Ecological network stability monitoring method based on machine learning

Abstract

The invention provides an ecological network stability monitoring method based on machine learning, and the method comprises the steps: collecting a real-time ecological data sequence from a wetland bird habitat environment, extracting a seasonal period frequency through employing a frequency domain analysis method according to the environment temperature change, and carrying out the calculation of the seasonal period frequency; decomposing the real-time ecological data sequence into a periodic fluctuation part and a long-term change part; for the temperature fluctuation data, calculating an amplitude peak value interval, comparing the amplitude peak value interval with historical data, performing matching judgment in combination with preset reference model parameters, and generating a classified fluctuation marking result including a potential disturbance type and a natural period; for trend data, trend line fitting slopes in all windows are calculated, slope change distribution is obtained, in combination with a fluctuation marking result and the slope change distribution, a long-term temperature trend deviation state is judged, dynamic monitoring parameters are adjusted, the deviation degree of current temperature data and historical data is monitored, and when the deviation degree exceeds the adjusted range, the dynamic monitoring parameters are determined. And triggering a habitat environment stability transition signal, extracting a persistence characteristic in the transition signal, calculating a persistence index, and recording and confirming a final offset signal according to the persistence index and an offset event, thereby completing refined monitoring and early warning of wetland temperature change.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for monitoring the stability of ecological networks based on machine learning. Background Technology

[0002] Ecological network stability monitoring is a crucial foundation for maintaining biodiversity and ecosystem function. It analyzes interspecies interactions and changes in key indicators to provide early warnings of potential risks, supporting ecological conservation decision-making. Currently, many monitoring methods rely on pre-set machine learning models for periodic updates. However, ecosystem data generation is not uniform; its inherent dynamism is characterized by a coexistence of strong seasonal fluctuations and slow long-term evolutionary trends. This presents a significant challenge to existing methods: if fixed data accumulation amounts or characteristic change thresholds are used to trigger model updates, it becomes difficult to distinguish whether data changes stem from normal ecological rhythms or genuine shifts indicating instability.

[0003] This challenge raises two closely related core technical difficulties. First, there's the multi-scale nature of ecological indicator variation; the amplitude and pattern of natural fluctuations in the same indicator differ significantly across different time scales such as daily, monthly, and yearly. Second, there's the heterogeneity of data components; monitoring data simultaneously includes long-term trend evolution, periodic seasonal fluctuations, and random interference noise. Because multi-scale variation cannot be effectively analyzed and heterogeneous components removed, fixed triggering mechanisms may misinterpret many natural fluctuations as anomalous signals when the ecosystem is in a stable seasonal cycle, leading to frequent and unnecessary model updates, consuming computational resources, and potentially introducing noise. Conversely, when the ecosystem is truly disturbed or enters a sensitive transition phase, the inherent threshold may be too insensitive to capture early, weak instability signals in time, causing a lag in monitoring response.

[0004] Therefore, designing a threshold adjustment mechanism that can automatically adapt to the dynamic changes of ecosystems at multiple scales, so that it can both filter out data disturbances caused by natural fluctuations such as seasonality and keenly capture the real signals that characterize stability changes, thus achieving a balance between resource conservation and monitoring sensitivity, has become a key issue in improving the practicality and reliability of ecological network stability monitoring methods. Summary of the Invention

[0005] This invention provides a machine learning-based method for monitoring the stability of ecological networks, mainly comprising:

[0006] By collecting real-time ecological data sequences, frequency domain analysis was applied to extract seasonal cycle frequencies for environmental temperature changes in specific wetland bird habitats. Combined with a time window segmentation method, the data was decomposed into periodic fluctuations and long-term variations, resulting in separated temperature fluctuation and trend data. Based on the separated temperature fluctuation data, the amplitude peak intervals were calculated and matched against preset benchmark model parameters using historical data comparison. If the amplitude peak interval exceeded the tolerance range of the benchmark model, it was labeled as a potential disturbance type using fluctuation anomaly marking logic; otherwise, it was classified as a natural cycle, resulting in categorized fluctuation marking results. For the categorized fluctuation marking results, a disturbance probability assessment method was introduced to calculate the probability distribution of deviations from the benchmark for potential disturbance types. Simultaneously, for the long-term variation portion, a sliding window length setting and window smoothing were applied to calculate the slope of the trend line fitting within each window, obtaining a preliminary slope change distribution. Based on the preliminary slope change distribution, the slope change rate between consecutive windows was calculated. If the slope change rate of multiple consecutive windows exceeded a preset offset threshold, a long-term temperature trend offset was determined using trend direction identification and continuous window verification methods, resulting in an offset confirmation status. Using the offset confirmation status and fluctuation marking results, dynamic monitoring parameters are adjusted based on the potential disturbance type and offset status. If the offset confirmation status shows an abnormal upward trend, the monitoring tolerance range is reduced through marking and classification logic to improve detection sensitivity, resulting in adjusted monitoring parameters. For the adjusted monitoring parameters, the deviation between current wetland temperature data and historical data is monitored. If the deviation exceeds the adjusted tolerance range, a habitat environmental stability change signal is triggered using the offset event recording method, determining the trigger status of the change signal. Based on the trigger status of the change signal, persistence features are extracted from the signal, and its persistence index is calculated using time series correlation analysis. If the persistence index is below a preset threshold, it is confirmed as a true offset through offset event recording and trend direction marking, obtaining the final offset confirmation signal.

[0007] Furthermore, historical data is compared with preset benchmark model parameters. If the peak amplitude range exceeds the tolerance range of the benchmark model, it is marked as a potential disturbance type through the fluctuation anomaly marking logic; otherwise, it is classified as a natural cycle, and the classified fluctuation marking results are obtained.

