into the trained artificial intelligence model, and the output result of the artificial intelligence model is the specific type of the sample to be identified.
- the reference sample and the test sample are used as two kinds of known samples, and the training data formed respectively after the multiple measurement results of the two are processed, and the trained artificial intelligence model can realize the effectiveness of the two known samples. distinguish.
- the artificial intelligence model can accurately identify the specific type of the sample to be identified.
- the measurement module 1 includes a perturbation mechanism 11. Before the measurement module 1 performs each measurement on the reference sample and the test sample, the perturbation mechanism 11 introduces a slight perturbation to create a rich-condition measurement environment, thereby increasing the sample measurement efficiency. Noise observation dimension, so that a different noise profile is included in the measurement for each sample measurement.
- the slight disturbance introduced by the perturbation mechanism 11 may be selected from, but not limited to, spatial perturbation, temporal perturbation, physical perturbation and environmental perturbation.
- Spatial perturbations include, but are not limited to: slightly displacing the measurement site, slightly rotating the measurement site; temporal perturbations include, but are not limited to: increasing the measurement duration, shortening the measurement duration, and changing the time interval between multiple measurements ; Physical perturbations include but are not limited to: vibrating the measurement equipment or sample, agitating the fluid sample during measurement; environmental perturbations include but are not limited to: changing the ambient temperature during measurement, changing the ambient humidity during measurement, and changing the temperature during measurement. Electromagnetic fields, and changing the air pressure during measurement.
- processing module 2 includes a normalization module 21 and a posterior probability module 22;
- the normalization module 21 normalizes the measurement results of the reference sample and the test sample, and outputs the normalized results respectively;
- the posterior probability module 22 establishes a posterior probability model framework based on the normalized results, and forms a The required training data of reference samples and test samples are used for subsequent artificial intelligence model training.
- the measurement results of the reference sample and the test sample are regarded as measurement values obtained by measuring a measurement target composed of a complex system.
- Equation (3) is the normalization condition.
- the normalization module 21 normalizes the measurement results of the reference sample and the test sample, and outputs the normalization process. unification result;
- Equation (2) is rewritten in the form of an ensemble:
- ⁇ S is the measured information entropy
- ⁇ P is the measured environmental change.
- Equation (8) is the posterior probability condition.
- the posterior probability module 22 in order to make the measurement results of the reference sample and the test sample satisfy the posterior probability condition of equation (8), the posterior probability module 22 establishes a posterior probability model framework based on the normalized results.
- ⁇ n * argmax ⁇ n P(H ⁇ n>
- the measurement result processed by the normalization module 21 and the posterior probability module 22 can satisfy the normalization condition of formula (3) and the posterior probability condition of formula (8).
- the measurement results satisfying the above two conditions can be used to realize the estimation of the statistical fluctuation of the complex system in Eq. (9).
- the measurement results that meet the above two conditions will be used as training data in the subsequent artificial intelligence model training steps.
- different noise profiles constitute the noise panorama or at least part of the noise panorama.
- the overall measurement results of the two types of samples and the measurement results will show stable statistical characteristics respectively; the statistical distribution pattern of noise will also tend to be stable with the construction of the noise panorama.
- the artificial intelligence model can be selected but not limited to: artificial neural network, perceptron, support vector machine, Bayesian classifier, Bayesian network, random forest model or clustering model.
- the estimation of the statistical fluctuation of the complex system in the above formula (9) will be realized by an artificial intelligence model.
- a signal analysis method based on acquiring and identifying a noise panoramic distribution model comprising the following steps:
- S1 In a rich conditional measurement environment, repeat the measurement on a variety of known samples to obtain multiple measurement results; each measurement result contains a signal and a different noise profile;
- rich conditions refer to: rich conditions refer to measurement conditions that are not aimed at maintaining the consistency of external conditions, do not involve noise suppression, and are natural and include real complex noise factors.
- the purpose of "Rich Conditional Measurement Environment” and “Repeated Measurements” is to obtain rich noise profiles that are sufficient to construct a noise panorama.
- the signal in the measurement results of each sample will always remain statistically unchanged, but the noise will vary due to changes in the environment, that is, changes in the environment will increase The observed dimension of the noise.
- the rich noise profile is the basis for constructing the noise panorama and identifying the noise based on the statistical laws of the data in the subsequent steps.
- S2 Process the measurement results of multiple known samples to form training data for each known sample respectively; wherein the training data includes a noise panorama or at least part of a noise panorama composed of multiple noise profiles;
- the noise panorama refers to that the noise distribution model can fully reflect its theoretical real distribution model;
- the partial noise panorama refers to that the noise distribution model cannot fully reflect its theoretical real distribution model, But the distribution model already has a level of precision that can be used for subsequent signal analysis.
- the training data of each known sample will be randomly allocated to learning data and detection data in a preset ratio.
- Use the learning data to train the artificial intelligence model, input the detection data into the trained artificial intelligence model, and calculate the signal recognition result. If the signal recognition accuracy rate is lower than the preset threshold, use the learning data to continue training. If the recognition accuracy rate is higher than the preset threshold, it is considered that the artificial intelligence model has completed the training.
- S4 Input the measurement result of the sample to be identified into the trained artificial intelligence model, and the output result of the artificial intelligence model is the specific type of the sample to be identified.
- the training data formed by processing multiple measurement results of multiple known samples, and the trained artificial intelligence model can effectively distinguish each known sample.
- the sample to be identified is one of a variety of known samples
- the artificial intelligence model can accurately identify the specific type of the sample to be identified.
- step S1 before each measurement of each known sample, by introducing a slight disturbance, a rich conditional measurement environment is created, thereby increasing the noise observation dimension, so that the measurement results of each measurement contain different noise profile.
- the slight perturbation introduced before each measurement of each known sample can be selected but not limited to spatial perturbation, temporal perturbation, physical perturbation and environmental perturbation.
- Spatial perturbations include, but are not limited to: slightly displacing the measurement site, slightly rotating the measurement site; temporal perturbations include, but are not limited to: increasing the measurement duration, shortening the measurement duration, and changing the time interval between multiple measurements ; Physical perturbations include but are not limited to: vibrating the measurement equipment or sample, agitating the fluid sample during measurement; environmental perturbations include but are not limited to: changing the ambient temperature during measurement, changing the ambient humidity during measurement, and changing the temperature during measurement. Electromagnetic fields, and changing the air pressure during measurement.
- step S2 the measurement results of a variety of known samples are processed, and the steps of forming training data for each known sample respectively include:
- step S22 Based on the normalized result of step S21, establish a posterior probability model framework.
- the measurement result of each known sample is regarded as a measurement value obtained by measuring a measurement target composed of a complex system.
- Equation (3) is the normalization condition.
- step S21 is adopted to normalize the measurement result of each known sample.
- ⁇ S is the measured information entropy
- ⁇ P is the measured environmental change.
- Equation (8) is the posterior probability condition.
- step S22 is adopted to establish a posterior probability model framework based on the normalized result obtained in step S21.
- ⁇ n * argmax ⁇ n P(H ⁇ n>
- the measurement result processed in steps S21-S22 can satisfy the normalization condition of formula (3) and the posterior probability condition of formula (8).
- the measurement results satisfying the above two conditions can be used to realize the estimation of the statistical fluctuation of the complex system in Eq. (9).
- the measurement results that meet the above two conditions will be used as training data in the subsequent artificial intelligence model training steps.
- the different noise profiles constitute the noise panorama or at least part of the noise panorama, and at the same time, the overall measurement results of each known sample, and the signals in the measurement results, both will show stable statistical characteristics respectively; the statistical distribution pattern of noise will also tend to be stable with the construction of noise panorama.
- the artificial intelligence model can be selected but not limited to: artificial neural network, perceptron, support vector machine, Bayesian classifier, Bayesian network, random forest model or clustering model.
