Disease monitoring platform intelligent detection system based on machine learning

Abstract

The invention relates to the technical field of machine learning, and particularly discloses a disease monitoring platform intelligent detection system based on machine learning, which comprises a signal acquisition and processing module, a signal enhancement and model prediction module and a model output optimization module, the system collects original physiological signal data in real time through a multi-modal sensor, and performs preprocessing and preliminary noise suppression to obtain a high-quality physiological signal data set, thereby providing reliable basic data support for disease monitoring; aiming at the initial weak signal, the system adopts a self-adaptive enhancement strategy to improve the identifiability of the initial weak signal, and performs classification processing according to a signal-to-noise ratio so as to generate an accurate disease prediction result; and on the basis of credibility of a prediction result output by the machine learning model, signal processing and machine learning model identification capabilities are continuously improved. According to the system, weak physiological features are effectively enhanced, the prediction accuracy is improved, and adaptive closed-loop optimization is realized.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to an intelligent detection system for a disease monitoring platform based on machine learning. Background Technology

[0002] With the application of machine learning technology, intelligent detection technology for disease monitoring platforms based on machine learning is developing in multiple dimensions in terms of technical path: current technology can be divided into two major directions: medical image analysis and physiological data monitoring. Medical image analysis technology learns a large amount of normal and diseased image data through models to identify lesion areas such as lung nodules and inflammation. Physiological data monitoring technology uses algorithms to analyze continuous physiological data such as heart rate, blood pressure, and heart rate variability to explore cardiovascular disease risk prediction. At the same time, some technologies attempt to combine multi-source data, such as electronic medical records and environmental data, to improve monitoring coverage.

[0003] For example, Chinese invention patent CN120069117A discloses a method, system, and device for model detection based on a machine learning platform. The method for model detection based on a machine learning platform includes: acquiring model detection data, wherein the model detection data includes prediction data, prediction results obtained after the prediction data is input into the model to be tested, and evaluation data corresponding to the prediction data; when the prediction results and / or evaluation data meet preset model detection conditions, performing feature analysis based on the prediction data to determine the feature analysis results; determining the model index evaluation results of the model to be tested based on the evaluation data; and determining the detection results of the model to be tested deployed in the machine learning platform based on the feature analysis results and the model index evaluation results.

[0004] For example, Chinese invention patent CN106779087B discloses a general machine learning data analysis platform, including an interface module, a data storage module, a preprocessing module, a feature extraction module, a feature transformation module, an algorithm module, and a selection and optimization module. The feature extraction module extracts feature parameters from the data to be analyzed based on user-defined feature parameters. The feature transformation module converts user-defined features into the representation form required by the user. The algorithm module includes multiple algorithm models for the user to choose from and allows the user to build at least one set of models. The selection and optimization module selects the optimal model and optimal parameters from the built models and then saves the optimal model. The data generated by each of the above modules is stored in the data storage module.

[0005] The aforementioned technologies suffer from at least the following technical problems: In disease monitoring scenarios, the appearance of weak signals stems from multiple real-world factors: In the early or mild stages of a disease, the abnormality in physiological indicators is inherently small, for example, insulin secretion fluctuations in prediabetes are only 5-10 μU / mL; During sensor acquisition, poor contact and insufficient power consumption of the device can cause significant signal attenuation, and a loose wrist heart rate sensor can reduce the PPG signal amplitude to 1 / 3 of the normal value; Environmental electromagnetic interference and human motion noise can also mask the target signal, such as 50Hz power frequency interference drowning out weak ECG features; These weak signals can lead to the misfiltering of effective signals or residual noise during preprocessing, the omission of real anomalies and the mis-extraction of noise pseudo-features during feature extraction, and the inability of the model to capture the features of weak signals during recognition, ultimately directly reducing the accuracy of machine learning model predictions. Summary of the Invention

[0006] To address the technical problem of insufficient prediction accuracy in machine learning models due to weak signals in existing technologies, this invention provides an intelligent detection system for a disease monitoring platform based on machine learning. The technical solution is as follows: The signal acquisition and processing module is used to acquire raw physiological signal data in real time through multimodal sensors and integrate it into a raw physiological signal dataset. This dataset undergoes preprocessing and initial noise suppression to obtain a physiological signal dataset. The physiological signal dataset is then segmented and analyzed according to a preset time window, and initial weak signals are automatically identified. This raw physiological signal dataset provides basic data support for disease monitoring. The signal enhancement and model prediction module enhances the identified initial weak signals to obtain weak signals, which are then input into a pre-trained machine learning model in the disease monitoring platform. Simultaneously, classification processing is performed based on the signal-to-noise ratio of the weak signals, resulting in the output of the machine learning model's disease prediction results. The model output optimization module monitors the output parameters of the machine learning model, determines the reliability of the disease prediction results, and solidifies the weak signal enhancement strategy and machine learning model parameters based on highly reliable disease prediction results. For disease prediction results with low reliability, the signal-model processing link is optimized in multiple dimensions, thereby achieving intelligent detection on the machine learning-based disease monitoring platform.

[0007] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: (1) This embodiment of the invention provides an intelligent detection system for a disease monitoring platform based on machine learning. The system forms a complete monitoring link through the coordinated cooperation of three major modules: data acquisition, signal processing, model prediction, and result optimization. From the data level, it ensures the high quality of basic data and the accurate positioning of key signals; from the model level, it realizes targeted processing of signals of different quality and dynamic optimization of prediction results; ultimately, it not only improves the ability to capture weak signals and the accuracy of prediction, but also takes into account the real-time nature of monitoring and the reliability of results, which helps to shorten the discovery cycle of weak signals and improve the timeliness of signal feature analysis. It is of great significance for improving the efficiency of weak signal monitoring and the accuracy of signal feature extraction.

[0008] (2) The original physiological signal dataset is collected and integrated in real time by multimodal sensors. After preprocessing to restore signal integrity and suppress initial noise, a high-quality physiological signal dataset is obtained. Then, it is segmented according to a preset time window and the initial weak signal is automatically identified. This scheme ensures data quality from the source, avoids the misleading effect of defects in the original signal on subsequent analysis, and provides reliable and focused basic data support for the entire monitoring system, reducing the interference of invalid data on the identification of weak signals.