[0008] Furthermore, a perturbation probability assessment method is introduced to calculate the probability distribution of its deviation from the baseline. At the same time, a sliding window length setting and window smoothing are applied to the long-term variation part to calculate the slope of the trend line fitting within each window, thus obtaining a preliminary slope variation distribution.

[0009] Furthermore, if the slope change rate of multiple consecutive windows exceeds the preset offset threshold standard, then by combining the trend direction indicator and the continuous window verification method, it is determined to be a long-term temperature trend offset, and an offset confirmation status is obtained.

[0010] Furthermore, the dynamic monitoring parameters are adjusted. If the offset confirmation status shows an abnormal upward trend, the monitoring tolerance range is reduced through the labeling and classification logic to improve detection sensitivity, resulting in the adjusted monitoring parameters.

[0011] Furthermore, if the deviation value exceeds the adjusted tolerance range, the habitat environmental stability change signal is triggered by combining the offset event recording method to determine the triggering state of the change signal.

[0012] Furthermore, time series correlation analysis is applied to calculate its persistence index. If the persistence index is lower than the preset threshold, it is confirmed as a true shift by offset event records and trend direction indicators, thus obtaining the final shift confirmation signal.

[0013] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0014] This invention discloses a dynamic monitoring and trend analysis method for environmental temperature changes in wetland bird habitats, aiming to solve the problem of accurately identifying periodic fluctuations and long-term trend shifts in temperature data in wetland ecosystems and providing early warning of abnormal disturbances. This invention extracts seasonal periodic frequencies through frequency domain analysis, decomposing temperature data into periodic fluctuations and long-term variations. Combining amplitude peak interval matching and probability distribution assessment, it accurately classifies fluctuation types and marks potential disturbances. Simultaneously, for the long-term variation component, it uses a sliding window and trend line fitting to calculate the slope change rate and determine the trend shift status. Based on this, the invention dynamically adjusts monitoring parameters to improve detection sensitivity and confirms the true shift through deviation analysis and persistence index assessment, triggering environmental stability change signals. Ultimately, this invention achieves refined monitoring and early warning of wetland temperature changes, providing a scientific basis and technical support for habitat ecological protection. Attached Figure Description

[0015] Figure 1 This is a flowchart of a machine learning-based method for monitoring the stability of ecological networks according to the present invention. Figure 2 This is a schematic diagram of a data decomposition method according to the present invention; Figure 3 This is a flowchart of a disturbance analysis and slope calculation method according to the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0017] like Figures 1-3 This embodiment of a machine learning-based method for monitoring the stability of ecological networks may specifically include:

[0018] Example 1

[0019] This invention provides a machine learning-based method for monitoring the stability of ecological networks.

[0020] Step S101: Decompose the data into a periodic fluctuation component and a long-term variation component.

[0021] In this embodiment, Figure 2 This is a schematic diagram of a data decomposition method according to an embodiment of the present invention, such as... Figure 2 As shown, by collecting real-time ecological data sequences, frequency domain analysis was applied to extract seasonal cycle frequencies for environmental temperature changes in specific wetland bird habitats. Combined with a time window segmentation method, the collected data was decomposed into periodic fluctuations and long-term variations, thus obtaining separated temperature fluctuation and trend data. This data decomposition aims to better isolate natural rhythms and potential abnormal trends in the environment.

[0022] Optionally, for specific wetland bird habitats, environmental temperature changes are recorded in real time using sensors or monitoring systems to obtain temperature sequence values ​​for the wetland bird habitat. A frequency domain graph is then plotted based on these values, and the frequency values ​​corresponding to seasonal cycles are determined from the graph. The frequency domain graph visually reflects the periodic characteristics of temperature changes, facilitating the identification of major seasonal fluctuation components. The length of the time window is defined using the frequency values. Based on the length of the time window, a decomposition method, such as a time series decomposition algorithm based on an additive or multiplicative model, is applied to the temperature sequence values ​​to obtain fluctuation and trend quantities. The advantage is that it can visually identify the major seasonal fluctuation components, avoiding misjudging normal natural rhythms as abnormal shifts in the ecosystem. The decomposition method can be a time series decomposition algorithm based on an additive or multiplicative model. This algorithm can decompose the original temperature sequence into fluctuation quantities representing short-term changes (for subsequent fluctuation labeling) and trend quantities representing long-term trends (for slope change analysis), thereby separating natural rhythms from abnormal trends.

[0023] In one possible implementation, for a specific wetland bird habitat, environmental temperature changes are recorded in real time using sensors or monitoring systems to obtain the time-domain signal of the wetland bird habitat's temperature sequence. Frequency domain analysis methods, such as Fast Fourier Transform, are applied to process this sequence, transforming it from a time dimension to a frequency dimension, and a frequency domain graph reflecting the periodic characteristics of temperature changes is plotted. By observing the frequency domain graph, the main energy peaks or frequency components in the graph are identified. Since seasonal fluctuations (such as daily and annual fluctuations) have significant characteristics in the frequency domain, the frequency values ​​corresponding to the seasonal cycles (i.e., the periodic frequency) can be determined. Step S102 yields the classified fluctuation labeling results.

[0024] Based on the separated temperature fluctuation data, the peak amplitude range is calculated and matched with preset benchmark model parameters using a historical data comparison method. In this embodiment, if the peak amplitude range exceeds the tolerance range of the benchmark model, it is labeled as a potential disturbance type through fluctuation anomaly labeling logic; otherwise, it is classified as a natural cycle, resulting in a categorized fluctuation labeling result. The preset benchmark model parameters are normal fluctuation thresholds set based on long-term historical observation data. These parameters constitute the benchmark model tolerance range, used to assess whether the fluctuations in current ecological data (such as temperature) are within natural rhythms. The thresholds are dynamically set and typically depend on regional differences, species tolerance levels, and historical statistical distributions.