- the estimation of the statistical fluctuation of the complex system in the above formula (9) will be realized by an artificial intelligence model.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Probability & Statistics with Applications (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Computational Mathematics (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Physics (AREA)
- Operations Research (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Software Systems (AREA)
- Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
- Image Processing (AREA)
Abstract
一种基于获取并识别噪声全景分布模型的信号分析方法及系统，涉及信号分析技术领域，所述方法包括如下步骤：在富条件测量环境下，对参照样本和测试样本进行重复测量，分别获得多个测量结果；对参照样本和测试样本的测量结果进行处理，分别形成参照样本和测试样本的训练数据；基于参照样本和测试样本的训练数据，使模型能够从测量结果中识别出信号与噪声，并区分参照样本和测试样本；将待识别样本的测量结果输入经训练的人工智能模型，输出结果为该待识别样本的具体类型。所述方法区别于现有技术中的噪声处理方案，从完全不同的技术角度，提供了一种基于获取并识别噪声全景分布模型的信号分析方法，以解决现有技术难以处理的降噪问题。
Description
本发明属于信号分析技术领域，尤其涉及一种基于获取并识别噪声全景分布模型的信号分析方法及系统。
应对于诸多实际领域的应用与需求，现有技术中已存在种类繁多的样本测量手段，例如生理检测领域的心电信号、脑电信号，物质检测领域的光谱信号等。然而，受限于测量环境、设备精度及样本自身属性等多元影响，无论选择怎样的样本测量手段，测量结果必然是混杂着信号与噪声的混合数据。对于低信噪比的测量结果而言，从中提取出信号的难度较大，导致难以对信号进行全面、有效的分析，这将直接影响对于样本的准确认知。
为了解决上述问题，现有技术中给出如下两个处理噪声的技术方向：1.在样本测量阶段，通过提高设备精度、改善测量环境等措施，对噪声进行控制或抑制，使信号强度远超噪声强度，从而获取高信噪比的测量结果；2.在结果分析阶段，采用数学方法，基于预先假设的噪声统计分布构建数学模型，利用数学模型去除噪声，进一步提高测量结果的整体信噪比。
上述两种方式可以一定程度上解决部分情况下的噪声问题。但始终存在两个难以克服的缺陷：
首先，样本测量过程中采集到的、夹杂在信号中的噪声可能是非线性的，还可能具有相当复杂的形式及内容，例如：对于图像样本的测量结果而言，其不同位置处的噪声可能存在差异；对于音频样本的测量结果而言，其不同音轨上、甚至同一音轨上不同时刻的噪声都可能不同。面对这种复杂噪声，现有工程技术中常规的数学降噪方法难以直接进行，即，难以通过一个或者几个常用数学模型进行去噪方案设计，导致测量结果的信噪比无法提高到具备分析水平的程度；
其次，在实际样本测量过程中，为了获取高信噪比的测量结果，可能会对样本测量环境、设备精度等条件进行全面优化。但即便如此，受限于样本自身属性 及其他客观因素，测量结果也可能具有如下属性：1.测量结果中的信号可以被检测到，即信号强度在测量设备的可测下限以上；2.信号极为微弱，其强度与噪声强度为同一量级甚至更低；3.信号自身的特征十分复杂。在具备以上属性的测量结果中，信号极大可能会被噪声所淹没。常规的数学降噪方法在处理这类测量结果时存在困难，因为难以建立合理的数学模型，对测量结果中混杂的噪声进行模拟、去除，使信号与噪声难以剥离，甚至导致高价值信号在去噪处理中随噪声一起被消除掉。
发明内容
本发明的主要目的在于提供一种基于获取并识别噪声全景分布模型的信号分析方法，旨在解决现有技术难以分析超低信噪比条件下的复杂信号的技术问题。
为实现上述目的，本申请的技术方案如下：
一种基于获取并识别噪声全景分布模型的信号分析方法，包括如下步骤：
S1：在富条件测量环境下，对参照样本和测试样本进行重复测量，分别获得多个测量结果；其中，每个测量结果均包含信号和不同的噪声侧写；其中，富条件是指不以保持外部条件一致性为目的的、不涉及抑制噪声的、自然的、包括真实的复杂噪声因素的测量条件。
S2：对参照样本和测试样本的测量结果进行处理，分别形成参照样本和测试样本的训练数据；其中，训练数据包括由多个噪声侧写构成的噪声全景或至少部分噪声全景；
S3：基于参照样本和测试样本的训练数据，以噪声的可观测性呈现为收敛目标，进行人工智能模型训练，使模型能够从测量结果中识别出信号与噪声，并区分参照样本和测试样本；
S4：将待识别样本的测量结果输入经训练的人工智能模型，人工智能模型的输出结果为该待识别样本的具体类型。
可选择地，在步骤S1中，在参照样本和测试样本的每次测量之前，通过引入轻微扰动，创造出富条件测量环境，由此增加噪声观测维度，使每次测量的测量结果中包含不同的噪声侧写。
进一步地，所述轻微扰动可以选择但不限于空间微扰、时间微扰、物理微扰及环境微扰；空间微扰包括但不限于：使测量位点发生轻微位移、使测量位点发 生轻微旋转；时间微扰包括但不限于：增加测量时长、缩短测量时长、以及改变多次测量之间的时间间隔；物理微扰包括但不限于：在测量时震动测量设备或样本、对流质样本进行搅动；环境微扰包括但不限于：改变测量时的环境温度、改变测量时的环境湿度、改变测量时的电磁场、以及改变测量时的气压。
进一步地，在步骤S2中，对参照样本和测试样本的测量结果进行处理，分别形成参照样本和测试样本的训练数据的步骤包括：
S21.对参照样本和测试样本的测量结果进行归一化处理；
S22.基于步骤S21的归一化结果，建立后验概率模型框架；
参照样本和测试样本的测量结果，经过步骤S21-22处理后，将分别形成符合要求的训练数据，用于后续的人工智能模型训练。
在参照样本和测试样本的测量结果形成训练数据的过程中，不同的噪声侧写构成噪声全景或至少部分噪声全景，同时，两类样本的整体测量结果、以及测量结果中的信号，都将分别呈现出稳定的统计特性；噪声呈现的统计分布模式，也将随着噪声全景的构建而趋于稳定。
在步骤S3中，人工智能模型可以选择但不限于：人工神经网络、感知机、支持向量机、贝叶斯分类器、贝叶斯网、随机森林模型或聚类模型。
在步骤S3中，在人工智能模型的训练过程中，模型将以迭代的方式，对训练数据内含的能够实现信号与噪声识别的特征，以及能够实现参照样本和测试样本区分的特征，进行大量经验性学习、归纳与收敛，并习得特征与预设标签之间的联系。
具体而言，能够实现信号识别的特征，包括多个测量结果经处理后呈现的、符合信号真实数学统计规律的统计分布模式；能够实现噪声识别的特征，包括由多样化的噪声侧写构建的噪声全景或至少部分噪声全景所呈现的、趋近于噪声真实数学统计规律的统计分布模式；能够实现参照样本和测试样本区分的特征，包括参照样本和测试样本的多个测量结果经处理后、分别呈现出的统计分布模式。
进一步地，预设标签包括输出标签与输入标签。其中，输出标签包括分别代表参照样本和测试样本的两个标签；输入标签是分别涉及参照样本和测试样本的训练数据的两组耦合性标签，每个耦合性标签分别与样本测量时所处的富条件测量环境相关联。
具体而言，不同组别的每个耦合性标签分别代表：在富条件测量环境中的每个独立的测量环境下，参照样本或测试样本的测量结果与噪声全景的耦合；其中，测量结果包含的噪声侧写，是这个独立的测量环境下获取的噪声侧写。
本发明还提供了一种基于获取并识别噪声全景分布模型的信号分析系统，包括测量模块、处理模块、训练模块与分析模块；