[0009] (3) First, the initial weak signal is enhanced to highlight the disease-related features. Then, the signal-to-noise ratio is combined with the classification processing and input into the trained machine learning model to output the prediction results. This scheme specifically solves the core pain point that the weak signals in the early stage of the disease are easily masked. The signal enhancement makes the potential disease features clearer. The classification processing based on the signal-to-noise ratio, such as high signal-to-noise ratio focusing on feature purification and low signal-to-noise ratio focusing on local feature amplification, can adapt to weak signal scenarios of different qualities. This makes the model more targeted in processing various weak signals, greatly improves the sensitivity of the machine learning model in capturing weak features and the accuracy of feature prediction results, and avoids the feature extraction error caused by a single processing method for weak signals with different signal-to-noise ratios.

[0010] (4) The reliability of the prediction results is judged by monitoring the output process parameters of the model. When the reliability is high, the weak signal enhancement strategy and model parameters are solidified to ensure the efficiency and stability of subsequent processing. When the reliability is low, the signal-model processing link is optimized in multiple dimensions, such as adjusting the signal window size and correcting the model gating coefficient. This scheme constructs a closed loop of result feedback and strategy optimization. It not only realizes the reuse of effective strategies through parameter solidification and reduces the time cost of repeated optimization, but also corrects the deviation of low reliability results in a timely manner through multi-dimensional optimization, avoiding the risk of misjudgment or omission of weak features, ensuring the reliability of the system's output results of weak signal features, and solving the problem of insufficient stability of weak signal processing caused by the lack of dynamic optimization after traditional model output. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of the intelligent detection system for a disease monitoring platform based on machine learning, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the time-series disease monitoring process of the machine learning model provided in this embodiment of the invention; Figure 3 This is a schematic diagram of the dual-branch feature extraction and classification of the machine learning model provided in this embodiment of the invention; Figure 4 This is a flowchart of weak signal enhancement and classification processing provided in an embodiment of the present invention; Figure 5 This is a flowchart of the weak signal feature vector component optimization and strategy effectiveness determination provided in the embodiments of the present invention; Figure 6 This is a flowchart of short-time window signal amplitude optimization and strategy effectiveness determination provided in an embodiment of the present invention; Figure 7 This is a flowchart of the accuracy judgment and subsequent processing of the model output results provided in the embodiments of the present invention; Figure 8 This is a flowchart of the model output final value judgment and multi-dimensional iterative optimization early warning process provided in the embodiments of the present invention. Detailed Implementation

[0013] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0014] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0015] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0016] This invention provides an intelligent detection system for a disease surveillance platform based on machine learning, such as... Figure 1The flowchart shown is of an intelligent detection system for a disease monitoring platform based on machine learning. The system includes: a signal acquisition and processing module, a signal enhancement and model prediction module, a model output optimization module, and a management database.

[0017] The signal acquisition and processing module is connected to the signal enhancement and model prediction module, which in turn is connected to the model output optimization module. All three modules—signal acquisition and processing, signal enhancement and model prediction, and model output optimization—are connected to a management database. This management database stores various parameters involved in the intelligent detection system of the machine learning-based disease monitoring platform.

[0018] In this invention, to achieve intelligent detection of the machine learning-based disease monitoring platform, a mapping rule system is configured based on historical operational data and multi-scenario experimental stress test results. This system integrates the correspondence between multi-dimensional indicators (such as signal fluctuation metrics) and multi-level adjustment parameters. Through data fitting and normalization, a segmented and scalable parameter mapping table is established. This system supports both static configuration and dynamic updates, ensuring the real-time adaptability and controllability of the control strategy. The system includes various mapping relationships, such as the mapping relationship between signal fluctuation metrics and the defined signal-to-noise ratio correction value, all of which are preset in the management database and can be automatically invoked. All mapping relationships involved in this invention can be obtained through this mapping rule system.

[0019] The signal acquisition and processing module is used to acquire raw physiological signal data in real time through multimodal sensors and integrate it into a raw physiological signal dataset. The module performs preprocessing and initial noise suppression on the raw physiological signal dataset to obtain a physiological signal dataset. The physiological signal dataset is then segmented and analyzed according to a preset time window, and initial weak signals are automatically identified. This raw physiological signal dataset provides basic data support for disease monitoring. The aforementioned raw physiological signal dataset refers to the unprocessed collection of multi-source physiological signals acquired in real time at the patient's end through multimodal sensors (such as ECG sensors, photoelectric blood oxygen sensors, etc.), specifically including ECG signals, blood oxygen signals, and respiratory signals. The aforementioned initial weak signals are short-term signal segments with low energy and small amplitude but potentially clinical significance from the raw physiological signals.

[0020] The aforementioned preprocessing of the original physiological signal dataset refers to performing neighborhood mean difference processing on the incomplete original physiological signal dataset to restore the integrity of the signal. The aforementioned preliminary noise suppression refers to suppressing high-frequency interference through methods such as wavelet denoising and bandpass filters. A signal fluctuation metric is obtained by weighting the instantaneous fluctuation value, cumulative fluctuation trend, and cross-modal consistency fluctuation rate of the signal. The weights of the instantaneous fluctuation value, cumulative fluctuation trend, and cross-modal consistency fluctuation rate are determined by professionals in the field. The signal fluctuation metric is compared with the preset stable range of the signal fluctuation metric in the management database. If the signal fluctuation metric is within the stable range, the current defined signal-to-noise ratio remains unchanged. If the signal fluctuation metric is not within the stable range, the current defined signal-to-noise ratio is corrected based on the signal fluctuation metric. The signal fluctuation metric is input into the management database, which can then match the corresponding defined signal-to-noise ratio correction value based on the preset signal fluctuation metric-defined signal-to-noise ratio correction value mapping table, thereby realizing the correction of the defined signal-to-noise ratio.

[0021] The aforementioned signal-to-noise ratio (SNR) correction value refers to the quantization parameter used to dynamically adjust the high / low SNR threshold when the signal fluctuation metric exceeds the preset stable range. Its core function is to ensure that the SNR definition standard is compatible with the overall fluctuation state of the current signal, thereby avoiding misjudgment of the SNR due to changes in signal stability.