[0025] Specifically, temperature fluctuation data is obtained from raw records, and the fluctuation sequences for each time period are separated using a segmented processing method to obtain a preliminary set of fluctuation segments. For each fluctuation segment set, the peak amplitude and corresponding peak range of each segment are calculated, and statistical tools are used to extract the maximum and minimum value ranges to determine the amplitude feature set. This amplitude feature set contains the extreme value information of temperature fluctuations within each time segment. Historical data archives are accessed, and the real-time extracted amplitude feature set is compared and matched numerically with the preset benchmark model parameters one by one to determine whether the current fluctuation value falls within the standard tolerance range set by the benchmark model. If the matching result exceeds the tolerance range, the segment is labeled using anomaly marking logic and classified as a potential disturbance type, obtaining an anomaly classification label. This means that the temperature fluctuation amplitude within this time period is abnormally large and may be affected by external interference. If the matching result is within the tolerance range, the segment is classified as a natural cycle type, identified using classification rules, and a cycle classification label is determined. This indicates that the temperature fluctuation within this time period belongs to a normal natural rhythm. By integrating the anomaly classification label and the cycle classification label, a complete fluctuation marking result is generated, and the final classification information is saved using data storage tools for subsequent steps. By using this historical benchmark-based matching method, the system can effectively distinguish seasonal fluctuations in wetland environments from real unstable offset signals.

[0026] Step S103 yields the comprehensive analysis results of the classified fluctuation labels and disturbance effects.

[0027] After obtaining the categorized fluctuation labels, a disturbance probability assessment method is introduced to calculate the probability distribution of deviations from the baseline for potential disturbance types. Simultaneously, a sliding window length setting and window smoothing are applied to the long-term variation portion, and the slope of the trend line fitting within each window is calculated to obtain a preliminary slope variation distribution.

[0028] Specifically, Figure 3 This is a flowchart of a disturbance analysis and slope calculation method according to an embodiment of the present invention, such as... Figure 3 As shown, the categorized fluctuation data is used to initially screen for disturbance types. Statistical analysis methods are employed to calculate the probability distribution of each type's deviation from the baseline, yielding the probability distribution results. Based on these probability distribution results, for disturbance types with significant deviations from the baseline, their corresponding time series data are obtained. A sliding window method is applied to divide long-term variation intervals, determining the window length range for each interval. This means the system does not randomly cut data but rather sequentially extracts data segments along the time axis through a moving window, thus discretizing the continuous long-term temperature change process into multiple continuous, calculable observation intervals. For the divided long-term variation intervals, smoothing techniques are used to denoise the data within the window, resulting in smoothed time series data. The smoothing technique can be moving average or exponential smoothing, used to eliminate the influence of random noise on trend analysis. From the smoothed time series data, for each data point within the sliding window, the slope value of the trend fit is calculated, obtaining the set of fitting slopes for each window. Based on the fitted slope set, the distribution characteristics of slope changes within each window are analyzed. If the slope value of a certain window deviates from the preset threshold of the overall distribution (for example, the Z-score method can be used, setting the threshold to 2 or 3 times the standard deviation of the overall slope mean), then the window is marked as an abnormal change interval, and the identifier of the abnormal change interval is obtained. For the identifier of the abnormal change interval, cross-validation is performed in conjunction with the probability distribution results to determine the correlation between the disturbance type and the slope change, and to determine the final disturbance impact range. Based on the final disturbance impact range, key fluctuation markers in long-term changes are extracted, and a mapping relationship between fluctuations and disturbance types is constructed through data modeling to obtain the comprehensive analysis results of the classified fluctuation markers and disturbance impact.

[0029] In one possible implementation, the probability distribution of each type of disturbance deviating from the baseline is initially calculated using the classified fluctuation label data to obtain the probability distribution results. Here, each type refers to the subdivision of potential disturbances under different dimensions, such as time dimension type (i.e., potential disturbance segments that occur during different time periods, such as different seasons or specific weather events), feature dimension type (disturbance subclasses based on features such as amplitude peak, duration or frequency of change), and correlation type (disturbances with different degrees of correlation with long-term trends (slope changes).

[0030] In one possible implementation, the data points within the defined long-term variation interval (sliding window) are first subjected to noise reduction and smoothing processing (such as moving average or exponential smoothing) to eliminate the influence of random noise on trend analysis. Then, a linear regression algorithm is used to calculate the slope value of the trend fit, and the set of fitted slopes for each window is obtained. The slope value calculated by linear regression can effectively reflect the rate (numerical magnitude) and direction (positive and negative) of temperature change within each specific time window.

[0031] In one possible implementation, the identification of abnormal change intervals is combined with probability distribution results for cross-validation to determine the correlation between the disturbance type and the slope change, thus determining the final disturbance impact range. The cross-validation aims to eliminate random noise and confirm the authenticity of abnormal signals through the alignment of multi-dimensional data. The specific steps are as follows: First, obtain the identifiers marked as abnormal change intervals in the slope analysis. Retrieve the corresponding disturbance probability distribution results within the same time period (i.e., the probability of fluctuations deviating from the baseline within that time period). Align and verify the time window where the abnormal slope occurs with the extreme deviation time points displayed in the probability distribution to see if they overlap or have continuity on the time axis. The correlation judgment is achieved by analyzing the coupling relationship between fluctuations (short-term) and trends (long-term). Specifically, the system analyzes the disturbance types in the probability distribution results that deviate significantly from the baseline to see if they are accompanied by significant slope deviations. If, within a certain time period, the deviation probability of temperature fluctuations is extremely high, and simultaneously, the fitted slope of that interval also deviates from the preset threshold of the overall distribution, then it is determined that the disturbance type and the slope change have a strong correlation. If there are only abnormal fluctuations with a stable slope, or slight changes in slope but fluctuation probabilities within the normal range, the correlation is considered weak, possibly belonging to isolated random noise or simple seasonal adjustments. Determining the final disturbance impact range is to quantify the temporal and degree of impact of the abnormal event. Specifically, based on the final determined correlation results, key fluctuation markers that truly cause trend shifts are extracted from long-term changes. A mapping relationship is constructed between these fluctuations and specific disturbance types (such as extreme weather, human interference, etc.). Based on the start and end points of the slope anomaly in the continuous window, combined with the time point of probability distribution returning to normal, the specific coverage interval of the disturbance in the time series is determined, thus obtaining the final disturbance impact range. This impact range determined through cross-validation provides accurate data support for the subsequent step S104 to determine whether it belongs to a long-term temperature trend shift.