在富条件测量环境下，测量模块对参照样本和测试样本进行重复测量，分别获得多个测量结果；其中，每个测量结果均包含信号和不同的噪声侧写；其中，富条件是指不以保持外部条件一致性为目的的、不涉及抑制噪声的、自然的、包括真实的复杂噪声因素的测量条件。
处理模块对参照样本和测试样本的测量结果进行处理，分别形成参照样本和测试样本的训练数据；其中，训练数据包括由多个噪声侧写构成的噪声全景或至少部分噪声全景；
基于参照样本和测试样本的训练数据，训练模块以噪声的可观测性呈现为收敛目标，进行人工智能模型训练，使模型能够从测量结果中识别出信号与噪声，并区分参照样本和测试样本；
针对待识别样本的测量结果，分析模块将其输入经训练的人工智能模型，人工智能模型的输出结果为该待识别样本的具体类型。
可选择地，测量模块包括微扰机构，在测量模块对参照样本和测试样本进行每次测量之前，微扰机构引入轻微扰动，创造出富条件测量环境，由此增加样本测量的噪声观测维度，使每次样本测量的测量结果中包含不同的噪声侧写。
进一步地，在参照样本和测试样本的每次测量前，微扰机构引入的轻微扰动，可以选择但不限于空间微扰、时间微扰、物理微扰及环境微扰。
空间微扰包括但不限于：使测量位点发生轻微位移、使测量位点发生轻微旋转；时间微扰包括但不限于：增加测量时长、缩短测量时长、以及改变多次测量之间的时间间隔；物理微扰包括但不限于：在测量时震动测量设备或样本、对流质样本进行搅动；环境微扰包括但不限于：改变测量时的环境温度、改变测量时的环境湿度、改变测量时的电磁场、以及改变测量时的气压。
进一步地，处理模块包括归一化模块和后验概率模块；
其中，归一化模块对参照样本和测试样本的测量结果进行归一化处理，分别 输出归一化结果；后验概率模块基于归一化结果，建立后验概率模型框架，分别形成符合要求的、参照样本和测试样本的训练数据，用于后续的人工智能模型训练。
在处理模块对参照样本和测试样本的测量结果进行处理、形成训练数据的过程中，不同的噪声侧写构成噪声全景或至少部分噪声全景，同时，两类样本的整体测量结果、以及测量结果中的信号，都将分别呈现出稳定的统计特性；噪声呈现的统计分布模式，也将随着噪声全景的构建而趋于稳定。
进一步地，人工智能模型可以选择但不限于：人工神经网络、感知机、支持向量机、贝叶斯分类器、贝叶斯网、随机森林模型或聚类模型。
在人工智能模型的训练过程中，模型将以迭代的方式，对训练数据内含的能够实现信号与噪声识别的特征，以及能够实现参照样本和测试样本区分的特征，进行大量经验性学习、归纳与收敛，并习得特征与预设标签之间的联系。
具体而言，能够实现信号识别的特征，包括多个测量结果经处理后呈现的、符合信号真实数学统计规律的统计分布模式；能够实现噪声识别的特征，包括由多样化的噪声侧写构建的噪声全景或至少部分噪声全景所呈现的、趋近于噪声真实数学统计规律的统计分布模式；能够实现参照样本和测试样本区分的特征，包括参照样本和测试样本的多个测量结果经处理后、分别呈现出的统计分布模式。
进一步地，预设标签包括输出标签与输入标签。其中，输出标签包括分别代表参照样本和测试样本的两个标签；输入标签是分别涉及参照样本和测试样本的训练数据的两组耦合性标签，每个耦合性标签分别与样本测量时、所处的富条件测量环境相关联。
具体而言，不同组别的每个耦合性标签分别代表：在富条件测量环境中的每个独立的测量环境下，参照样本或测试样本的测量结果与噪声全景的耦合；其中，测量结果包含的噪声侧写，是这个独立的测量环境下获取的噪声侧写。
本申请的有益效果为：
1.本发明区别于现有技术中的噪声处理方案，从完全不同的技术角度，提供了一种基于获取并识别噪声全景分布模型的信号分析方法，以解决现有技术难以处理的降噪问题。
信号检测领域经常涉及具备分析价值、但因自身强度极为微弱和/或自身特 征极为复杂，从而淹没在噪声里的信号。在这种情况下，噪声的分布模型无法被合理假设，导致现有的、对噪声信号进行数学建模的降噪方法难以有效实践。
本发明提供的信号分析方法基于数学统计原理，并不直接进行信号与噪声的剥离，但仍可以有效区分噪声与信号，并实现基于不同测量样本的多个独立信号的成功识别，由此开展样本检测、物质分类等实际应用。此外，本发明利用人工智能技术，对噪声及淹没于其中的信号进行混合建模。即使噪声不存在数学假设，经训练的人工智能模型也可以深度挖掘测量结果中隐藏的数学统计规律，准确获取信号与噪声的数学分布模型。
2.本发明在样本测量阶段，并未对测量条件进行一致性设定，而是通过营造多样性的测量条件，形成多个因测量条件变化而产生差异的噪声侧写，并将多个噪声侧写结合构成噪声全景或至少部分噪声全景，继而识别噪声的数学分布模型。该操作手段不会导致信号的丢失与误删，避免现有技术中常用的去噪步骤对信号自身的影响。
在本发明公开的技术方案中，通过多样性测量条件下的大量重复测量，使相对稳定的、信号的数据分布形式能够得到较为明确的呈现，由此改善了测量结果中信号的可见性，有益于后续的信号提取与分析。另一方面，多样性测量条件提供的微扰环境，为每次样本测量提供不同的噪声观测维度，确保了噪声的样本随机性。在此基础上，通过大量重复测量能够获取噪声全景，即大量的噪声“样本”能够近乎全面地覆盖噪声自身的所有可能性。同时，噪声的分布模型也将趋于其真实的分布形式。
本发明能够从样本测量结果的混合数据分布形式中，发现噪声的数学统计规律，从数据分布模型的角度实现噪声和信号的区分、以及不同类型的信号的识别。基于该技术思路，在样本测量结果中，噪声与信号将分别呈现出其真实的数学统计规律。相较于现有工程技术中直接进行的噪声去除及信号提取，本发明对噪声及信号的数学统计规律进行深度发掘，能够避免去噪操作对信号的误消除，保证数据的有效性。因此，噪声对信号分析不会造成干扰，也不会影响独立信号间的识别分类。由此可见，对于现有技术无法解决的、从混杂的样本测量结果中去除噪声本身，或提取出信号本身的难题，本发明提供了一种有效的解决方案。
3.由于噪声是实际样本测量过程中无法避免的影响因素，因此，即使在现阶 段最为优异的样本测量条件下，每次样本测量获得的测量结果可能会无限地接近真实的信号，但始终只能在信号附近“统计稳定”地变化着，且这种测量结果及其中噪声的“统计稳定”的变化是具有不可预见性的，即无法对下一次样本测量结果的准确值进行假设。
然而，在多次重复采集后，大量的测量结果将整体呈现出趋于稳定的数据分布模型。这个稳定的数据分布模型，代表了测量结果中所有成分相互影响的宏观集合。即，除了真实的信号外，诸如环境复杂度、设备精度、采样手段固有影响等可能引发噪声的干扰因素，均被融入了上述测量结果的整体分布模型中。因此，该测量结果的整体分布模型能够充分反映其自身特征。在测量结果的分布模型趋于稳定的同时，由大量噪声侧写构成的噪声全景或至少部分噪声全景，也将呈现出特有的数学统计规律，并趋近于噪声真实的分布模型。本发明通过识别完整的噪声数学模型，实现噪声与信号的区分，该识别方案将获取更为全面、准确的识别结果。
4.采用人工智能模型对噪声的数据分布模型进行深度挖掘，能够得到高度经验性的、更为精准的分析结果。
多样性测量条件下开展的大量的样本重复测量，其获取的噪声侧写能够至少构建出部分噪声全景，且噪声的数据分布模型已经趋于能够反映噪声理论上的真实分布模型。在此情况下，本发明采用人工智能技术对噪声的分布模型进行发掘。
本发明公开的技术方案中，样本测量阶段涉及在多样化微扰条件下进行的多次的样本测量，由此获取了大量的、混合着信号与噪声的测量结果。上述操作在有利于获取噪声全景的同时，庞大的测量值也为人工智能模型的训练提供了足够的数据基础。经训练的人工智能模型能够在高背景噪声数据中发现符合实验者分析需求或满足其分析目的的真实特征，能够提供更为高效的数学运算，并实时地输出高度经验性的、更为精准的分析结果。
5.本发明通过不同的微扰引入手段营造出富条件测量环境，以增加样本测量中的噪声观测维度，使大量重复测量条件下的测量结果能够显现出完整的噪声全景，或可以为后续信号分析提供足够精准程度的至少部分噪声全景。不同的微扰引入手段的实践难度存在差异，在增加噪声观测维度方面也可能导致不同影响。
在本发明涉及的技术方案的实际应用中，出于样本自身特性、样本测量手段、 测量精度要求等多方面因素的综合考虑，实验者完全可以根据实际需求、从本发明提供的微扰引入手段中进行选择。本发明公开的多样化的微扰引入手段为实验者提供了广阔的选择余地，也在一定程度上降低了本发明的适用难度，使该技术方案更加具有推广应用价值。