[0022] The instantaneous fluctuation value of the above signal represents the short-term change amplitude of the signal at each sampling point, which can be obtained by calculating the absolute difference or differential of consecutive sampling points; the cumulative fluctuation trend represents the overall cumulative fluctuation of the signal within the time window, which can be obtained by calculating the cumulative deviation of the signal from the average value within the window; the cross-modal consistency volatility represents the consistency of fluctuation of different modal signals in the same time period, which can be obtained by calculating the correlation coefficient of each modal signal within the window and taking the average.

[0023] The signal enhancement and model prediction module enhances the identified initial weak signal to obtain a weak signal, which is then input into a pre-trained machine learning model in the disease monitoring platform. Simultaneously, it performs classification processing based on the signal-to-noise ratio of the weak signal, outputting the disease prediction result from the machine learning model. The model output optimization module monitors the output parameters of the machine learning model, determining the reliability of the disease prediction result. Based on highly reliable disease prediction results, it solidifies the weak signal enhancement strategy and machine learning model parameters. Based on less reliable disease prediction results, it performs multi-dimensional optimization of the signal-model processing link, thereby achieving intelligent detection on the machine learning-based disease monitoring platform.

[0024] The trained machine learning model mentioned above, such as Figure 2 The time-series disease monitoring process of the machine learning model provided in this embodiment of the invention is illustrated in the diagram. t This refers to multimodal physiological signal fragments, including but not limited to ECG, SpO2, respiration, and blood pressure. t For temporal feature representation (Transformer encoding), o t The disease risk probability output is [0, 1], and Unfold expands and processes multiple signal segments in the time dimension (parameters are shared). X t-1 X t+1 This represents the data input into the network at times t-1 and t+1. In disease surveillance scenarios, it refers to the multimodal physiological signal segments collected over the time interval [t-1, t+1]. t-1 h t+1 U represents the hidden state / temporal feature representation calculated by the network at times t-1 and t+1. It carries information about historical signals and passes it to the next time step, corresponding to the temporal feature vector after Transformer encoding. U is the convolutional feature extraction weight, which is shared across all time steps to ensure that the same convolutional kernel is used to extract local waveform features such as ECG and blood oxygenation in different time segments. V is the time-dependent modeling (recurrent / attention weight), which is shared across all time steps to ensure that the integration rules for historical information are consistent at each step (whether it is a recurrent RNN or the attention of the Transformer). W is the classification fully connected mapping weight, which is shared across all time steps to ensure that the temporal features at each time step are projected onto the same disease risk output space. U, V, and W are shared at each time step, and the parameters used at all time steps in the same network are exactly the same.

[0025] like Figure 3The schematic diagram of the dual-branch feature extraction and classification of the machine learning model provided in this embodiment of the invention is shown. Layer C is a convolutional layer, corresponding to signal feature extraction at different levels; layer S is a downsampling layer, corresponding to signal compression and noise reduction enhancement; layer n is a fully connected layer, corresponding to fusion risk prediction; Input (1×28×28): The input is a multimodal physiological signal matrix (time×channel), such as ECG, SpO2, respiration, and blood pressure signals spliced ​​into a matrix according to time slices; C1 (Feature maps 4×24×24) is the first convolutional layer, i.e., local waveform feature extraction, extracting basic patterns such as ECG P wave, QRS complex, and small-cycle fluctuations of respiration; S1 (Feature maps 4×12×12) is the first downsampling, i.e., time / frequency domain compression, reducing data dimensionality, retaining principal components, and enhancing noise reduction capability; C2 (Feature maps 8×8×8) is the second convolutional layer, i.e., pattern combination feature, learning cross-signal combinations, and the coupling anomaly of ECG rhythm + SpO2 fluctuations; S2 (Feature maps The 8×4×4 layer is the second downsampling layer, i.e., feature aggregation, which further compresses the spatiotemporal resolution and retains key anomaly information; n1 (Fully Connected Layer 1) is the feature fusion layer, which fuses the convolutional features of the high signal-to-noise / low signal-to-noise dual branches to form a unified feature representation; n2 (Output Layer) is the disease risk probability output y∈[0,1], which is the final prediction result, representing the risk value of a certain disease.

[0026] Specifically, the initial weak signal is enhanced by matching the adaptive gain amplification local amplitude of the weak signal based on the signal optimization factor, thereby increasing the strength of the initial weak signal and finally obtaining the weak signal. The matching process for matching the adaptive gain amplification local amplitude of the weak signal based on the signal optimization factor is as follows: the signal optimization factor is input into the management database, and the management database can match the corresponding adaptive gain amplification local amplitude based on the preset signal optimization factor-adaptive gain amplification local amplitude mapping relationship.

[0027] The signal optimization factor is obtained by weighted averaging of local signal energy, local volatility, and cross-modal consistency coefficient, and is used to characterize the comprehensive characteristics of the signal in three dimensions: strength, stability, and multimodal coherence. The weights of the local signal energy, local volatility, and cross-modal consistency coefficient are determined by professionals in the field.

[0028] The aforementioned local signal energy, derived from a single modal signal in the physiological signal dataset, represents the intensity of a weak signal within a specific time window and can be calculated by the sum of squares of the signals at each sampling point within the window. The aforementioned local fluctuation, derived from the time-domain statistical characteristics of the physiological signal dataset after segmentation, represents the amplitude of signal fluctuation within a local time window and can be obtained by averaging the absolute values ​​of the deviations of each sampling point from the average value within the window. The aforementioned cross-modal consistency coefficient is calculated from the coupling relationship between multimodal signals in the physiological signal dataset and represents the consistency of fluctuations of different modal signals within the same time window.

[0029] Furthermore, classification processing is performed based on the signal-to-noise ratio (SNR) of weak signals. The specific analysis process is as follows: When the weak signal has a high SNR, it is input into a high SNR chain within the machine learning model for feature extraction, thereby obtaining the weak signal features. The activation entropy of the weak signal feature extraction process is monitored, and it is determined whether the feature vector component optimization strategy of the weak signal is triggered. The activation entropy is used to characterize the uncertainty of the activation distribution of the feature vector components during the weak signal feature extraction process. The aforementioned high SNR chain extracts key features and enhances or weights them for weak signals with high SNR, without blindly performing large-scale enhancements to avoid amplifying noise. The process of obtaining the activation entropy involves: normalizing the squared feature vector of the weak signal to a probability distribution, and then calculating the uncertainty of feature activation using the information entropy formula.

[0030] When the weak signal has a low signal-to-noise ratio (SNR), it is input into the low SNR chain in the machine learning model for feature extraction. This process extracts features from the weak signal, monitors the signal response energy during feature extraction, and determines whether a short-time window signal amplitude optimization strategy is triggered. The signal response energy characterizes the energy accumulation of the weak signal within a local time window after feature extraction. The aforementioned low SNR chain, for weak signals with low SNR, is responsible for segmented local enhancement and amplification of weak features while avoiding excessive noise amplification. The process of obtaining the signal response energy involves summing the squares of the convolutional features of the weak signal within a short-time window to obtain local feature energy for evaluating signal strength.

[0031] The aforementioned high signal-to-noise ratio (SNR) refers to a weak signal whose SNR is greater than or equal to the preset threshold SNR in the management database. The aforementioned low signal-to-noise ratio (SNR) refers to a weak signal whose SNR is less than the threshold SNR. The threshold SNR is a value used to distinguish between high and low SNR.

[0032] like Figure 4The flowchart of weak signal enhancement and classification processing provided in this embodiment of the invention shows that: first, the original physiological signal dataset is acquired, then preprocessing and preliminary noise suppression are performed, and then the initial weak signal is identified; the initial weak signal is enhanced based on the signal optimization factor to obtain the weak signal; it is determined whether the weak signal has a high signal-to-noise ratio (SNR); if so, it is input into the high SNR link in the machine learning model for feature extraction processing; then it is determined whether the activation entropy is not lower than the activation entropy threshold; if so, the feature vector component optimization strategy of the weak signal is triggered; finally, the signal optimization factor is corrected based on the activation entropy deviation value; if the activation entropy is lower than the activation entropy threshold, the feature vector component optimization strategy of the weak signal is not triggered; if the weak signal is not high SNR, it is input into the low SNR link in the machine learning model for feature extraction processing; it is determined whether the signal response energy is lower than the signal response energy threshold; if so, the short-time window signal amplitude optimization strategy is triggered; if not, the short-time window signal amplitude optimization strategy is not triggered.

[0033] Specifically, the process for determining whether to trigger the feature vector component optimization strategy for weak signals is as follows: the activation entropy of the weak signal feature extraction process in the high signal-to-noise ratio chain is compared with the preset activation entropy threshold in the management database. The aforementioned activation entropy threshold represents the maximum value of the activation entropy within a specified range.

[0034] When the activation entropy is not lower than the activation entropy threshold, it is determined that the feature vector component optimization strategy for weak signals is triggered. After the feature vector component optimization strategy for weak signals is completed, the activation entropy is reacquired and the feature vector component optimization strategy is adjusted and processed. At the same time, based on the activation entropy and the activation entropy threshold, the activation entropy deviation value is obtained, and the signal optimization factor is corrected based on the activation entropy deviation value. When the activation entropy is lower than the activation entropy threshold, it is determined that the feature vector component optimization strategy for weak signals is not triggered.

[0035] The aforementioned optimization strategy for the eigenvector components of weak signals refers to weighting and amplifying the eigenvector components of weak signals to increase their contribution to weak features. Specifically: , where f j fk represents the j-th component in the eigenvector. j The weighted feature components are j=1, 2, 3, ..., n, where n is the total number of components in the feature vector; w j Here, HL is the weighting coefficient for the j-th feature component, HT is the activation entropy threshold, and γ is the amplification factor. The specific values ​​are determined by professionals. Substituting the obtained activation entropy into this formula yields the weighted feature components of the weak signal. Through weighted amplification, the amplitude proportion of the corresponding component of the weak signal feature can be directionally increased, significantly widening the gap between it and irrelevant components. This allows the originally hidden useful features to stand out from the uniform distribution, greatly reducing the difficulty of model recognition and avoiding the omission of effective information.

[0036] The above-mentioned acquisition of the activation entropy deviation value refers to subtracting the activation entropy threshold from the activation entropy. The above-mentioned correction of the signal optimization factor based on the activation entropy deviation value is as follows: FN is the corrected signal optimization factor, FO is the current signal optimization factor, and DF is the activation entropy deviation value. Substituting the obtained activation entropy deviation value into the formula yields the corrected signal optimization factor. This enables adaptive enhancement and suppression of weak signals, preventing weak features from being submerged by averaging, thereby improving the model's sensitivity to identifying early features of potential diseases.

[0037] Specifically, the activation entropy is reacquired and the optimization strategy for adjusting the feature vector components is judged and processed. The specific analysis process is as follows: if the optimized activation entropy decreases and is lower than the activation entropy threshold, the feature vector component optimization strategy is marked as an effective strategy, and the high signal-to-noise ratio chain is solidified. Specifically, the weight coefficients, filter parameters and other processing link parameters of the current feature vector components are written into the management database and used as the priority call parameters for subsequent weak signal segments of the same category.

[0038] If the optimized activation entropy decreases but does not fall below the activation entropy threshold, the feature vector component optimization strategy is marked as a strategy to be optimized and continuously optimized. Continuous optimization refers to iteratively screening the feature vector components of weak signals, including gradually eliminating components with low contribution and increasing the weight ratio of high contribution components, until the maximum number of iterations is reached.

[0039] If the optimized activation entropy increases, the feature vector component optimization strategy is marked as invalid, and parameter rollback is performed. This means undoing the adjustments made to the feature vector component weights, signal optimization factors, etc., during the current optimization process, restoring them to the baseline parameters before optimization. This ensures the stability and security of the feature extraction chain, prevents a single failed optimization from damaging system performance, and provides a reliable starting point for the next round of optimization. After parameter rollback, the system can continue to try new optimization strategies, use default parameters or backup processing chains, or trigger anomaly flags in case of consecutive failures to ensure the continuity and effectiveness of signal processing.

[0040] like Figure 5 As shown in the flowchart of the weak signal feature vector component optimization and strategy effectiveness determination provided in this embodiment of the invention, after triggering the weak signal feature vector component optimization strategy, the optimized activation entropy is obtained. If the optimized activation entropy decreases and is lower than the activation entropy threshold, it is marked as an effective strategy, and the high signal-to-noise ratio chain is solidified. If the optimized activation entropy decreases but is not lower than the activation entropy threshold, it is marked as a strategy to be optimized, and the weak signal feature vector components are iteratively screened. If the optimized activation entropy increases, it is marked as an invalid strategy, and the parameters are rolled back. Finally, the accuracy index of the model output result is obtained.

[0041] Furthermore, to determine whether the short-term window signal amplitude optimization strategy is triggered, the specific determination process is as follows: the signal response energy of the weak signal feature extraction process in the low signal-to-noise ratio chain is compared with the preset signal response energy threshold in the management database; the aforementioned signal response energy threshold represents the minimum value of the signal response energy within the specified range.

[0042] When the signal response energy is lower than the signal response energy threshold, it is determined that the short-time window signal amplitude optimization strategy is triggered. The optimized signal response energy is obtained, and the short-time window signal amplitude optimization strategy is adjusted and processed. At the same time, based on the signal response energy and the signal response energy threshold, the signal response energy deviation value is obtained, and the signal optimization factor is corrected based on the signal response energy deviation value. When the signal response energy is not lower than the signal response energy threshold, it is determined that the short-time window signal amplitude optimization strategy is not triggered.

[0043] The aforementioned short-time window signal amplitude optimization strategy refers to dividing weak signals into short-time windows and amplifying the local windows. Specifically: α is the local window magnification factor, β is the amplification intensity adjustment parameter used to control the amplification amplitude sensitivity. The specific value is determined by professional technicians based on actual needs. ET is the signal response energy threshold, and ER is the signal response energy. Substituting the obtained signal response energy into the formula yields the local window magnification factor. Multiplying the obtained local window magnification factor by the original local window amplitude gives the optimized local window amplitude. This increases the amplitude of potentially useful signal segments, significantly improves the local signal-to-noise ratio, and makes weak signals stand out from noise, avoiding being misjudged as invalid signals by hardware / algorithms due to excessively low amplitude. It also simultaneously clarifies signal features, greatly reducing errors in subsequent feature extraction and parameter calculation, and improving the accuracy of weak signal analysis.

[0044] The aforementioned acquisition of the signal response energy deviation value refers to subtracting the signal response energy from the signal response energy threshold; the aforementioned correction of the signal optimization factor based on the signal response energy deviation value specifically involves: EN is the corrected signal optimization factor, EO is the current signal optimization factor, and DE is the signal response energy deviation value. Substituting the obtained signal response energy deviation value into the formula yields the corrected signal optimization factor. The larger the signal response energy deviation value, the larger the value of the optimized factor after correction, avoiding improper processing of signals of different weakness levels with a fixed optimization factor, such as insufficient amplification of extremely weak signals or excessive amplification of slightly weak signals. At the same time, the correction process is directly related to the actual energy state of the signal, enabling real-time calibration of the optimization strategy. This ensures that the energy of weak signals reaches the standard after amplification, meeting the requirements for subsequent feature extraction, while avoiding simultaneous noise enhancement due to blind amplification. This improves signal recognizability while ensuring processing stability.

[0045] Specifically, the optimized signal response energy is obtained, and the short-time window signal amplitude optimization strategy is judged and processed. The specific analysis process is as follows: if the optimized signal response energy increases but is not lower than the signal response energy value, the short-time window signal amplitude optimization strategy is marked as an effective strategy, and the low signal-to-noise ratio chain is solidified. The currently effective strategy parameters are written into the system configuration library, including the short-time window length, sliding step size, etc.; if the optimized signal response energy increases but is lower than the signal response energy threshold, the short-time window signal amplitude optimization strategy is marked as a strategy to be optimized, and continuous optimization is carried out. Continuous optimization refers to using signal reconstruction technology to recover the characteristics of weak signals. The above-mentioned increase in optimized signal response energy means that the optimized signal response energy is higher than the signal response energy before optimization.

[0046] If the optimized signal response energy decreases, the short-time window signal amplitude optimization strategy is marked as invalid, and parameter rollback is performed. This means undoing the adjustments made to the short-time window signal amplitude, signal optimization factor, etc., during the current optimization process and restoring them to the baseline parameters before optimization.

[0047] like Figure 6 As shown in the flowchart of the short-time window signal amplitude optimization and strategy effectiveness determination provided in this embodiment of the invention, after triggering the short-time window signal amplitude optimization strategy, the optimized signal response energy is obtained. If the optimized signal response energy increases and is not lower than the signal response energy threshold, it is marked as an effective strategy, and the low signal-to-noise ratio chain is solidified. If the optimized signal response energy increases but is lower than the signal response energy threshold, it is marked as a strategy to be optimized, and optimization continues. If the optimized signal response energy decreases, it is marked as an invalid strategy, and the parameters are rolled back. Finally, the accuracy index of the model output result is obtained.

[0048] Specifically, the output parameters of the machine learning model are monitored, and the analysis process is as follows: The output parameters of a machine learning model include the model response energy proportionality coefficient, the activation entropy proportionality coefficient, and the local signal-to-noise ratio proportionality coefficient. By pre-setting the effect coefficients of these coefficients in a management database, the weighted contribution of each coefficient to the model's output accuracy index is quantified. Finally, a weighted average fusion algorithm is used to synthesize the model's output accuracy index, which represents the accuracy of the disease prediction results output by the machine learning model. The model response energy proportionality coefficient is the ratio of the model response energy to the bounded model response energy; the activation entropy proportionality coefficient is the ratio of the activation entropy to the bounded activation entropy; and the local signal-to-noise ratio proportionality coefficient is the ratio of the local signal-to-noise ratio to the bounded local signal-to-noise ratio.

[0049] The above-mentioned model response energy represents the activation intensity of the machine learning model for the input signal features, and is obtained by calculating the sum of squares of the model's output feature vectors; the above-mentioned activation entropy characterizes the uncertainty of the activation distribution of feature vector components during the extraction of weak signal features, that is, the larger the activation entropy, the more dispersed the features, and the smaller the activation entropy, the more concentrated the features, and is obtained by calculating the entropy value after normalizing the feature vectors to a probability distribution; the above-mentioned local signal-to-noise ratio reflects the ratio of effective feature power to noise power in a signal segment, and is used to measure the identifiability of the segment's features, and is obtained by calculating and normalizing the ratio of signal power to noise power within a time window.

[0050] The above definition of model response energy represents the minimum value of model response energy within a specified range; the above definition of activation entropy represents the maximum value of activation entropy within a specified range; the above definition of partial signal-to-noise ratio represents the minimum value of partial signal-to-noise ratio within a specified range.

[0051] If the local signal-to-noise ratio is high, it means that the power of the effective features in the signal segment is much greater than the power of the noise, the features are highly identifiable, the feature quality input to the model is better, the model activates the effective features more fully, the model response energy is often higher, and the concentrated feature distribution makes the activation entropy lower. If the local signal-to-noise ratio is low, that is, the noise interference is strong, the model may mistakenly activate the noise, resulting in abnormally high or low response energy, and the dispersed feature distribution makes the activation entropy higher.

[0052] A higher model response energy ratio indicates that the model has a stronger activation intensity for effective signal features, captures more effective features, and helps improve the accuracy index of the model output results. A higher activation entropy ratio means that the feature distribution is more dispersed, the model's focus on disease features decreases, and it is easy to lead to a decrease in the accuracy index of the model output results. A higher local signal-to-noise ratio ratio indicates that the power of effective features in the signal segment is much greater than the power of noise, the feature identifiability is stronger, the quality of the model input features is better, and the accuracy index of the model output results is greater.

[0053] The accuracy index of the model output results is evaluated using the following method: ; In the formula, moai is the accuracy index of the model output result, mrerc is the model response energy proportional coefficient, lsnrc is the local signal-to-noise ratio proportional coefficient, aerc is the activation entropy proportional coefficient, gh is the effect coefficient corresponding to the model response energy proportional coefficient preset in the management database, gl is the effect coefficient corresponding to the local signal-to-noise ratio proportional coefficient preset in the management database, and gt is the effect coefficient corresponding to the activation entropy proportional coefficient preset in the management database.

[0054] The effect coefficients corresponding to the model response energy proportionality coefficients mentioned above indicate that when the model response energy proportionality coefficient changes by a unit amplitude, the accuracy index of the model output result will change accordingly. The effect coefficients corresponding to the local signal-to-noise ratio proportionality coefficients mentioned above indicate that when the local signal-to-noise ratio proportionality coefficient changes by a unit amplitude, the accuracy index of the model output result will change accordingly. The effect coefficients corresponding to the activation entropy proportionality coefficients mentioned above indicate that when the activation entropy proportionality coefficient changes by a unit amplitude, the accuracy index of the model output result will change accordingly.

[0055] The management database stores the mapping relationships between the model response energy ratio coefficient and its corresponding effect coefficient, the local signal-to-noise ratio ratio coefficient and its corresponding effect coefficient, and the activation entropy ratio coefficient and its corresponding effect coefficient. In this embodiment, the mapping relationship is a mapping table. For example, when the model response energy ratio coefficient, the local signal-to-noise ratio ratio coefficient, and the activation entropy ratio coefficient are input into the management database, the management database can match the corresponding effect coefficients of the model response energy ratio coefficient, the local signal-to-noise ratio ratio coefficient, and the activation entropy ratio coefficient based on the preset mapping relationship table. The numerical range of each effect coefficient is strictly controlled between 0 and 1.

[0056] The accuracy index of the model output result is compared with the accuracy threshold of the model output result. The accuracy threshold of the model output result refers to the minimum value of the accuracy index of the model output result within a specified range preset in the management database. When the accuracy index of the model output result is not lower than the accuracy threshold of the model output result, the disease prediction result is judged to have high credibility. When the accuracy index of the model output result is lower than the accuracy threshold of the model output result, the disease prediction result is judged to have low credibility.

[0057] like Figure 7As shown in the flowchart of the model output result accuracy judgment and subsequent processing provided in this embodiment of the invention, after obtaining the model output result accuracy index, it is determined whether the model output result accuracy index is not lower than the model output result accuracy threshold. If so, the disease prediction result has high credibility; if not, the disease prediction result does not have high credibility. Based on the model output result accuracy deviation value, the time window length for weak signal feature extraction is reduced, the gating in the machine learning model is increased, and finally the model output result accuracy final value is obtained.

[0058] like Figure 8 The flowchart for judging the final value of the model output result and the early warning of multi-dimensional iterative optimization provided in this embodiment of the invention determines whether the accurate final value of the model output result is not lower than the accurate threshold of the model output result after obtaining the accurate final value of the model output result. If yes, multi-dimensional iterative optimization is not triggered; if no, multi-dimensional iterative optimization is triggered. Then, it is determined whether the number of multi-dimensional iterative optimizations is not lower than the maximum number of multi-dimensional iterative optimizations. If yes, an early warning is issued to the machine learning model; if no, no early warning is issued to the machine learning model.

[0059] Furthermore, based on disease prediction results lacking high reliability, the signal-model processing link is optimized in multiple dimensions. Specifically, the optimization process involves: obtaining the model output accuracy deviation value based on the model output accuracy index and the model output accuracy threshold; matching the model output accuracy deviation value with a time window reduction coefficient to reduce the time window length for weak signal feature extraction, focusing on shorter signal segments, accurately locking the time interval of weak features, reducing the averaging interference of non-target signals (such as noise) on effective features, avoiding cross-time period signal aliasing, and inputting high-value features that better reflect the instantaneous changes of the disease into the model, thus improving prediction accuracy from the signal source; simultaneously, matching the model output accuracy deviation value with a gate control amplification coefficient to increase the gate control in the machine learning model, strengthening the gate control's feature selection intensity, amplifying the weight of effective weak features, suppressing the influence of irrelevant features (such as noise), solving the problem of insufficient model attention to effective features, improving the sensitivity of disease signal recognition from the model processing stage, reducing prediction bias caused by weight imbalance, and gradually improving the reliability of the results.

[0060] The aforementioned acquisition of the model output result accuracy deviation value refers to subtracting the model output result accuracy index from the model output result accuracy threshold. The aforementioned matching of the time window reduction coefficient based on the model output result accuracy deviation value is as follows: The management database presets the time window reduction coefficient corresponding to the accuracy deviation value interval of each model output result. The obtained model output result accuracy deviation value is input into the management database, and the management database can match the corresponding model output result accuracy deviation value interval. The time window reduction coefficient corresponding to this interval is the required reduction coefficient. The obtained time window reduction coefficient is multiplied by the original time window length, and the result is the time window length to be adjusted for weak signal feature extraction. The aforementioned time window reduction coefficient is less than 1, indicating that the time window length needs to be reduced by a certain percentage.

[0061] The above-mentioned matching of gate amplification factors based on the accuracy deviation value of the model output results is as follows: The management database presets the gate amplification factors corresponding to the accuracy deviation value intervals of each model output result. The obtained accuracy deviation values ​​of the model output results are input into the management database, which then matches the corresponding accuracy deviation value intervals. The gate amplification factor corresponding to this interval is the required amplification factor. Multiplying the obtained gate amplification factor by the original gate size yields the gate size in the machine learning model that needs to be adjusted. A gate amplification factor greater than 1 indicates that the gate in the machine learning model needs to be amplified by a certain factor. The accuracy deviation value intervals of each model output result are divided by domain experts or experienced engineers based on past experience. For example, in scenarios with high accuracy requirements, the intervals may be divided more finely to more precisely adjust the gate size.

[0062] Obtain the accuracy index of the model output after multi-dimensional optimization, and mark it as the final accurate value of the model output to determine whether multi-dimensional iterative optimization is triggered.

[0063] The determination of whether to trigger multi-dimensional iterative optimization is as follows: compare the accurate final value of the model output with the accurate threshold of the model output; if the accurate final value of the model output is not lower than the accurate threshold of the model output, it is determined that multi-dimensional iterative optimization is not triggered; if the accurate final value of the model output is lower than the accurate threshold of the model output, it is determined that multi-dimensional iterative optimization is triggered; the above-mentioned triggering of multi-dimensional iterative optimization refers to performing multi-dimensional optimization on the signal-model processing link, and the specific optimization process is consistent with the multi-dimensional optimization of the signal-model processing link mentioned above.

[0064] The number of multi-dimensional iteration optimizations is counted and compared with the preset maximum number of multi-dimensional iteration optimizations in the management database. When the number of multi-dimensional iteration optimizations is not less than the maximum number of multi-dimensional iteration optimizations, an early warning is issued for the machine learning model. When the number of multi-dimensional iteration optimizations is less than the maximum number of multi-dimensional iteration optimizations, no early warning is issued for the machine learning model. The aforementioned early warning for the machine learning model refers to pushing visual early warning information to the management terminal (such as the operation and maintenance platform). The early warning information includes, but is not limited to, the final time window length, the gating increase coefficient, etc.

[0065] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0066] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0067] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A machine learning-based intelligent detection system for disease surveillance platforms, characterized in that: include: The signal acquisition and processing module is used to acquire raw physiological signal data in real time through multimodal sensors and integrate it into a raw physiological signal dataset. The raw physiological signal dataset is preprocessed and subjected to preliminary noise suppression to obtain a physiological signal dataset. The physiological signal dataset is segmented and analyzed according to a preset time window, and the initial weak signal is automatically identified. The raw physiological signal dataset is used to provide basic data support for disease monitoring. The signal enhancement and model prediction module is used to enhance the identified initial weak signal to obtain a weak signal. The weak signal is then input into the machine learning model trained in the disease monitoring platform. At the same time, the weak signal is classified based on its signal-to-noise ratio, thereby outputting the disease prediction result of the machine learning model. The model output optimization module is used to monitor the output process parameters of the machine learning model, determine whether the disease prediction results have high credibility, solidify the weak signal enhancement strategy and machine learning model parameters based on the disease prediction results with high credibility, and optimize the signal-model processing link in multiple dimensions based on the disease prediction results with low credibility, thereby realizing intelligent detection of the disease monitoring platform based on machine learning.

2. The intelligent detection system for a disease monitoring platform based on machine learning according to claim 1, characterized in that, The specific analysis process for enhancing the identified initial weak signal is as follows: The weak signal is amplified by adaptive gain based on signal optimization factor matching, thereby increasing the strength of the weak signal and finally obtaining the weak signal. The signal optimization factor is obtained by weighted averaging of local signal energy, local volatility, and cross-modal consistency coefficient, and is used to characterize the comprehensive characteristics of the signal in three dimensions: strength, stability, and multimodal cooperation.

3. The intelligent detection system for a disease monitoring platform based on machine learning according to claim 1, characterized in that, The classification process based on the signal-to-noise ratio of weak signals is as follows: When the weak signal has a high signal-to-noise ratio, the weak signal is input into the high signal-to-noise ratio chain in the machine learning model for feature extraction processing, thereby obtaining the weak signal features. The activation entropy of the weak signal feature extraction process is monitored, and it is determined whether the feature vector component optimization strategy of the weak signal is triggered. The activation entropy is used to characterize the uncertainty of the activation distribution of the feature vector components during the weak signal feature extraction process. When the weak signal has a low signal-to-noise ratio, the weak signal is input into the low signal-to-noise ratio chain in the machine learning model for feature extraction processing to obtain the weak signal features. The signal response energy of the weak signal feature extraction process is monitored, and it is determined whether the short-time window signal amplitude optimization strategy is triggered. The signal response energy is used to characterize the degree of energy accumulation of the weak signal in the local time window after feature extraction. The term "high signal-to-noise ratio" refers to a signal-to-noise ratio of a weak signal that is greater than or equal to the defined signal-to-noise ratio. The term "low signal-to-noise ratio" refers to a signal-to-noise ratio of a weak signal that is less than the defined signal-to-noise ratio. The defined signal-to-noise ratio is a value used to distinguish between high and low signal-to-noise ratios.

4. The intelligent detection system for a machine learning-based disease monitoring platform according to claim 3, characterized in that, The specific determination process for the feature vector component optimization strategy to determine whether a weak signal is triggered is as follows: The activation entropy of the weak signal feature extraction process in the high signal-to-noise ratio chain is compared with the activation entropy threshold, where the activation entropy threshold represents the maximum value of the activation entropy within a specified range. When the activation entropy is not lower than the activation entropy threshold, the feature vector component optimization strategy for weak signals is determined. After the feature vector component optimization strategy for weak signals is completed, the activation entropy is re-acquired and the feature vector component optimization strategy is determined and processed. At the same time, based on the activation entropy and the activation entropy threshold, the activation entropy deviation value is obtained, and the signal optimization factor is corrected based on the activation entropy deviation value. When the activation entropy is lower than the activation entropy threshold, it is determined that the feature vector component optimization strategy for weak signals will not be triggered.

5. The intelligent detection system for a machine learning-based disease monitoring platform according to claim 4, characterized in that, The specific analysis process for re-acquiring the activation entropy and determining and processing the feature vector component optimization strategy is as follows: If the optimized activation entropy decreases and falls below the activation entropy threshold, the feature vector component optimization strategy is marked as an effective strategy, and the high signal-to-noise ratio chain is solidified. If the optimized activation entropy decreases but does not fall below the activation entropy threshold, the feature vector component optimization strategy is marked as a strategy to be optimized, and the weak signal is continuously optimized. The continuous optimization refers to iteratively filtering the feature vector components of the weak signal. If the optimized activation entropy increases, the feature vector component optimization strategy is marked as invalid, and the parameters are rolled back.

6. The intelligent detection system for a machine learning-based disease monitoring platform according to claim 3, characterized in that, The specific process for determining whether the short-time window signal amplitude optimization strategy is triggered is as follows: The signal response energy of the weak signal feature extraction process in the low signal-to-noise ratio chain is compared with the signal response energy threshold. When the signal response energy is lower than the signal response energy threshold, it is determined that the short-time window signal amplitude optimization strategy is triggered, the optimized signal response energy is obtained, and the short-time window signal amplitude optimization strategy is judged and processed. At the same time, based on the signal response energy and the signal response energy threshold, the signal response energy deviation value is obtained, and the signal optimization factor is corrected based on the signal response energy deviation value. If the signal response energy is not lower than the signal response energy threshold, then the short-time window signal amplitude optimization strategy will not be triggered.

7. The intelligent detection system for a disease monitoring platform based on machine learning according to claim 6, characterized in that, The specific analysis process for obtaining the optimized signal response energy and determining and processing the short-time window signal amplitude optimization strategy is as follows: If the optimized signal response energy increases but is not lower than the signal response energy threshold, the short-time window signal amplitude optimization strategy is marked as an effective strategy, and the low signal-to-noise ratio chain is solidified. If the optimized signal response energy increases but is below the signal response energy threshold, the short-time window signal amplitude optimization strategy is marked as a strategy to be optimized and continuously optimized. The continuous optimization refers to using signal reconstruction technology to recover the characteristics of weak signals. If the optimized signal response energy decreases, the short-time window signal amplitude optimization strategy is marked as invalid, and the parameters are rolled back.

8. The intelligent detection system for a machine learning-based disease monitoring platform according to claim 1, characterized in that, The specific analysis process for the output parameters of the monitoring machine learning model is as follows: The output parameters of the machine learning model include the model response energy ratio coefficient, the activation entropy ratio coefficient, and the local signal-to-noise ratio ratio coefficient. By pre-setting the effect coefficients of the model response energy ratio coefficient, the activation entropy ratio coefficient, and the local signal-to-noise ratio ratio coefficient in the management database, the weight contribution values ​​of the model response energy ratio coefficient, the activation entropy ratio coefficient, and the local signal-to-noise ratio ratio coefficient to the accuracy index of the model output result are quantified. Finally, a weighted average fusion algorithm is used to synthesize the accuracy index of the model output result, where the accuracy index of the model output result represents the accuracy of the disease prediction results output by the machine learning model. The model response energy ratio coefficient refers to the ratio of the model response energy to the bounded model response energy; the activation entropy ratio coefficient represents the ratio of the activation entropy to the bounded activation entropy; and the local signal-to-noise ratio ratio coefficient represents the ratio of the local signal-to-noise ratio to the bounded local signal-to-noise ratio. The accuracy index of the model output result is compared with the accuracy threshold of the model output result. The accuracy threshold of the model output result refers to the minimum value of the model output result accuracy index within a specified range preset in the management database. When the accuracy index of the model output result is not lower than the accuracy threshold of the model output result, the disease prediction result is considered to have high credibility. When the accuracy index of the model output is lower than the accuracy threshold of the model output, the disease prediction results are not considered to have high reliability.

9. The intelligent detection system for a disease monitoring platform based on machine learning according to claim 1, characterized in that, The signal-model processing link is optimized in multiple dimensions based on disease prediction results that lack high reliability. The specific optimization process is as follows: Based on the accuracy index and accuracy threshold of the model output results, the accuracy deviation value of the model output results is obtained. Based on the accuracy deviation value of the model output results, a time window reduction coefficient is matched to reduce the time window length for weak signal feature extraction. At the same time, based on the accuracy deviation value of the model output results, a gate control increase coefficient is matched to increase the gate control in the machine learning model. Obtain the accuracy index of the model output after multi-dimensional optimization, and mark it as the final accurate value of the model output to determine whether multi-dimensional iterative optimization is triggered.

10. The intelligent detection system for a disease monitoring platform based on machine learning according to claim 9, characterized in that, The specific process for determining whether multi-dimensional iterative optimization has been triggered is as follows: Compare the accurate final value of the model output with the accurate threshold of the model output; If the final accurate value of the model output is not lower than the accurate threshold of the model output, then it is determined that multi-dimensional iterative optimization will not be triggered. When the final accurate value of the model output is lower than the accurate threshold of the model output, it is determined that multi-dimensional iterative optimization is triggered. The number of multi-dimensional iterations was counted and compared with the preset maximum number of multi-dimensional iterations. If the number of multi-dimensional iterative optimizations is not less than the maximum number of multi-dimensional iterative optimizations, then an early warning will be issued for the machine learning model. If the number of multi-dimensional iterations is less than the maximum number of multi-dimensional iterations, no warning will be issued for the machine learning model.

Images