[0032] In one possible implementation, key fluctuation markers in long-term changes are extracted based on the final disturbance impact range. A mapping relationship between fluctuations and disturbance types is constructed through data modeling to obtain a comprehensive analysis result of the classified fluctuation markers and disturbance impacts. For example, a descriptive marker is assigned to each signal segment confirmed as a true shift. For instance, if a fluctuation has the characteristics of "high amplitude, low probability, and strong slope correlation," it is mapped to a "strong disturbance event." Then, a feature matrix is ​​constructed: the input includes the identifier of the key fluctuation, the time range of its impact, the probability distribution value, and the rate of change of the associated slope. The output is the corresponding classification type (potential disturbance or natural cycle). Through the above mapping relationship, the system finally generates a comprehensive analysis result of the classified fluctuation markers and disturbance impacts. This result not only records what happened during the fluctuation but also quantifies the specific degree of impact of the fluctuation on the stability (long-term trend) of the wetland ecosystem.

[0033] Step S104: Obtain the offset confirmation status.

[0034] The slope change rate between consecutive windows is calculated based on the initial slope change distribution. In this embodiment, if the slope change rate of multiple consecutive windows exceeds a preset offset threshold standard, the trend direction indicator and the consecutive window verification method are combined to determine that it is a long-term temperature trend offset, and an offset confirmation status is obtained.

[0035] Specifically, temperature time-series data is acquired, and the time series is segmented using a sliding window method to obtain multiple continuous data windows. For each data window, linear regression is used to calculate the slope of the data within that window, resulting in a window slope sequence. Linear regression effectively quantifies the rate and direction of temperature change within each window. The difference in slope between adjacent windows is calculated to obtain a sequence of slope change rates between windows, reflecting the acceleration or deceleration of the temperature change trend. If the slope change rate of three consecutive windows is greater than a preset positive offset threshold, a positive trend offset candidate is identified, which usually indicates that the temperature is accelerating upwards. If the slope change rate of three consecutive windows is less than a preset negative offset threshold, a negative trend offset candidate is identified, which usually indicates that the temperature is accelerating downwards. The thresholds need to be set based on historical background data of specific wetland bird habitats. The positive offset threshold is used to determine whether the temperature is in a state of accelerating upwards; the negative offset threshold is used to determine whether the temperature is in a state of accelerating downwards. Based on the slope change direction corresponding to the offset candidate window, the confirmation status of the long-term temperature trend offset is determined, thus providing a basis for subsequent adjustment of monitoring parameters.

[0036] In one possible implementation, the acquisition of temperature time series data involves using a sliding window method to segment the time series, resulting in multiple continuous data windows. This aims to discretize long-term continuous temperature data into comparable analytical units. Specifically, the length of the time window is defined using the seasonal cycle frequency determined through frequency domain analysis in step S101. This ensures that each window covers a complete natural rhythm (such as a season or a specific biological cycle). Furthermore, a resliding or adjacent sliding method is typically used to determine the sliding step size to guarantee continuity between windows. For example, the system acquires preprocessed (e.g., noise reduction and smoothing) wetland temperature time series data. According to the set window length, data is extracted starting from the beginning of the sequence, and then the window moves forward along the time axis, sequentially extracting multiple continuous data windows. The core of this segmentation method lies in using the cycle frequency to scientifically set the window size, thereby accurately capturing long-term trend shift signals representing system instability amidst complex seasonal fluctuations.

[0037] Step S105: Adjust the dynamic monitoring parameters.

[0038] In this embodiment, the offset confirmation status and fluctuation marking results are used to adjust the dynamic monitoring parameters according to the potential disturbance type and offset status. If the offset confirmation status shows an abnormal upward trend, the monitoring tolerance range is reduced through the marking classification logic to improve the detection sensitivity, and the adjusted monitoring parameters are obtained.

[0039] Specifically, for the data on disturbance type and offset state, the corresponding historical fluctuation marker records are obtained from the pre-established database. The comparative analysis method is used to determine whether the current offset state conforms to the abnormal upward trend direction. The comparative analysis method can effectively identify whether the current trend deviates from the historical normal pattern.

[0040] In one possible implementation, the pre-established database is constructed based on historical fluctuation marker records. Its specific construction basis includes: long-term observation data accumulation: integrating historical long-term monitoring data of environmental temperature in specific wetland bird habitats; fluctuation classification results: including classification labels of natural cycle types and potential disturbance types generated in previous steps (such as S102) and their corresponding amplitude feature sets; normal pattern benchmark: the database stores the normal fluctuation thresholds and historical evolution patterns of the ecosystem in a stable state, as a reference system for determining whether the current state deviates from the normal state.

[0041] The specific implementation method for determining whether the current offset state conforms to the abnormal upward trend direction using the comparative analysis method is as follows: First, the system extracts and aligns the corresponding historical fluctuation marker records from the database for the current disturbance type and offset state data. The comparative analysis method compares the "current offset state" with the "historical normal pattern" one by one in terms of both value and direction. For example, the system checks whether the currently identified offset candidates (from S104, e.g., three consecutive windows with a slope change rate greater than the positive offset threshold) significantly deviate from the historical variation pattern under similar backgrounds. If the current offset state shows a continuous positive increase in direction, and its rate of increase (slope change rate) or magnitude exceeds the historical normal range recorded in the database, then the current trend direction is determined to be abnormally upward.