图1为本发明提供的一种基于获取并识别噪声全景分布模型的信号分析方法的流程示意图。
图2为说明书附图1所示的信号分析方法中步骤2的流程示意图。
图3为本发明提供的一种基于获取并识别噪声全景分布模型的信号分析系统的结构示意图。
图4为多个噪声侧写构成噪声全景或至少部分噪声全景的原理示意图。
信号处理的目的是从样本测量结果中抽取有用信息，例如，具备研究价值的内容，或者是区别于其他信号的差异特征。受限于样本测量过程中的诸多不确定因素，这些“具备研究价值的内容”和“差异特征”往往无法被独立的数值所代表，而是由信号整体的数据统计分布所体现。
实际样本测量过程中获取的测量结果，其中除了反映真实特性的信号之外，必然混合着噪声。现有技术公开的信号处理方案中，或是从测量结果中消除噪声的影响，或是从测量结果中提取出信号。然而，对于测量结果中混杂的噪声无法通过“已知”的数学模型进行模拟时，消除噪声或者提取信号都具有极大的难度。所以，本发明的技术目的在于：基于数据统计原理，从测量结果中找出信号和噪声的“未知”的分布模型，从而实现信号与噪声的有效区分。此外，当测量结果来源于不同的测量样本时，通过找出差异化的数据分布模型，实现样本类型的准确识别。
对于每次样本测量而言，获取的测量结果中的信号和噪声都会与以往测量获取的结果略有差异。从样本与抽样的角度对此进行解释，每次的样本测量都相当于是在样本总体中进行一次随机抽样，随机抽样对应的测量结果是无法反映真实特征的。然而，在信号和噪声分别具有特定的数据统计规律、符合特定的数据分布模型的前提下，当测量次数增加，采样范围扩大到趋近于样本总体的时候，大 量的测试结果反映出的整体数据统计规律才能趋于反映出真实情况。
更具体地，1.对于测量结果中的信号而言，由于信号反映了测试样本自身的固有特性，其必然具有明确的统计分布模型。这个明确的统计分布模型可以通过大量的数据采样而清晰地呈现出来；2.对于测量结果中的噪声而言，现有技术通常认为噪声的“理想”数学统计规律符合高斯分布。然而，实际的样本测量过程通常无法营造出“理想”的噪声情况。此外，即使通过提高设备精度、改善材料纯度等手段，将样本测量条件尽可能优化，仍可能获取不到具备理想分析条件的测量结果。即，待分析的目标信号因强度微弱而淹没于噪声中，或是信号的特征极为复杂难以分析。
对这类测量结果进行数据分布模型的挖掘，往往存在较大难度，甚至完全无法假设其中噪声的分布模型。在这样的情况下，本发明通过营造多样化的测量条件，实现了微扰环境下的样本测量，并从多个观测维度生成噪声侧写。当样本测量次数足够多时，便能够体现对于噪声的全方位观测，即构建出了能够显现出完整数据统计分布模型的噪声全景，且这个数据统计分布模型会无限趋近于噪声的真实分布。由此可见，从统计学的角度，通过微扰环境下的噪声全景构造，显现出噪声的统计分布模型，是完全理论可行的。
在获取了至少部分噪声全景，且噪声已呈现出趋于清晰、稳定的分布模型的情况下，本发明采用人工智能技术对噪声的统计规律进行深度发掘。人工智能技术是适用于各种数据分析、并解决经验性数据处理的有效手段。例如，人工智能深度学习模型能够模拟人类的学习过程，快速总结人类经验性的数据处理办法，从而实现信号识别与判断行为。在本发明中，经大数据训练后的人工智能模型，其输出的经验性分析结果的精确性是可以保证的，从而有效识别噪声的数学分布模型，并据此开展后续噪声分离、信号分类等具体分析工作。
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合实施例，对本发明进行进一步详细说明。应当理解，此处所描述的具体实施例仅用于解释本发明，并不用于限定本发明。
实施例1
一种基于获取并识别噪声全景分布模型的信号分析方法，其流程示意图参见说明书附图1，该方法包括如下步骤：
S1：在富条件测量环境下，对参照样本和测试样本进行重复测量，分别获得多个测量结果；其中，每个测量结果均包含信号和不同的噪声侧写；
在信号采集与分析的技术领域，保持样本测量过程中的外部条件一致性，是一种减小噪声波动、形成良好信噪比的常规手段，且重复进行样本测量也被认定是减少随机误差的有效方式。然而，在本发明实施例中，并不涉及外部条件的一致性设定，而是在富条件测量环境下，开展参照样本和测试样本的重复测量。其中，富条件是指：富条件是指不以保持外部条件一致性为目的的、不涉及抑制噪声的、自然的、包括真实的复杂噪声因素的测量条件。“富条件测量环境”与“重复测量”的目的都是为了获取足够构建噪声全景的、丰富的噪声侧写。
具体而言，在富条件测量环境下，受限于样本自身属性，参照样本与测试样本的测量结果中的信号将始终保持统计不变，但噪声会因环境变化而出现差别，即环境的变化会增加噪声的观测维度。基于多方位、多角度及多时空特点的噪声观测维度，进行参照样本和测试样本的重复测量，将形成丰富的噪声侧写。丰富的噪声侧写是后续步骤中构建噪声全景、基于数据统计规律识别噪声的基础。
S2：对参照样本和测试样本的测量结果进行处理，分别形成参照样本和测试样本的训练数据；其中，训练数据包括由多个噪声侧写构成的噪声全景或至少部分噪声全景；
如说明书附图4所示，在单一的噪声观测维度下，只能获取反映噪声局部的噪声侧写，无法获取全面的噪声观测结果，即噪声无法呈现出完整的、符合其真实分布特性的数据统计规律。然而，在本发明实施例中，参照样本与测试样本的重复测量在富条件测量环境下进行。不同噪声观测维度下获取的丰富的噪声侧写，将足够构建噪声全景或至少部分噪声全景。在构建噪声全景的同时，噪声的数据统计规律将趋于其真实的数学统计规律。
在本发明实施例中，噪声全景是指：噪声的分布模型已经能够全面反映其理论上的真实分布模型；部分噪声全景是指：噪声的分布模型无法完整地反映其理论上的真实分布模型，但分布模型已经具备能够用于后续信号分析的精准程度。
S3：基于参照样本和测试样本的训练数据，以噪声的可观测性呈现为收敛目标，进行人工智能模型训练，使模型能够从测量结果中识别出信号与噪声，并区分参照样本和测试样本；
在本发明实施例中，参照样本和测试样本的训练数据将分别以预设比例，随机分配为学习数据和检测数据。采用学习数据对人工智能模型进行训练，并将检测数据输入经训练的人工智能模型中，计算得出信号识别结果，若信号识别准确率低于预设阈值，采用学习数据继续进行训练，若信号识别准确率高于预设阈值，视为人工智能模型完成训练。
S4：将待识别样本的测量结果输入经训练的人工智能模型，人工智能模型的输出结果为该待识别样本的具体类型。
在本发明实施例中，参照样本和测试样本作为两种已知样本，二者的多个测量结果经处理后分别形成的训练数据，训练出的人工智能模型能够实现两种已知样本的有效区分。当待识别样本是两种已知样本中的某一种时，人工智能模型能够对待识别样本的具体类型进行准确识别。
可选择地，在步骤S1中，在参照样本和测试样本的每次测量之前，通过引入轻微扰动，创造出富条件测量环境，由此增加噪声观测维度，使每次测量的测量结果中包含不同的噪声侧写。
进一步地，在参照样本和测试样本的每次测量前引入的轻微扰动，可以选择但不限于空间微扰、时间微扰、物理微扰及环境微扰。
空间微扰包括但不限于：使测量位点发生轻微位移、使测量位点发生轻微旋转；时间微扰包括但不限于：增加测量时长、缩短测量时长、以及改变多次测量之间的时间间隔；物理微扰包括但不限于：在测量时震动测量设备或样本、对流质样本进行搅动；环境微扰包括但不限于：改变测量时的环境温度、改变测量时的环境湿度、改变测量时的电磁场、以及改变测量时的气压。
参见说明书附图2，在步骤S2中，对参照样本和测试样本的测量结果进行处理，分别形成参照样本和测试样本的训练数据的步骤包括：
S21.对参照样本和测试样本的测量结果进行归一化处理；
S22.基于步骤S21的归一化结果，建立后验概率模型框架。
参照样本和测试样本的测量结果，经过步骤S21-22处理后，将分别形成符合要求的训练数据，用于后续的人工智能模型训练。
具体而言，在本发明实施例中，对于参照样本和测试样本的测量结果，将其视为对复杂体系构成的测量目标进行测量而得到的测量值。
定义测量密度函数为
其中，S为测量空间维度；V为测量环境；则测量目标中，体系数目为N，N由式(1)定义：
定义B(V)为测量函数，则测量值
有：
其中，
式(3)为归一化条件。本发明实施例中，为了使参照样本和测试样本的测量结果满足式(3)的归一化条件，采用步骤S21，对参照样本和测试样本的测量结果进行归一化处理。
由于参照样本和测试样本的测量是重复进行的，重复过程采用离散的方式来表示，将式(2)改写为系综的形式：
定义H为系综密度函数，则有：
<V>＝H<n> (5)
复杂体系的统计涨落则为：
其中,对应于重复测量而言，δS为测量的信息熵，δP为测量的环境变化量。
将δP作为噪声全景的统计空间，而δS作为信号的统计空间。所以根据贝叶斯公式，有：
在式(7)中，定义
为式(8)，式(8)是后验概率条件。本发明实施例中，为了使参照样本和测试样本的测量结果满足式(8)的后验概率条件，采用步骤S22，基于步骤S21获得的归一化结果建立后验概率模型框架。
则对复杂体系的统计涨落的估计δn
*为：
δn
*＝argmax
δnP(H<n>|δn)P(δn) (9)
在本发明实施例中，经过步骤S21-S22处理的测量结果，能够满足式(3)的归一化条件和式(8)的后验概率条件。满足了上述两个条件的测量结果，能够用 于实现式(9)中的、对复杂体系的统计涨落的估计。满足了上述两个条件的测量结果，将作为训练数据，用于后续的人工智能模型训练步骤中。
在参照样本和测试样本的测量结果形成训练数据的过程中，不同的噪声侧写构成噪声全景或至少部分噪声全景，同时，两类样本的整体测量结果、以及测量结果中的信号，都将分别呈现出稳定的统计特性；噪声呈现的统计分布模式，也将随着噪声全景的构建而趋于稳定。
在步骤S3中，人工智能模型可以选择但不限于：人工神经网络、感知机、支持向量机、贝叶斯分类器、贝叶斯网、随机森林模型或聚类模型。
在本发明实施例中，上述式(9)中的、对复杂体系的统计涨落的估计，将由人工智能模型实现。
实施例2
一种基于获取并识别噪声全景分布模型的信号分析系统，其结构示意图参见说明书附图3，系统包括测量模块1、处理模块2、训练模块3与分析模块4；
在富条件测量环境下，测量模块1对参照样本和测试样本进行重复测量，分别获得多个测量结果；其中，每个测量结果均包含信号和不同的噪声侧写；
在信号采集与分析的技术领域，保持样本测量过程中的外部条件一致性，是一种减小噪声波动、形成良好信噪比的常规手段，且重复进行样本测量也被认定是减少随机误差的有效方式。然而，在本发明实施例中，并不涉及外部条件的一致性设定，而是在富条件测量环境下，由测量模块1开展参照样本和测试样本的重复测量。其中，富条件是指：富条件是指不以保持外部条件一致性为目的的、不涉及抑制噪声的、自然的、包括真实的复杂噪声因素的测量条件。“富条件测量环境”与“重复测量”的目的都是为了获取足够构建噪声全景的、丰富的噪声侧写。
具体而言，在富条件测量环境下，受限于样本自身属性，参照样本与测试样本的测量结果中的信号将始终保持统计不变，但噪声会因环境变化而出现差别，即环境的变化会增加噪声的观测维度。基于多方位、多角度及多时空特点的噪声观测维度，由测量模块1进行参照样本和测试样本的重复测量，将形成丰富的噪声侧写。丰富的噪声侧写是后续构建噪声全景、基于数据统计规律识别噪声的基础。
处理模块2对参照样本和测试样本的测量结果进行处理，分别形成参照样本和测试样本的训练数据；其中，训练数据包括由多个噪声侧写构成的噪声全景或至少部分噪声全景；
如说明书附图4所示，在单一的噪声观测维度下，只能获取反映噪声局部的噪声侧写，无法获取全面的噪声观测结果，即噪声无法呈现出完整的、符合其真实分布特性的数据统计规律。然而，在本发明实施例中，参照样本与测试样本的重复测量在富条件测量环境下进行。不同噪声观测维度下获取的丰富的噪声侧写，将足够处理模块构建噪声全景或至少部分噪声全景。在处理模块2构建噪声全景的同时，噪声的数据统计规律将趋于其真实的数学统计规律。
在本发明实施例中，噪声全景是指：噪声的分布模型已经能够全面反映其理论上的真实分布模型；部分噪声全景是指：噪声的分布模型无法完整地反映其理论上的真实分布模型，但分布模型已经具备能够用于后续信号分析的精准程度。
基于参照样本和测试样本的训练数据，训练模块3以噪声的可观测性呈现为收敛目标，进行人工智能模型训练，使模型能够从测量结果中识别出信号与噪声，并区分参照样本和测试样本；
在本发明实施例中，参照样本和测试样本的训练数据将分别以预设比例，随机分配为学习数据和检测数据。训练模块3采用学习数据对人工智能模型进行训练，并将检测数据输入经训练的人工智能模型中，计算得出信号识别结果，若信号识别准确率低于预设阈值，采用学习数据继续进行训练，若信号识别准确率高于预设阈值，视为人工智能模型完成训练。
针对待识别样本的测量结果，分析模块4将其输入经训练的人工智能模型，人工智能模型的输出结果为该待识别样本的具体类型。
在本发明实施例中，参照样本和测试样本作为两种已知样本，二者的多个测量结果经处理后分别形成的训练数据，训练出的人工智能模型能够实现两种已知样本的有效区分。当待识别样本是两种已知样本中的某一种时，人工智能模型能够对待识别样本的具体类型进行准确识别。
可选择地，测量模块1包括微扰机构11，在测量模块1对参照样本和测试样本进行每次测量之前，微扰机构11引入轻微扰动，创造出富条件测量环境，由此增加样本测量的噪声观测维度，使每次样本测量的测量结果中包含不同的噪 声侧写。
进一步地，在参照样本和测试样本的每次测量前，微扰机构11引入的轻微扰动，可以选择但不限于空间微扰、时间微扰、物理微扰及环境微扰。
空间微扰包括但不限于：使测量位点发生轻微位移、使测量位点发生轻微旋转；时间微扰包括但不限于：增加测量时长、缩短测量时长、以及改变多次测量之间的时间间隔；物理微扰包括但不限于：在测量时震动测量设备或样本、对流质样本进行搅动；环境微扰包括但不限于：改变测量时的环境温度、改变测量时的环境湿度、改变测量时的电磁场、以及改变测量时的气压。
进一步地，处理模块2包括归一化模块21和后验概率模块22；
其中，归一化模块21对参照样本和测试样本的测量结果进行归一化处理，分别输出归一化结果；后验概率模块22基于归一化结果，建立后验概率模型框架，分别形成符合要求的、参照样本和测试样本的训练数据，用于后续的人工智能模型训练。
具体而言，在本发明实施例中，对于参照样本和测试样本的测量结果，将其视为对复杂体系构成的测量目标进行测量而得到的测量值。
定义测量密度函数为
其中，S为测量空间维度；V为测量环境；则测量目标中，体系数目为N，N由式(1)定义：
定义B(V)为测量函数，则测量值
有：
其中，
式(3)为归一化条件。本发明实施例中，为了使参照样本和测试样本的测量结果满足式(3)的归一化条件，由归一化模块21对参照样本和测试样本的测量结果进行归一化处理，输出归一化结果；
由于参照样本和测试样本的测量是重复进行的，重复过程采用离散的方式来表示，将式(2)改写为系综的形式：
定义H为系综密度函数，则有：
<V>＝H<n> (5)
复杂体系的统计涨落则为：
其中,对应于重复测量而言，δS为测量的信息熵，δP为测量的环境变化量。
将δP作为噪声全景的统计空间，而δS作为信号的统计空间。所以根据贝叶斯公式，有：
在式(7)中，定义
为式(8)，式(8)是后验概率条件。本发明实施例中，为了使参照样本和测试样本的测量结果满足式(8)的后验概率条件，后验概率模块22基于归一化结果，建立后验概率模型框架。
则对复杂体系的统计涨落的估计δn
*为：
δn
*＝argmax
δnP(H<n>|δn)P(δn) (9)
在本发明实施例中，经过归一化模块21和后验概率模块22处理的测量结果，能够满足式(3)的归一化条件和式(8)的后验概率条件。满足了上述两个条件的测量结果，能够用于实现式(9)中的、对复杂体系的统计涨落的估计。满足了上述两个条件的测量结果，将作为训练数据，用于后续的人工智能模型训练步骤中。
在处理模块2对参照样本和测试样本的测量结果进行处理、形成训练数据的过程中，不同的噪声侧写构成噪声全景或至少部分噪声全景，同时，两类样本的整体测量结果、以及测量结果中的信号，都将分别呈现出稳定的统计特性；噪声呈现的统计分布模式，也将随着噪声全景的构建而趋于稳定。
进一步地，人工智能模型可以选择但不限于：人工神经网络、感知机、支持向量机、贝叶斯分类器、贝叶斯网、随机森林模型或聚类模型。
在本发明实施例中，上述式(9)中的、对复杂体系的统计涨落的估计，将由人工智能模型实现。
实施例3
一种基于获取并识别噪声全景分布模型的信号分析方法，该方法包括如下步骤：