[0042] Furthermore, if the trend is judged to be abnormally upward, the monitoring tolerance is narrowed through classification and labeling logic to obtain an adjusted monitoring tolerance range. This means the system will be more sensitive to minute temperature changes, thus enabling it to detect potential ecological risks earlier. Based on the adjusted monitoring tolerance range, combined with the dynamic monitoring mechanism, the configuration of monitoring parameters is updated, such as tightening the tolerance, increasing sensitivity, and triggering closed-loop, to obtain updated monitoring parameter values. For the updated monitoring parameter values, a support vector machine algorithm is used to optimize and adjust the detection sensitivity, resulting in an optimized detection sensitivity configuration. Through the optimized detection sensitivity configuration, the real-time data stream is dynamically monitored to determine if there are any potential abnormal disturbance types. If an abnormal disturbance type is detected during dynamic monitoring, the parameter adjustment process is triggered according to preset threshold rules to obtain the final monitoring parameter adjustment scheme.

[0043] The final monitoring parameter adjustment scheme is applied to the real-time monitoring system to continuously track and update the offset status and trend direction, ensuring that the monitoring system can adapt to dynamic changes in the environment.

[0044] In one possible implementation, the specific method for updating the monitoring parameter configuration based on the adjusted monitoring tolerance range and combined with the dynamic monitoring mechanism is as follows: tightening the tolerance, that is, when S104 confirms an abnormal increase in the long-term trend, the system considers the ecosystem to have entered a sensitive or disturbed period, and at this time, the tolerance range is reduced. Reduced tolerance means the system is more sensitive to small temperature changes; the purpose of this adjustment is to capture potential ecological risks or instability signals earlier. The updated parameters are applied to the real-time monitoring system. If the deviation value of the real-time data exceeds this adjusted, stricter tolerance range, the system will formally trigger a habitat environmental stability change signal using a migration event recording method. The monitoring parameters are a comprehensive configuration scheme centered on monitoring tolerance, combined with detection sensitivity (SVM optimization) and dynamic triggering rules. By adjusting these parameters, the system transitions from a conventional monitoring mode to a high-sensitivity alert mode. Furthermore, a training sample set is constructed based on the updated monitoring parameter values; a support vector machine regression algorithm is used, with the monitoring parameter values ​​as input features and the historical best detection sensitivity value as the output target, to train the model; the current monitoring parameter values ​​are input into the trained model to calculate the recommended detection sensitivity value, and this value is used as the optimized detection sensitivity configuration.

[0045] In one possible implementation, the optimized detection sensitivity configuration is used to dynamically monitor the real-time data stream and determine whether there are potential abnormal disturbance types. If an abnormal disturbance type is detected during dynamic monitoring, a parameter adjustment process is triggered according to a preset threshold rule to obtain the final monitoring parameter adjustment scheme. Determining whether there are potential abnormal disturbance types is a dynamic process combining machine learning optimization and real-time data comparison. Specifically, the adjusted monitoring parameters (including tightened monitoring tolerance and optimized detection sensitivity) are used to continuously monitor the real-time wetland temperature data stream. The system uses a support vector machine algorithm to optimize the detection sensitivity configuration. Through this algorithm, the system can identify minute numerical fluctuations in the real-time data stream and compare them with historical normal fluctuation patterns. The deviation between the current real-time temperature data and historical benchmark data is monitored. If the characteristics exhibited by the deviation (such as rate of change and amplitude) match the potential disturbance characteristics defined in the previous steps (S102 / S103), the system determines that there is a potential abnormal disturbance type.

[0046] The preset threshold rules are the logical standards for triggering the final monitoring parameter adjustment scheme. Specifically, they include: Deviation triggering rules: If the real-time deviation value exceeds the new monitoring tolerance range after processing based on the offset confirmation status (e.g., abnormal increase), the triggering condition is met. Trend linkage rules: The threshold rules consider not only a single value but also the trend direction indicator. For example, if the current fluctuation is marked as a potential disturbance and simultaneously meets the condition of an abnormally rising trend, the rule will force the system to enter a higher frequency or lower tolerance alert state. Parameter adjustment triggering logic: Once the data performance in dynamic monitoring violates the above-mentioned preset "tolerance + trend" combined threshold, the system will automatically trigger the parameter adjustment process. This includes further optimizing the sensitivity configuration or updating the monitoring frequency to obtain the final monitoring parameter adjustment scheme, ensuring that the monitoring system can adapt to continuous dynamic changes in the environment.

[0047] In summary, the core of anomaly detection lies in using the sensitivity optimized by SVM to capture minute deviations; while the preset threshold rules are based on the dual constraints of the narrowed tolerance range and the trend direction, used to determine whether further adjustments to the system's monitoring configuration are needed.

[0048] Step S106: Determine the trigger state of the transition signal.

[0049] For the adjusted monitoring parameters, monitor the degree of deviation between the current wetland temperature data and historical data. If the deviation value exceeds the adjusted tolerance range, trigger the habitat environmental stability change signal by combining the offset event recording method, and determine the triggering state of the change signal.

[0050] Specifically, real-time wetland temperature data is acquired from the monitoring system and compared with pre-stored historical data to calculate the deviation, yielding preliminary deviation analysis results. This deviation reflects the degree of deviation between the current environmental state and historical norms. If the deviation exceeds a preset tolerance range, it is recorded as an offset event, and the relevant timestamp and data details are stored in the system log to determine the event's content. Based on the offset event's content, a preset recording method is used to classify the event. The classified data is then matched with environmental stability assessment criteria to determine if environmental stability has been affected. If the environmental stability assessment shows an anomaly, a habitat environmental change signal is triggered. Further, signal data is transmitted through an internal message queue to confirm the signal's trigger status. After obtaining the signal's trigger status, the frequency and pattern of signal triggering are analyzed in conjunction with historical offset event data to obtain the signal triggering regularity characteristics. Based on the analysis of these regularity characteristics, a support vector machine algorithm is used to predict potential future change signals, and the prediction results are stored in a database for subsequent queries. Based on the prediction results, early warning information on changes in environmental stability is generated, and the early warning information is pushed to relevant monitoring modules through the system interface to complete the closed-loop processing triggered by the signal.

[0051] In one possible implementation, based on the recorded content of the offset events, a preset recording method, namely system log storage and attribute annotation, is used to classify the events. The classified data is then matched with the environmental stability evaluation criteria. The classification process includes: classification by offset nature: combining the results of S104 and S105, events are classified into abnormal upward trends, accelerated downward trends, or sudden disturbances; classification by severity: based on the calculated deviation value, events are divided into different levels; and classification by frequency pattern: combining the data records of historical offset events, the frequency and pattern of signal triggering (such as isolated events or continuous sequences) are analyzed to obtain the regularity characteristics of signal triggering. The environmental stability assessment criteria are mainly divided into: Tolerance redundancy criteria: the adjusted monitoring tolerance range is used as the core red line. If the real-time data fluctuates outside this range, it is considered a challenge to stability; Regularity characteristic criteria: the frequency and pattern of signal triggering are analyzed using the support vector machine (SVM) algorithm. If the signal shows obvious regularity shifts rather than random noise, it meets the criteria for stability impairment; Persistence characteristic indicators: environmental stability also depends on the persistence of the signal. If the calculated autocorrelation coefficient decay rate (persistence indicator) is low, it means that the shift has long-term persistence, which is a key criterion for judging stability changes.

[0052] In one possible implementation, after acquiring the trigger state of the transition signal, the system synchronously calls the time series records of historical offset events in the database to analyze the frequency and pattern of signal triggering. This involves counting the number of signal triggers within a specific time period, calculating the time interval between triggers, and identifying whether the signal is an occasional or high-frequency, dense trigger. The system also analyzes the correlation between signal triggering and external environmental cycles (such as seasons or day / night cycles) to identify the intensity sequence of signal triggers. For example, it observes whether the deviation value exhibits increasing, decreasing, or cyclical characteristics. Through this analysis, the scattered signals are transformed into descriptive, regular feature vectors. For example, a signal triggering frequency of "3 times / week" and a triggering pattern of "periodic fluctuations accompanied by abnormal temperature increases" collectively constitute a regular profile of the signal.

[0053] In one possible implementation, the specific steps for predicting potential future change signals using a Support Vector Machine (SVM) algorithm based on the analysis of signal triggering regularity features are as follows: Constructing a training sample set: The signal triggering regularity features obtained in the previous step (such as frequency, pattern, historical offset amplitude, slope change rate, etc.) are used as feature inputs to the SVM model. Based on historical results, the data is labeled as either true stability changes or non-change disturbances. SVM model training and classification: The SVM is used to find an optimal hyperplane in the multi-dimensional feature space to maximize the distinction between normal ecological fluctuations and signal features indicating change. The model learns historical patterns to establish a mapping relationship between existing feature combinations and future change probabilities. Prediction execution: The system inputs the latest regularity features obtained from real-time analysis into the trained SVM model. The model outputs a classification result or score to determine whether the current signal sequence is pointing to a potential future stability change event. Conversion into early warning information: If the SVM prediction result shows that the change probability exceeds a preset value, the system will convert the prediction result into early warning information for environmental stability changes, achieving a leap from post-event recording to pre-event warning.

[0054] This method, which combines historical pattern analysis with machine learning classification and prediction, significantly improves the system's lead time and accuracy in detecting signals of ecosystem instability.

[0055] Step S107: Obtain the final offset confirmation signal.

[0056] By analyzing the triggering state of the transition signal, the persistence features in the signal are extracted, and the persistence index is calculated using time series correlation analysis. If the persistence index is lower than the preset threshold, the offset event record and trend direction indicator are used to confirm that it is a true offset, thus obtaining the final offset confirmation signal.

[0057] Specifically, the original time-series signal is acquired, and for the transition trigger point in the signal, signal segments within the time window before and after the trigger point are extracted. The signal segments are analyzed using the autocorrelation function to calculate their autocorrelation coefficient sequence, thereby obtaining the persistence characteristics of the signal. The persistence characteristics reflect the continuity and stability of the signal over time.

[0058] Based on the autocorrelation coefficient sequence, its decay rate is calculated. The decay rate is defined as a persistence index. If the persistence index is lower than a preset threshold, the signal segment is determined to correspond to a candidate offset event. This indicates that the signal is not a short-term random fluctuation, but has a certain persistence.

[0059] The offset event record item corresponding to the time period of the signal segment is obtained from the historical database, and the signal trend direction within the time period is calculated. If the offset event record item exists and the signal trend direction is consistent with the offset direction identifier, the offset is determined to be the true offset, and the offset confirmation signal corresponding to the true offset is output.

[0060] In one possible implementation, the signal segment is analyzed using an autocorrelation function to calculate its autocorrelation coefficient sequence and obtain the persistence characteristics of the signal. When processing a signal segment (denoted as a time series X = {1, 2, ..., n}), the following formula is typically used:

[0061] (1) The autocorrelation function measures the correlation between a signal and its offset over a time interval (lag order) of k. For discrete signals, its calculation formula is usually expressed as:

[0062]

[0063] Where k is the lag time step, μ is the mean of the signal segment, and n is the total number of data points in the signal segment.

[0064] (2) The autocorrelation coefficient sequence is a set of R(k) values ​​calculated when k takes different values ​​(e.g., k = 1, 2, 3...). This sequence reflects how the correlation of the signal disappears over time.

[0065]

[0066] After obtaining the autocorrelation coefficient sequence, the system extracts persistence features by analyzing the decay rate of the sequence. The specific judgment logic is as follows:

[0067] Calculating the decay rate: The persistence characteristic mainly depends on how quickly the autocorrelation coefficient decreases as the lag order k increases. If R(k) decreases very slowly as k increases (long-range correlation), it indicates that the signal has strong memory. If R(k) drops rapidly to near 0 (short-range correlation), it indicates that the signal is random or transient.

[0068] The system extracts persistence indicators, namely, high persistence (stable transition signal): If the autocorrelation coefficient sequence exhibits a low decay rate, it means that the signal is not caused by accidental short-term disturbances, but has some inherent continuous dynamics. In this case, the system determines that the signal has persistent characteristics. Low persistence (random noise): If the decay rate is extremely fast, it is considered that the shift is only temporary and does not represent long-term environmental changes. Finally, the system combines this persistence characteristic with the SVM prediction result in S106 for a comprehensive judgment: If the SVM predicts future changes, and the current signal's autocorrelation coefficient decay rate is low (proving that the signal has persistence), then it is finally confirmed as a real environmental stability shift, triggering an alarm.

[0069] In one possible implementation, the decay rate is calculated based on the autocorrelation coefficient sequence. If the persistence index is lower than a preset threshold, the signal segment is determined to correspond to a candidate offset event. The decay rate reflects how quickly the autocorrelation coefficient R(k) decreases with increasing time step (lag order k). In practical implementation, one of the following two methods is usually adopted:

[0070] Slope fitting method: Perform linear regression on the autocorrelation coefficient sequence {R(1),R(2),...,R(k)}, and the absolute value of the slope of the fitted line is the decay rate.

[0071]

[0072] Integral time-scale method: Calculates the time step taken for the autocorrelation coefficient to drop to a certain threshold (e.g., 1 / e ≈ 0.37 or 0). A longer step size results in a lower decay rate and stronger signal persistence. This implies that the current temperature shift is not a random fluctuation but rather exhibits a certain inertial trend.

[0073] The preset threshold is usually set based on the autocorrelation characteristic distribution of the wetland's historical normal years. For example, the lower quantile (such as the top 5%) of the historical normal data decay rate distribution is taken as the critical line for judging the change.

[0074] In one possible implementation, the process involves retrieving offset event records (including the slope marked in S104, the tolerance adjustment records in S105, etc.) from a historical database for the corresponding time period of the signal segment, and calculating the signal trend direction within that time period. If the offset event record exists and the signal trend direction matches the offset direction identifier, then the offset is determined to be a true offset. The specific calculation method is as follows:

[0075] 1. Segmented slope mean method: Calculate the weighted average of the fitted slopes S of all sliding windows within the time period.

[0076] 2. First-order difference method: The rate of change of the whole is calculated by comparing the smoothed values ​​at the beginning and end of the time period.

[0077] Judgment logic: If the calculation result is positive, the signal trend is upward. If the calculation result is negative, the signal trend is downward.

[0078] The system performs consistency matching between the calculated signal trend direction and the SVM-predicted transition direction: if the autocorrelation analysis proves that the signal is persistent and the extracted trend direction is consistent with the future transition direction predicted by the SVM, then the system finally confirms it as a real environmental stability shift and triggers the final transition signal.

[0079] This method significantly reduces the false alarm rate in ecological monitoring through dual verification of temporal persistence (autocorrelation) and spatial / logical directional consistency (SVM + historical comparison).

[0080] Example 2

[0081] According to an embodiment of the present invention, a monitoring system is also provided for implementing the machine learning-based ecological network stability monitoring method shown in steps S101 to S107 above.

[0082] The monitoring system may include a data acquisition and frequency domain analysis module, a fluctuation amplitude classification module, a disturbance probability and trend slope calculation module, a trend offset judgment module, a monitoring parameter adjustment module, a deviation monitoring and signal triggering module, and an offset confirmation module.

[0083] The data acquisition and frequency domain analysis module is used to collect real-time ecological data sequences, target the environmental temperature changes of specific wetland bird habitats, extract seasonal periodic frequencies using frequency domain analysis, and combine time window segmentation methods to decompose the data into periodic fluctuation parts and long-term change parts, obtaining separated temperature fluctuation data and trend data.

[0084] The fluctuation amplitude classification module is used to calculate the peak amplitude range based on the separated temperature fluctuation data. It matches the peak amplitude range with the preset benchmark model parameters using a historical data comparison method. If the peak amplitude range exceeds the tolerance range of the benchmark model, it is marked as a potential disturbance type through the fluctuation anomaly marking logic. Otherwise, it is classified as a natural cycle, and the fluctuation marking results are obtained after classification.

[0085] The perturbation probability and trend slope calculation module is used to obtain the classified fluctuation labeling results. For potential perturbation types, a perturbation probability assessment method is introduced to calculate the probability distribution of its deviation from the baseline. At the same time, for the long-term change part, the sliding window length setting and window smoothing processing are applied to calculate the trend line fitting slope in each window to obtain the preliminary slope change distribution.

[0086] The trend offset judgment module is used to calculate the slope change rate between consecutive windows based on the preliminary slope change distribution. If the slope change rate of multiple consecutive windows exceeds the preset offset threshold standard, it is judged as a long-term temperature trend offset by combining the trend direction indicator and the consecutive window verification method, and the offset confirmation status is obtained.

[0087] The monitoring parameter adjustment module is used to adjust the dynamic monitoring parameters based on the offset confirmation status and fluctuation marking results, targeting the potential disturbance type and offset status. If the offset confirmation status shows an abnormal upward trend, the monitoring tolerance range is reduced through the marking classification logic to improve detection sensitivity, thus obtaining the adjusted monitoring parameters.

[0088] The deviation monitoring and signal triggering module is used to monitor the degree of deviation between the current wetland temperature data and historical data for the adjusted monitoring parameters. If the deviation value exceeds the adjusted tolerance range, the module will trigger a habitat environmental stability change signal in combination with the offset event recording method to determine the triggering status of the change signal.

[0089] The offset confirmation module is used to extract the persistence features in the signal by the trigger state of the transition signal, and calculate its persistence index by applying time series correlation analysis. If the persistence index is lower than the preset threshold, it is confirmed as a real offset by offset event records and trend direction indicators, and the final offset confirmation signal is obtained.

[0090] It should be noted that the above modules correspond to steps S101 to S107 in Embodiment 1. The seven modules and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in Embodiment 1.

[0091] Obviously, those skilled in the art can make various modifications and variations to the embodiments of this application without departing from the spirit and scope of the embodiments of this application. Therefore, if these modifications and variations to the embodiments of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include these modifications and variations.

Claims

1. A machine learning-based method for monitoring the stability of ecological networks, characterized in that, include: Real-time ecological data sequences were collected from wetland bird habitats. For changes in ambient temperature, frequency domain analysis was used to extract seasonal periodic frequencies. Combined with time window division, the real-time ecological data sequences were decomposed into periodic fluctuation parts and long-term change parts, resulting in separated temperature fluctuation data and trend data. For the temperature fluctuation data, its amplitude peak range is calculated. By comparing it with historical data and combining it with preset benchmark model parameters, a matching judgment is made. When the amplitude peak range exceeds the preset range, it is marked as a potential disturbance type. When the amplitude peak range is within the preset range, it is classified as a natural cycle, and a classified fluctuation labeling result is generated. For the aforementioned trend data, the sliding window method is applied to calculate the slope of the trend line fitting within each window, obtaining the slope change distribution. Combining the fluctuation marking results with the slope change distribution, the long-term temperature trend shift status is determined, dynamic monitoring parameters are adjusted, and the degree of deviation between the current temperature data and historical data is monitored. When the degree of deviation exceeds the adjusted range, a habitat environmental stability change signal is triggered. The persistence characteristics in the change signal are extracted, and the persistence index is calculated. Based on the persistence index and the shift event record, the final shift signal is confirmed.

2. The method for monitoring the stability of ecological networks based on machine learning according to claim 1, characterized in that, The step of calculating the peak range of the temperature fluctuation data and comparing it with historical data, combined with preset benchmark model parameters, includes: Extract the fluctuation sequences of each time period from the temperature fluctuation data to generate a set of fluctuation segments. For each segment in the set of fluctuation segments, calculate its peak amplitude and corresponding interval range to generate a set of amplitude features. Based on the amplitude feature set, historical data archives are called for comparison processing, and numerical matching is performed in combination with preset benchmark model parameters. When the matching result exceeds the preset range, the segment is labeled by the labeling logic and classified as a potential disturbance type. When the matching result is within the preset range, the segment is classified as a natural cycle type, and the classified fluctuation labeling result is generated.

3. The method for monitoring the stability of ecological networks based on machine learning according to claim 1, characterized in that, For the aforementioned trend data, the sliding window method is applied to calculate the slope of the trend line fitting within each window, obtaining the slope variation distribution, including: Extract long-term variation intervals from the trend data, divide the long-term variation intervals using the sliding window method, and set the window length range for each interval. For each window after partitioning, smoothing is performed to reduce data noise and generate smoothed time series data; From the smoothed time series data, the trend fitting slope value within each window is calculated to generate a set of fitting slopes for each window; Based on the fitted slope set, the distribution characteristics of slope changes within each window are analyzed. When the slope value of a certain window deviates from the overall distribution range, the window is marked as an abnormal change interval, and an identifier for the abnormal change interval is generated.

4. The machine learning-based ecological network stability monitoring method according to claim 3, characterized in that, The step of analyzing the slope variation distribution characteristics within each window based on the fitted slope set includes: Extract the slope values ​​of adjacent windows from the fitted slope set, calculate the slope change rate between adjacent windows, and generate a change rate sequence; When the rate of change of multiple consecutive windows exceeds the preset range, the trend direction indicator and the continuous window verification logic are combined to determine that it is a long-term temperature trend deviation, and a deviation confirmation status is generated.

5. The machine learning-based ecological network stability monitoring method according to claim 4, characterized in that, The step of combining trend direction indicators and continuous window verification logic to determine a long-term temperature trend shift and generate a shift confirmation status also includes: Based on the offset confirmation status and fluctuation marking results, adjust the dynamic monitoring parameters according to the potential disturbance type and offset status; When the offset confirmation status shows an abnormal trend direction, the monitoring range is narrowed through classification and marking logic, an adjusted monitoring parameter configuration is generated, and the monitoring mechanism is updated based on the adjusted monitoring parameter configuration.

6. The machine learning-based ecological network stability monitoring method according to claim 5, characterized in that, The configuration update mechanism based on the adjusted monitoring parameters includes: Based on the adjusted monitoring parameters, monitor the deviation between the current wetland temperature data and historical data; When the deviation value exceeds the adjusted range, the deviation is recorded as an offset event, and matched with the environmental stability assessment standard through classification processing to trigger a habitat environmental stability change signal, and the triggering status of the change signal is confirmed.

7. The machine learning-based ecological network stability monitoring method according to claim 6, characterized in that, The triggering of habitat environmental stability change signals and confirmation of the triggering status of the change signals further include: Persistence features are extracted from the triggering state of the transition signal, and persistence indices are calculated using time series correlation analysis. When the persistence index is below a preset range, the transition signal is confirmed as a true offset by combining the offset event record and the trend direction indicator, and a final offset confirmation signal is generated.

Images