S1：在富条件测量环境下，对多种已知样本进行重复测量，分别获得多个测 量结果；其中，每个测量结果均包含信号和不同的噪声侧写；
在信号采集与分析的技术领域，保持样本测量过程中的外部条件一致性，是一种减小噪声波动、形成良好信噪比的常规手段，且重复进行样本测量也被认定是减少随机误差的有效方式。然而，在本发明实施例中，并不涉及外部条件的一致性设定，而是在富条件测量环境下，开展多种已知样本的重复测量。其中，富条件是指：富条件是指不以保持外部条件一致性为目的的、不涉及抑制噪声的、自然的、包括真实的复杂噪声因素的测量条件。“富条件测量环境”与“重复测量”的目的都是为了获取足够构建噪声全景的、丰富的噪声侧写。
具体而言，在富条件测量环境下，受限于样本自身属性，每种样本的测量结果中的信号将始终保持统计不变，但噪声会因环境变化而出现差别，即环境的变化会增加噪声的观测维度。基于多方位、多角度及多时空特点的噪声观测维度，进行每种已知样本的重复测量，将形成丰富的噪声侧写。丰富的噪声侧写是后续步骤中构建噪声全景、基于数据统计规律识别噪声的基础。
S2：对多种已知样本的测量结果进行处理，分别形成每种已知样本的训练数据；其中，训练数据包括由多个噪声侧写构成的噪声全景或至少部分噪声全景；
如说明书附图4所示，在单一的噪声观测维度下，只能获取反映噪声局部的噪声侧写，无法获取全面的噪声观测结果，即噪声无法呈现出完整的、符合其真实分布特性的数据统计规律。然而，在本发明实施例中，多种已知样本的重复测量在富条件测量环境下进行。不同噪声观测维度下获取的丰富的噪声侧写，将足够构建噪声全景或至少部分噪声全景。在构建噪声全景的同时，噪声的数据统计规律将趋于其真实的数学统计规律。
在本发明实施例中，噪声全景是指：噪声的分布模型已经能够全面反映其理论上的真实分布模型；部分噪声全景是指：噪声的分布模型无法完整地反映其理论上的真实分布模型，但分布模型已经具备能够用于后续信号分析的精准程度。
S3：基于多种已知样本的训练数据，以噪声的可观测性呈现为收敛目标，进行人工智能模型训练，使模型能够从测量结果中识别出信号与噪声，并区分多种已知样本；
在本发明实施例中，每种已知样本的训练数据将分别以预设比例，随机分配为学习数据和检测数据。采用学习数据对人工智能模型进行训练，并将检测数据 输入经训练的人工智能模型中，计算得出信号识别结果，若信号识别准确率低于预设阈值，采用学习数据继续进行训练，若信号识别准确率高于预设阈值，视为人工智能模型完成训练。
S4：将待识别样本的测量结果输入经训练的人工智能模型，人工智能模型的输出结果为该待识别样本的具体类型。
在本发明实施例中，多种已知样本的多个测量结果，经处理后分别形成的训练数据，训练出的人工智能模型能够实现每种已知样本的有效区分。当待识别样本是多种已知样本中的某一种时，人工智能模型能够对待识别样本的具体类型进行准确识别。
可选择地，在步骤S1中，在每种已知样本的每次测量之前，通过引入轻微扰动，创造出富条件测量环境，由此增加噪声观测维度，使每次测量的测量结果中包含不同的噪声侧写。
进一步地，在每种已知样本的每次测量前引入的轻微扰动，可以选择但不限于空间微扰、时间微扰、物理微扰及环境微扰。
空间微扰包括但不限于：使测量位点发生轻微位移、使测量位点发生轻微旋转；时间微扰包括但不限于：增加测量时长、缩短测量时长、以及改变多次测量之间的时间间隔；物理微扰包括但不限于：在测量时震动测量设备或样本、对流质样本进行搅动；环境微扰包括但不限于：改变测量时的环境温度、改变测量时的环境湿度、改变测量时的电磁场、以及改变测量时的气压。
在步骤S2中，对多种已知样本的测量结果进行处理，分别形成每种已知样本的训练数据的步骤包括：
S21.对每种已知样本的测量结果进行归一化处理；
S22.基于步骤S21的归一化结果，建立后验概率模型框架。
多种已知样本的测量结果，经过步骤S21-22处理后，将分别形成符合要求的训练数据，用于后续的人工智能模型训练。
具体而言，在本发明实施例中，对于每种已知样本的测量结果，将其视为对复杂体系构成的测量目标进行测量而得到的测量值。
定义测量密度函数为
其中，S为测量空间维度；V为测量环境；则测量目标中，体系数目为N，N由式(1)定义：
定义B(V)为测量函数，则测量值
有：
其中，
式(3)为归一化条件。本发明实施例中，为了使每种已知样本的测量结果满足式(3)的归一化条件，采用步骤S21，对每种已知样本的测量结果进行归一化处理。
由于每种已知样本的测量是重复进行的，重复过程采用离散的方式来表示，将式(2)改写为系综的形式：
定义H为系综密度函数，则有：
<V>＝H<n> (5)
复杂体系的统计涨落则为：
其中,对应于重复测量而言，δS为测量的信息熵，δP为测量的环境变化量。
将δP作为噪声全景的统计空间，而δS作为信号的统计空间。所以根据贝叶斯公式，有：
在式(7)中，定义
为式(8)，式(8)是后验概率条件。本发明实施例中，为了使每种已知样本的测量结果满足式(8)的后验概率条件，采用步骤S22，基于步骤S21获得的归一化结果建立后验概率模型框架。
则对复杂体系的统计涨落的估计δn
*为：
δn
*＝argmax
δnP(H<n>|δn)P(δn) (9)
在本发明实施例中，经过步骤S21-S22处理的测量结果，能够满足式(3)的归一化条件和式(8)的后验概率条件。满足了上述两个条件的测量结果，能够用于实现式(9)中的、对复杂体系的统计涨落的估计。满足了上述两个条件的测量结果，将作为训练数据，用于后续的人工智能模型训练步骤中。
在每种已知样本的测量结果形成训练数据的过程中，不同的噪声侧写构成噪声全景或至少部分噪声全景，同时，每种已知样本的整体测量结果、以及测量结果中的信号，都将分别呈现出稳定的统计特性；噪声呈现的统计分布模式，也将随着噪声全景的构建而趋于稳定。
在步骤S3中，人工智能模型可以选择但不限于：人工神经网络、感知机、支持向量机、贝叶斯分类器、贝叶斯网、随机森林模型或聚类模型。
在本发明实施例中，上述式(9)中的、对复杂体系的统计涨落的估计，将由人工智能模型实现。
以上实施例仅用以说明本发明的技术方案，而非对其限制。本领域技术人员在本发明揭露的技术范围内，可轻易想到的变化或替换，都应涵盖在本发明的保护范围之内。因此，本发明的保护范围应该以权利要求所界定的保护范围为准。
Claims (10)
- 一种基于获取并识别噪声全景分布模型的信号分析方法，其特征在于：包括如下步骤：S1：在富条件测量环境下，对参照样本和测试样本进行重复测量，分别获得多个测量结果；其中，每个测量结果均包含信号和不同的噪声侧写；富条件是指不以保持外部条件一致性为目的的、不涉及抑制噪声的、自然的、包括真实的复杂噪声因素的测量条件；S2：对参照样本和测试样本的测量结果进行处理，分别形成参照样本和测试样本的训练数据；其中，训练数据包括由多个噪声侧写构成的噪声全景或至少部分噪声全景；S3：基于参照样本和测试样本的训练数据，以噪声的可观测性呈现为收敛目标，进行人工智能模型训练，使模型能够从测量结果中识别出信号与噪声，并区分参照样本和测试样本；S4：将待识别样本的测量结果输入经训练的人工智能模型，人工智能模型的输出结果为该待识别样本的具体类型。
- 根据权利要求1所述的基于获取并识别噪声全景分布模型的信号分析方法，其特征在于：在步骤S1中，在参照样本和测试样本的每次测量之前，通过引入轻微扰动，创造出富条件测量环境，由此增加噪声观测维度，使每次测量的测量结果中包含不同的噪声侧写。
- 根据权利要求2所述的基于获取并识别噪声全景分布模型的信号分析方法，其特征在于：所述轻微扰动可以选择但不限于空间微扰、时间微扰、物理微扰及环境微扰；其中，空间微扰包括但不限于：使测量位点发生轻微位移、使测量位点发生轻微旋转；时间微扰包括但不限于：增加测量时长、缩短测量时长、以及改变多次测量之间的时间间隔；物理微扰包括但不限于：在测量时震动测量设备或样本、对流质样本进行搅动；环境微扰包括但不限于：改变测量时的环境温度、改变测量时的环境湿度、改变测量时的电磁场、以及改变测量时的气压。
- 根据权利要求1所述的基于获取并识别噪声全景分布模型的信号分析方法，其特征在于：在步骤S2中，对参照样本和测试样本的测量结果进行处理，分别形成参照样本和测试样本的训练数据的步骤包括：S21.对参照样本和测试样本的测量结果进行归一化处理；S22.基于步骤S21的归一化结果，建立后验概率模型框架；参照样本和测试样本的测量结果，经过步骤S21-22处理后，将分别形成符合要求的训练数据，用于后续的人工智能模型训练。
- 根据权利要求1所述的基于获取并识别噪声全景分布模型的信号分析方法，其特征在于：在步骤S3中，人工智能模型可以选择但不限于：人工神经网络、感知机、支持向量机、贝叶斯分类器、贝叶斯网、随机森林模型或聚类模型。
- 根据权利要求1所述的基于获取并识别噪声全景分布模型的信号分析方法，其特征在于：在步骤S3中，在人工智能模型的训练过程中，模型将以迭代的方式，对训练数据内含的能够实现信号与噪声识别的特征，以及能够实现参照样本和测试样本区分的特征，进行大量经验性学习、归纳与收敛，并习得特征与所述预设标签之间的联系；其中，能够实现信号识别的特征，包括多个测量结果经处理后呈现的、符合信号真实数学统计规律的统计分布模式；能够实现噪声识别的特征，包括由多样化的噪声侧写构建的噪声全景所呈现的、趋近于噪声真实数学统计规律的统计分布模式；能够实现参照样本和测试样本区分的特征，包括参照样本和测试样本的多个测量结果经处理后、分别呈现出的统计分布模式。
- 根据权利要求6所述的基于获取并识别噪声全景分布模型的信号分析方法，其特征在于：预设标签包括输出标签与输入标签；其中，输出标签包括分别代表参照样本和测试样本的两个标签；输入标签是分别涉及参照样本和测试样本的训练数据的两组耦合性标签，每个耦合性标签分别与样本测量时、所处的富条件测量环境相关联；不同组别的每个耦合性标签分别代表：在富条件测量环境中的每个独立的测量环境下，参照样本或测试样本的测量结果与噪声全景的耦合；其中，测量结果包含的噪声侧写，是这个独立的测量环境下获取的噪声侧写。
- 一种基于获取并识别噪声全景分布模型的信号分析系统，其特征在于：包括测量模块、处理模块、训练模块与分析模块；在富条件测量环境下，测量模块对参照样本和测试样本进行重复测量，分别获得多个测量结果；其中，每个测量结果均包含信号和不同的噪声侧写；处理模块对参照样本和测试样本的测量结果进行处理，分别形成参照样本和测试样本的训练数据；其中，训练数据包括由多个噪声侧写构成的噪声全景或至少部分噪声全景；基于参照样本和测试样本的训练数据，训练模块以噪声的可观测性呈现为收敛目标，进行人工智能模型训练，使模型能够从测量结果中识别出信号与噪声，并区分参照样本和测试样 本；针对待识别样本的测量结果，分析模块将其输入经训练的人工智能模型，人工智能模型的输出结果为该待识别样本的具体类型。
- 根据权利要求8所述的基于获取并识别噪声全景分布模型的信号分析系统，其特征在于：测量模块包括微扰机构，在测量模块对参照样本和测试样本进行每次测量之前，微扰机构引入轻微扰动，创造出富条件测量环境，由此增加样本测量的噪声观测维度，使每次样本测量的测量结果中包含不同的噪声侧写；其中，在参照样本和测试样本的每次测量前，微扰机构引入的轻微扰动，可以选择但不限于空间微扰、时间微扰、物理微扰及环境微扰。
- 根据权利要求8所述的基于获取并识别噪声全景分布模型的信号分析系统，其特征在于：处理模块包括归一化模块和后验概率模块；其中，归一化模块对参照样本和测试样本的测量结果进行归一化处理，分别输出归一化结果；后验概率模块基于归一化结果，建立后验概率模型框架，分别形成符合要求的、参照样本和测试样本的训练数据，用于后续的人工智能模型训练。
Priority Applications (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2023504132A JP2023535905A (ja) | 2020-12-04 | 2021-06-10 | ノイズパノラマ分布モデルの取得と認識に基づく信号解析方法及びシステム |
| US18/247,842 US20230385378A1 (en) | 2020-12-04 | 2021-06-10 | Signal analysis method and system based on model for acquiringand identifying noise panoramic distribution |
| EP21899310.3A EP4167128A4 (en) | 2020-12-04 | 2021-06-10 | METHOD AND SYSTEM FOR ANALYZING MODEL-BASED SIGNALS FOR ACQUIRING AND IDENTIFYING PANORAMIC DISTRIBUTION OF NOISE |
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011411339.2 | 2020-12-04 | ||
| CN202011411339.2A CN114662522A (zh) | 2020-12-04 | 2020-12-04 | 基于获取并识别噪声全景分布模型的信号分析方法及系统 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2022116508A1 true WO2022116508A1 (zh) | 2022-06-09 |
Family
ID=81852928
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2021/099384 WO2022116508A1 (zh) | 2020-12-04 | 2021-06-10 | 基于获取并识别噪声全景分布模型的信号分析方法及系统 |
Country Status (5)
| Country | Link |
|---|---|
| US (1) | US20230385378A1 (zh) |
| EP (1) | EP4167128A4 (zh) |
| JP (1) | JP2023535905A (zh) |
| CN (1) | CN114662522A (zh) |
| WO (1) | WO2022116508A1 (zh) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117349596B (zh) * | 2023-12-04 | 2024-03-29 | 深圳汉德霍尔科技有限公司 | 基于多传感器的电池异常状态监测预警系统 |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130064423A1 (en) * | 2011-09-09 | 2013-03-14 | Sony Corporation | Feature extraction and processing from signals of sensor arrays |
| CN107808098A (zh) * | 2017-09-07 | 2018-03-16 | 阿里巴巴集团控股有限公司 | 一种模型安全检测方法、装置以及电子设备 |
| CN108111294A (zh) * | 2017-12-13 | 2018-06-01 | 南京航空航天大学 | 一种基于ML-kNN的保护隐私的多标记分类方法 |
| CN109508740A (zh) * | 2018-11-09 | 2019-03-22 | 郑州轻工业学院 | 基于高斯混合噪声生成式对抗网络的物体硬度识别方法 |
| CN111436929A (zh) * | 2019-01-17 | 2020-07-24 | 复旦大学 | 一种神经生理信号的生成和识别方法 |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018013247A2 (en) * | 2016-06-02 | 2018-01-18 | Brown University | Physics informed learning machine |
-
2020
- 2020-12-04 CN CN202011411339.2A patent/CN114662522A/zh active Pending
-
2021
- 2021-06-10 WO PCT/CN2021/099384 patent/WO2022116508A1/zh active Application Filing
- 2021-06-10 JP JP2023504132A patent/JP2023535905A/ja active Pending
- 2021-06-10 EP EP21899310.3A patent/EP4167128A4/en active Pending
- 2021-06-10 US US18/247,842 patent/US20230385378A1/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130064423A1 (en) * | 2011-09-09 | 2013-03-14 | Sony Corporation | Feature extraction and processing from signals of sensor arrays |
| CN107808098A (zh) * | 2017-09-07 | 2018-03-16 | 阿里巴巴集团控股有限公司 | 一种模型安全检测方法、装置以及电子设备 |
| CN108111294A (zh) * | 2017-12-13 | 2018-06-01 | 南京航空航天大学 | 一种基于ML-kNN的保护隐私的多标记分类方法 |
| CN109508740A (zh) * | 2018-11-09 | 2019-03-22 | 郑州轻工业学院 | 基于高斯混合噪声生成式对抗网络的物体硬度识别方法 |
| CN111436929A (zh) * | 2019-01-17 | 2020-07-24 | 复旦大学 | 一种神经生理信号的生成和识别方法 |
Non-Patent Citations (1)
| Title |
|---|
| See also references of EP4167128A4 * |
Also Published As
| Publication number | Publication date |
|---|---|
| EP4167128A4 (en) | 2024-05-08 |
| JP2023535905A (ja) | 2023-08-22 |
| CN114662522A (zh) | 2022-06-24 |
| EP4167128A1 (en) | 2023-04-19 |
| US20230385378A1 (en) | 2023-11-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Sahasrabudhe et al. | Self-supervised nuclei segmentation in histopathological images using attention | |
| WO2021253510A1 (zh) | 基于双向交互网络的行人搜索方法、系统、装置 | |
| CN112183643B (zh) | 基于声发射的硬岩拉剪破裂识别方法及装置 | |
| CN108664986B (zh) | 基于lp范数正则化的多任务学习图像分类方法及系统 | |
| CN112529005B (zh) | 基于语义特征一致性监督金字塔网络的目标检测方法 | |
| Zhang et al. | Explainability metrics of deep convolutional networks for photoplethysmography quality assessment | |
| CN109242010A (zh) | 一种稀疏学习rcs序列特征提取方法 | |
| WO2022116508A1 (zh) | 基于获取并识别噪声全景分布模型的信号分析方法及系统 | |
| CN112200238A (zh) | 基于声响特征的硬岩拉剪破裂识别方法与装置 | |
| CN115358337A (zh) | 一种小样本故障诊断方法、装置及存储介质 | |
| CN108805181B (zh) | 一种基于多分类模型的图像分类装置及分类方法 | |
| CN111383217B (zh) | 大脑成瘾性状评估的可视化方法、装置及介质 | |
| CN112861881A (zh) | 一种基于改进MobileNet模型的蜂窝肺识别方法 | |
| Altinok et al. | Activity analysis in microtubule videos by mixture of hidden Markov models | |
| Yang et al. | A semantic information decomposition network for accurate segmentation of texture defects | |
| CN116011307A (zh) | 基于获取并识别噪声全景分布模型的信号分析方法及系统 | |
| Li et al. | How to identify pollen like a palynologist: A prior knowledge-guided deep feature learning for real-world pollen classification | |
| Alam et al. | Synthetic Brain Image Generation for ADHD prediction based on Progressive Growing Generative Adversarial Network | |
| AltundoĞan et al. | Cracked Wall Image Classification Based on Deep Neural Network Using Visibility Graph Features | |
| CN104573746A (zh) | 基于磁共振成像的实蝇种类识别方法 | |
| CN113780084B (zh) | 基于生成式对抗网络的人脸数据扩增方法、电子设备和存储介质 | |
| JPWO2022116508A5 (zh) | ||
| Yang et al. | Efficient pattern unmixing of multiplex proteins based on variable weighting of texture descriptors | |
| Li et al. | Robustness Analysis for VGG-16 Model in Image Classification of Post-Hurricane Buildings | |
| Perala et al. | Optical Character Recognition for Test Automation Using LabVIEW |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 21899310 Country of ref document: EP Kind code of ref document: A1 |
|
| ENP | Entry into the national phase |
Ref document number: 2023504132 Country of ref document: JP Kind code of ref document: A |
|
| ENP | Entry into the national phase |
Ref document number: 2021899310 Country of ref document: EP Effective date: 20230112 |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 18247842 Country of ref document: US |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |