Conductivity pollution intelligent diagnosis method based on multi-modal feature fusion and adaptive density clustering

Abstract

The conductivity pollution intelligent diagnosis method based on multi-modal feature fusion and adaptive density clustering is as follows: (1) data processing is carried out through an intelligent data preprocessing module; (2) conductivity data are detected through a multi-scale anomaly detection engine; (3) system extraction is carried out through a multi-modal feature extraction system; (4) data are optimized through an adaptive density clustering algorithm; and (5) diagnosis is carried out through an intelligent diagnosis decision system. The present application provides a conductivity pollution intelligent diagnosis method based on multi-modal feature fusion and adaptive density clustering, which is suitable for application scenarios such as smart water affairs, basin water quality monitoring, industrial discharge supervision and the like.

Description

TECHNICAL FIELD

[0001] The present application belongs to the cross field of environmental monitoring and machine learning, and particularly relates to a conductivity pollution intelligent diagnosis method based on multi-modal feature fusion and adaptive density clustering, which is suitable for smart water, watershed water quality monitoring, industrial emission supervision and other application scenarios. BACKGROUND

[0002] Although the existing multi-parameter water quality monitoring system (pH, ammonia nitrogen, COD, etc.) has comprehensive data, it is high in cost and complex in equipment maintenance. In contrast, the conductivity sensor is widely used in practical engineering due to its low price, simple structure, fast response and strong anti-interference.

[0003] However, the single indicator of conductivity has the following limitations: 1. Insufficient specificity: conductivity cannot directly distinguish the type of pollutants (such as industrial wastewater and initial rain flushing), and needs to rely on pattern recognition algorithms for supplementation. 2. Temperature drift and electrode aging: conductivity is significantly affected by water temperature, and long-term operation may also cause baseline drift due to electrode contamination and aging. 3. High false alarm rate of anomaly detection: traditional threshold-based detection methods (such as setting upper and lower limits) cannot adapt to complex dynamic environments, especially in high-noise and frequently changing seasonal scenarios.

[0004] In addition, traditional anomaly detection algorithms (such as static threshold and mean-variance-based detection) have poor adaptability to conductivity time series data, with a high false alarm rate in high-noise environments, and lack comprehensive analysis capabilities in multiple scales and multiple modalities.

[0005] Therefore, the single conductivity detection also has the following technical bottlenecks: 1. Static threshold limitations: fixed upper and lower limits cannot reflect day-night / seasonal dynamic characteristics. 2. Lack of multi-scale analysis: existing methods rarely consider micro (second-level changes), meso (minute-level fluctuations), and macro (day-level trends) simultaneously. 3. No feature fusion mechanism: time domain, frequency domain, and spatial information are not used cooperatively.

[0006] Based on the above status quo, the present application proposes a new solution that solves the above problems through multi-modal feature fusion and adaptive density clustering. SUMMARY

[0007] In view of the deficiencies in the prior art, the present application provides a conductivity pollution intelligent diagnosis method based on multi-modal feature fusion and adaptive density clustering, which is suitable for smart water, watershed water quality monitoring, industrial emission supervision and other application scenarios.

[0008] The present application is solved by the following technical solutions.

[0009] The intelligent diagnosis method for conductivity pollution based on multi-modal feature fusion and adaptive density clustering is as follows: (1) data processing is performed through an intelligent data preprocessing module, including the following steps: S10: temperature compensation: obtaining original conductivity data Craw, converting the original conductivity data Craw into compensated conductivity C25 through a temperature compensation algorithm; S20: drift correction: analyzing the conductivity time series based on an ARIMA model, comparing the actual value and the predicted value residual, and identifying the correction sensor drift.

[0010] (2) conductivity data is detected through a multi-scale anomaly detection engine, including the following steps: S30: real-time tracking of the conductivity change rate is performed through microscopic scale detection, and the change rate exceeding the threshold is regarded as an anomaly; S40: anomaly detection is performed through mesoscopic scale detection; S50: macroscopic scale detection is performed to compare the cumulative deviation with the threshold to determine the anomaly.

[0011] (3) system extraction is performed through a multi-modal feature extraction system, including the following steps: S60: at least one of time domain features, frequency domain features, and spatiotemporal features is used for system extraction.

[0012] (4) data optimization is performed through an adaptive density clustering algorithm, including the following steps: S70: at least one of a parameter adaptive mechanism, a hybrid distance metric, and hierarchical clustering optimization is used for data optimization.

[0013] (5) diagnosis is performed through an intelligent diagnosis decision system, including the following steps: S80: diagnosis is performed through a multi-expert fusion model and dynamic confidence evaluation.

[0014] In a preferred embodiment, in step S10, the temperature compensation algorithm uses a third-order polynomial fitting to correct the temperature drift, and the mathematical model is:

[0015] C 25 =C raw ·[1+a(T-T ref )+b(T-T ref ) 2 +c(T-T ref ) 3 ];

[0016] wherein: C 25 is the conductivity compensated to 25℃; C raw is the original conductivity data; T is the current water temperature; T ref is the reference temperature 25℃; a, b, and c are coefficients obtained by least squares fitting a large amount of historical data.

[0017] In a preferred embodiment, the parameters a, b, and c in step S10 are obtained through the following algorithm: S11: collecting historical data T and Craw For the data, perform multinomial regression on the data to minimize the sum of squared residuals, as shown in the following formula:

[0018] , thus obtaining parameters a, b, c.

[0019] In a preferred embodiment, the drift correction in step S20 is achieved by the following method: S21: Analyze the conductivity time series based on the ARIMA model, compare the residual between the actual value and the predicted value, identify and correct sensor drift, and ensure stable and accurate data.

[0020] In a preferred embodiment, in step S30, the microscale detection is as follows: an improved Kalman filter is used to track the rate of change of conductivity in real time, and a dynamic threshold is set based on the historical standard deviation. A rate of change exceeding the threshold is considered abnormal. In step S40, the mesoscale detection is as follows: the Robust Z-score is calculated through an adaptive sliding window, outliers are removed using a Hampel identifier, and anomalies within the window are detected. In step S50, the macroscale detection is as follows: a 7-day periodic dynamic baseline is constructed, and a nonparametric CUSUM algorithm is used to compare the cumulative deviation with the threshold to determine anomalies.

[0021] In a preferred embodiment, in step S60: the time-domain feature is the manifestation of the pollution event in the conductivity time-series data; the frequency-domain feature includes EEMD decomposition, intrinsic mode function entropy value, and energy ratio; the spatiotemporal feature includes network topology modeling and spatial feature extraction.

[0022] In a preferred embodiment, in step S70, the parameter adaptation mechanism is as follows: parameters are determined based on the data density distribution analysis using the OPTICS algorithm, and the initial centroid selection is improved using DBSCAN++ to enhance clustering accuracy; the hybrid distance metric is as follows: the temporal DTW distance and spatial GraphEMD distance are combined and weighted according to the coefficients optimized by the gradient boosting tree to accurately measure data similarity; the hierarchical clustering is as follows: initial clustering in the original feature space, fine clustering after UMAP dimensionality reduction, and then merging similar clusters according to Wasserstein distance to optimize the clustering results.

[0023] In a preferred embodiment, in step S80, the multi-expert fusion model is: employing a multi-model collaborative expert fusion strategy, designing dedicated sub-models for different feature types, and improving overall performance through weighted integration; the dynamic confidence level is: introducing a three-level dynamic confidence interval mechanism.

[0024] In a preferred embodiment, in step S80, the processing method of the multi-expert fusion model includes structured feature processing, temporal feature processing, spatial feature processing, and fusion strategy.

[0025] In a preferred embodiment, in step S80, the structured feature processing is: classifying structured environmental parameters using random forest; the temporal feature processing is: mining local patterns in the conductivity time series using a one-dimensional convolutional neural network; the spatial feature processing is: processing the spatial topology relationships and node conductivity data of the sensor network using a graph attention network; and the fusion strategy is: fusing the outputs of the three sub-models—structured feature processing, temporal feature processing, and spatial feature processing—through weighted voting.

[0026] Compared with the prior art, the present invention has the following beneficial effects: (1) Improved detection performance: event recognition rate > 92%, false alarm rate < 8%, which is a significant improvement over traditional methods. (2) Economic advantages: equipment cost is reduced by 90%, maintenance cycle is extended to 6 months, saving costs and manpower. (3) Adaptable innovation: supports operation from -20℃ to 60℃, tolerates electrode aging deviation < 15%, and adapts to various environments. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the various module layers in one embodiment of this application.

[0028] Figure 2 This is a flowchart illustrating one embodiment of this application. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to specific embodiments. The intelligent diagnostic method for conductivity pollution based on multimodal feature fusion and adaptive density clustering in the present invention is as follows.

[0030] 1. Data processing is performed through an intelligent data preprocessing module, specifically including the following methods and algorithms.

[0031] (1) Temperature compensation algorithm:

[0032] Electrical conductivity exhibits a non-linear relationship with temperature, and traditional linear compensation is insufficient to guarantee accuracy under large temperature differences. This invention employs a third-order polynomial fitting to correct temperature drift; the mathematical model is as follows:

[0033] C 25 =C raw ·[1+a(TT ref )+b(TT ref ) 2 +c(TT ref ) 3 ];

[0034] Where: C 25 To compensate for conductivity at 25℃; C rawThis is the raw conductivity data; T is the current water temperature; T ref The reference temperature is 25℃; a, b, and c are coefficients obtained by fitting a large amount of historical data using least squares.

[0035] The algorithm steps of this temperature compensation algorithm are as follows: First, collect historical T and C values. raw The dataset is then processed using multinomial regression to obtain a, b, and c, which are then applied online for real-time compensation. The formula for minimizing the sum of squared residuals is as follows.

[0036]

[0037] (2) Drift correction model:

[0038] Conductivity time series analysis based on the ARIMA model compares the residuals between actual and predicted values ​​to identify and correct sensor drift, ensuring stable and accurate data. The residual calculation formula is as follows:

[0039]

[0040] The correction formula is as follows:

[0041]

[0042] in It is the moving average of the most recent NNN residuals.

[0043] The algorithm steps of this drift correction model are as follows:

[0044] First, establish the ARIMA model:

[0045]

[0046] Recalculate the predicted value , find the residual .

[0047] Finally, if It is suspected that drift exists, and a correction is performed.

[0048] 2. The multi-scale anomaly detection engine in this application includes the following schemes.

[0049] (1) Microscale detection:

[0050] An improved Kalman filter tracks the rate of change in conductivity in real time, setting a dynamic threshold based on historical standard deviations; a rate of change exceeding the threshold is considered abnormal. The formula for calculating the rate of change in conductivity is as follows.

[0051] .

[0052] The formula for the dynamic threshold is as follows, where k=3 (empirical value).

[0053] .

[0054] (2) Mesoscale detection:

[0055] Robust Z-score is calculated using an adaptive sliding window, and outliers are removed using the Hampel identifier to detect anomalies within the window. The algorithm formula for Robust Z-score is as follows.

[0056] .

[0057] The Hampel identifier algorithm formula is as follows.

[0058] .

[0059] (3) Macro-scale detection:

[0060] A 7-day dynamic baseline is constructed, and a non-parametric CUSUM algorithm is used to compare the cumulative deviation with a threshold to identify anomalies. The formula for the cumulative deviation is as follows, where... is the historical dynamic baseline, and h is the threshold. .

[0061] 3. The multimodal feature extraction system in this application includes the following feature extraction methods.

[0062] (1) Temporal characteristics: Pollution events (such as initial rain wash-off and industrial wastewater discharge) are usually manifested in conductivity time series data as: sudden change: conductivity rises or falls sharply in a short period of time; gradual change: conductivity increases or decreases slowly; periodic disturbance: periodic fluctuations caused by intermittent industrial emissions.

[0063] Temporal feature extraction is performed using the following feature extraction formula.

[0064] Ascent time: t rise =t 90% -t 10%

[0065] t 90% , t 10% The time taken for the conductivity signal to travel from baseline to 10% and 90% of its peak value.

[0066] Peak duration: .

[0067] Attenuation rate: .

[0068] Skewness: .

[0069] Kurtosis: .

[0070] The specific steps for time-domain feature extraction are as follows: First, perform sliding window processing on the conductivity signal C(t), with the window length determined based on the sampling frequency (e.g., 5 minutes). Then, calculate the statistical characteristics within each window: , , ,S,K).

[0071] Then output the feature vector: .

[0072] (2) Frequency domain characteristics: including EEMD decomposition, intrinsic mode function (IMF) entropy value, and energy ratio.

[0073] Specifically, EEMD decomposes the signal C(t) into several intrinsic mode functions (IMFs), as shown in the following formula.

[0074] ;in Let r(t) be the i-th IMF and r(t) be the residual.

[0075] Specifically, the frequency domain feature formula includes sample entropy and energy ratio, where sample entropy measures the complexity of the IMF, as shown in the following formula.

[0076] ;

[0077] m: Embedding dimension; r: Tolerance threshold (usually 0.2σ); A, B: Number of matching templates.

[0078] The Energy Ratio is calculated as the proportion of energy from each IMF to the total energy, using the following formula:

[0079] .

[0080] The specific formula for the feature extraction step is as follows: First, perform EEMD on C(t) to obtain... Then for each calculate( , Finally, the frequency domain feature vector is output, as shown in the following formula:

[0081] .

[0082] (3) Spatiotemporal characteristics: First, network topology modeling is performed, with sensor nodes as nodes V in graph G=(V,E), and edges E representing hydraulic connections. Edge weights Distance-weighted + traffic-weighted, the formula is as follows. .

[0083] Then, spatial feature extraction is performed:

[0084] First, perform Dynamic Time Warped Distance (DTW) to quantify the similarity of the conductivity time series of nodes i and j, as shown in the following formula.

[0085] ;

[0086] Next, we perform node centrality and betweenness centrality calculations, using the following formulas: : The number of shortest paths between nodes s and t : The number of shortest paths passing through v.

[0087] .

[0088] 4. The adaptive density clustering algorithm in this application includes the following schemes.

[0089] (1) Parameter adaptive mechanism: Based on the analysis of data density distribution by OPTICS algorithm, parameters are determined, and DBSCAN++ is used to improve the selection of initial centroids and improve clustering accuracy.

[0090] Specifically, the local density is estimated using KDE, as shown in the following formula, where K is the kernel function (Gaussian kernel) and h is the bandwidth, which is automatically optimized.

[0091] .

[0092] Then, DBSCAN ϵ is adaptively adjusted based on the local density, as shown in the following formula.

[0093] .

[0094] (2) Hybrid distance metric: Combining temporal DTW distance and spatial GraphEMD distance, weighted and fused according to the coefficients of gradient boosting tree optimization, to accurately measure data similarity, the formula is as follows.

[0095] .

[0096] (3) Hierarchical clustering optimization: initial clustering in the original feature space, fine clustering after UMAP dimensionality reduction, and then merging similar clusters according to Wasserstein distance to optimize the clustering results.

[0097] Specifically, after initial clustering, UMAP is used for dimensionality reduction, as shown in the following formula.

[0098] .

[0099] Specifically, similar clusters are then merged based on Wasserstein distance, as shown in the following formula.

[0100] .

[0101] 5. The intelligent diagnostic decision-making system in this application includes the following solutions.

[0102] (1) Multi-expert fusion model: In order to improve the accuracy of conductivity pollution event diagnosis, this system adopts a multi-model collaborative expert fusion strategy, designs special sub-models for different feature types, and improves the overall performance through weighted integration.

[0103] In this process, structured feature processing uses Random Forest to classify structured environmental parameters (such as temperature, flow rate, historical mean, etc.). Random Forest determines the outcome through voting among multiple decision trees, as shown in the formula below, where T is the number of trees. This is the output of the t-th tree.

[0104] .

[0105] In this process, a one-dimensional convolutional neural network (1D-CNN) is used for temporal feature processing to mine local patterns in the conductivity time series. The core convolution operation formula is as follows: : Convolution kernel weights; k: Convolution kernel length; b: Bias term.

[0106] .

[0107] Spatial feature processing employs a Graph Attention Network (GAT) to handle the spatial topology and node conductivity data of the sensor network. This is achieved through attention weights. The information aggregation of nodes i and j is calculated using the following formula, where... : Attention coefficients of nodes i and j; W: Weight matrix; σ: Activation function.

[0108] .

[0109] The fusion strategy combines the outputs of each sub-model through weighted voting, as shown in the following formula. The model predicts a certain type of probability; The weights of each model are dynamically optimized during training.

[0110]

[0111] (2) Dynamic Confidence Assessment: To ensure diagnostic reliability, a three-level dynamic confidence interval mechanism is introduced. This includes the following steps: First, calculate the confidence level Conf(y) of the prediction result, defined as the weighted average of the output probabilities of multiple models, as shown in the following formula:

[0112]

[0113] The system then classifies the levels based on confidence level: High confidence level (Conf≥0.8): Automatic confirmation; Medium confidence level (0.5≤Conf<0.8): Marked as requiring attention by the system; Low confidence level (Conf<0.5): Manual review is triggered.

[0114] Finally, human review and feedback are used as training data to periodically update the model parameters (online learning).

[0115] In the specific implementation of this application, the hardware deployment is as follows: the conductivity sensor network is networked using the Modbus RTU protocol, and the node spacing is adaptively configured according to the watershed area, which facilitates efficient data transmission and comprehensive monitoring.

[0116] Software Architecture: Edge Computing Layer: Real-time anomaly detection with a latency of <1 second, enabling rapid response to contamination events. Cloud Analytics Layer: Completes feature extraction and clustering every 5 minutes, providing a basis for diagnosis. Application Layer: Provides visualized alarms and diagnostic reports for easy user viewing of results.

[0117] See Figure 1 and Figure 2 , Figure 1 As shown, the perception layer consists of sensors installed in the pipeline to collect data. The collected data stream enters the edge computing layer, where it undergoes real-time processing through an intelligent data preprocessing module and a multi-scale anomaly detection engine to obtain preprocessed data and signals. These signals are then transmitted to the cloud analysis layer for intelligent analysis and processing. Additionally, the application layer is used for visual interactive operations, alarm operations, and diagnostic report output.

[0118] Figure 2 The diagram shown is a flowchart of one embodiment of this application, which corresponds to the specific method steps in this application.

[0119] The multi-stage joint analysis framework in this application is described in detail below.

[0120] I. Dynamic Anomaly Detection:

[0121] (1) Sliding quantile calculation: The quantiles of conductivity data are dynamically calculated based on a time window. For example, a 24-hour window is commonly used to calculate the 25% (Q1) and 75% (Q3) quantiles and the IQR (interquartile range). Kernel density estimation is used to smooth the quantile calculation, thereby solving the problem of quantile jitter that may occur with sparse data, making the calculation results more stable and reliable. At the same time, the boundary of anomaly judgment is dynamically adjusted based on the permutation entropy of the data (an indicator that measures the degree of disorder of the data). When the permutation entropy of the data exceeds the preset threshold, the boundary of anomaly judgment is automatically widened to avoid misjudging the normal fluctuation of the data as an anomaly, thus preparing for the subsequent accurate detection of anomalies from the basic level.

[0122] (2) Adaptive threshold adjustment: Its core lies in monitoring the volatility of the data in real time, measuring this volatility through various statistical methods and indicators, and then dynamically changing the threshold range for anomaly detection based on the changes in volatility. For example, during periods when conductivity data fluctuates drastically, the threshold is appropriately relaxed, while during periods when the data is relatively stable, the threshold is tightened, thereby flexibly adapting to the dynamic changes in the data, minimizing false judgments, and ensuring the accuracy of anomaly detection.

[0123] (3) Multi-scale verification: The anomalies detected in the preliminary tests are verified in a comprehensive and detailed manner at three different scales: micro, meso, and macro. At the micro level, the consistency of gradient changes in the three sampling points before and after the anomaly is checked, that is, whether the trend of conductivity change near the anomaly is consistent with the change pattern caused by normal pollution events. If it is just an anomaly caused by accidental local fluctuations, it can be eliminated by this method. At the meso level, the statistical characteristic deviation within the sliding window (generally set to about 5 minutes) where the anomaly is located is verified. By observing whether the mean, variance and other statistical indicators of the data within the window show obvious deviations from the normal range, the authenticity of the anomaly is further confirmed. At the macro level, the percentile interval of the historical data of the same period of the day is compared. The data detected at the current moment is compared with the distribution of historical data of the same period in the past. If the current data exceeds the normal fluctuation range of the same period in the past, then the anomaly is more likely to represent a real pollution event or other abnormal situation. Through such progressive and multi-dimensional verification, the accuracy and reliability of anomaly detection are greatly improved.

[0124] II. Feature Extraction:

[0125] (1) Temporal characteristics: On the one hand, the morphological characteristics of the conductivity signal waveform are extracted, such as rise time (the time from the start of the waveform to the peak value of 10%-90%), peak duration (the duration of the waveform in the peak state), and decay rate (the rate at which the conductivity decreases after the peak). These morphological characteristics intuitively reflect the dynamic process of conductivity changing over time when a pollution event occurs. Different types of pollution events often lead to different morphological characteristics of the conductivity waveform. On the other hand, statistical characteristics within the sliding window are calculated, such as skewness (measuring the degree of asymmetry in the data distribution) and kurtosis (describing the peak degree of the data distribution). These statistical indicators further characterize the conductivity data from the perspective of data distribution, and help to judge the specific type and degree of pollution event. The combination of the two allows for a comprehensive mining of pollution-related information contained in the conductivity data from the temporal perspective.

[0126] (2) Frequency Domain Characteristics: First, a suitable transformation method, such as wavelet transform, is used to convert the conductivity signal from the time domain to the frequency domain, decomposing the complex time-domain signal into components of different frequencies. Based on this, the energy distribution of each frequency band is calculated, that is, the proportion of energy of different frequency components in the whole signal, as well as the Shannon entropy (used to measure the complexity and uncertainty of the signal in the frequency domain). Different pollution events usually exhibit different energy distribution and entropy characteristics in the frequency domain. By analyzing these frequency domain characteristics, different types of pollution events can be effectively distinguished. For example, some periodic pollution may have obvious energy concentration in a specific frequency band, while sudden pollution may show different frequency domain energy distribution patterns.

[0127] (3) Spatial correlation characteristics: First, a sensor network topology map within the watershed is constructed, with each sensor location considered as a node. Edges are constructed based on the actual physical connections between sensors (such as water flow direction, distance, etc.) to form a graph structure that reflects the spatial structure of the sensor network. Then, the dynamic time warping (DTW) distance between the target monitoring point and adjacent nodes is calculated. The DTW distance can measure the similarity of two time series under conditions of stretching, deformation, etc. on the time axis. By analyzing the DTW distance of conductivity time series between different nodes, the spatial diffusion trend of pollution can be inferred. For example, nodes that are close to each other and have a small DTW distance are likely to be in the same pollution diffusion area. Furthermore, a pollution propagation direction index can be established based on the Granger causality test. By determining whether there is a causal relationship between the changes in conductivity of different nodes, the direction of pollution propagation within the watershed can be determined. Additionally, the betweenness centrality of nodes in complex network theory can be used to quantify the spatial influence weight of each sensor node on pollution propagation. The higher the betweenness centrality of a node, the greater its influence on the pollution propagation process. By combining these spatial correlation characteristics, the propagation and distribution of pollution in the spatial dimension can be analyzed more comprehensively, providing a strong basis for accurate judgment of pollution events in the future.

[0128] III. Improved density clustering:

[0129] (1) Adaptive density parameters: Abandoning the traditional fixed density parameter setting method, the kernel density estimation (KDE) method is adopted to automatically determine the density parameters required for clustering based on the actual distribution of the data. Kernel density estimation is to estimate the density of the local area around the data points and dynamically adjust the clustering density requirements according to the sparsity of the data distribution in the feature space. For example, in areas with relatively sparse data, the density threshold is appropriately reduced so that data points that might be misclassified into different classes due to excessively high density standards can be grouped into one class; while in areas with dense data, a relatively high density standard is maintained to ensure the accuracy and rationality of clustering. Through this adaptive approach, the clustering process can better fit the inherent structure and distribution characteristics of the actual data and improve the clustering effect.

[0130] (2) Composite distance metric: For features of different dimensions, appropriate distance metric methods are adopted respectively, and then they are weighted and fused to form a comprehensive distance metric index. In terms of time domain features, a subsequence matching variant of dynamic time warping distance (DTW) is adopted. This variant can more flexibly handle complex issues such as subsequence similarity and time-axis scaling in conductivity time series data, and more accurately measure the similarity between different data at the time domain feature level. For frequency domain features, the Bhattacharyya distance between wavelet coefficients is used. It can well measure the similarity between two probability distributions (corresponding to the frequency domain feature distribution reflected by wavelet coefficients). By comparing the differences in wavelet coefficient distribution of different data in the frequency domain, the similarity or difference between them in frequency domain features can be accurately judged. In terms of spatial features, graph editing distance based on network topology is adopted. By calculating the minimum editing operations (such as node deletion, addition, edge modification, etc.) required to convert one graph structure (corresponding to the constructed watershed sensor network topology graph) into another graph structure, the similarity in spatial structure is reflected, thereby accurately measuring the distance between different data in terms of spatial features. Finally, these three different distance measures are weighted and combined according to the weights determined by feature importance analysis, so that the composite distance measure can comprehensively consider multiple features, more comprehensively and accurately reflect the overall similarity between data, and thus improve the quality and accuracy of clustering.

[0131] (3) Hierarchical Clustering Optimization: The entire clustering process is optimized in three stages. First, preliminary density clustering is performed in the original feature space. Based on the previously determined adaptive density parameters and composite distance metric, the data points in the high-dimensional space formed by the extracted multimodal features are initially divided, marking those points in the core clustering region and initially determining the possible cluster categories, providing a basic framework for further refinement of clustering. Next, for the points in the initial clustering that are at the boundary and whose classification is not clear (i.e., boundary points), secondary clustering is performed after local linear embedding projection. The local linear embedding method can project the boundary points in the high-dimensional feature space to the low-dimensional space, reducing the dimensionality while preserving the local structural features of the data. Then, in the low-dimensional space, these boundary points are clustered again based on more suitable distance metrics and other methods to further accurately determine their classification categories, refine the clustering results, and reduce the inaccuracy of classification caused by boundary ambiguity. Finally, hierarchical merging is performed based on the inter-cluster similarity matrix (constructed by calculating the Wasserstein distance). The Wasserstein distance can effectively measure the differences between different probability distributions (corresponding to the data distributions of different clusters). By calculating the Wasserstein distance between different clusters, their similarity is determined, and clusters with high similarity and similar features are merged to optimize the hierarchical structure of clustering as a whole, resulting in a reasonable and accurate clustering partition, which facilitates effective event classification in the future.

[0132] IV. Event Classification:

[0133] (1) Physical Rule Engine: A tree-like discriminant tree is constructed, with different judgment conditions set at each node. These conditions are mainly based on the characteristic differences of conductivity data in the time domain (such as rising slope, duration, etc.) and spatial correlation (such as correlation with surrounding nodes, etc.) to set corresponding threshold rules. Starting from the root node of the discriminant tree, the data to be classified is compared with the threshold conditions of each node according to its characteristics, and the judgment is made step by step along the branches of the tree. Finally, the data is classified into the leaf node of the corresponding specific event type, realizing the classification and differentiation of different events such as initial rain erosion, industrial pollution, and sensor failure. In addition, in order to avoid misclassification due to the data being in the boundary ambiguity area during the judgment process, a confidence decay factor is also set at each node. When the data is close to the judgment boundary of the node, the confidence level of the judgment is appropriately reduced by the confidence decay factor, making the whole classification process more robust and reliable, and improving the accuracy of classification.

[0134] (2) Dynamic weight allocation: The importance weight of each feature in the classification process is dynamically adjusted according to different external factors. For example, based on seasonal factors, during the rainy season, due to the scouring and dilution of water bodies by rainfall, the morphological characteristics of pollution events in terms of conductivity changes will be more prominent. Therefore, the weight of morphological features will be strengthened accordingly, and more emphasis will be placed on judging the event type based on the morphological characteristics of the waveform. In the dry season, the water body is relatively stable, and the periodic and long-term changes reflected by some frequency domain features are more helpful in distinguishing different pollution events. Therefore, the weight allocation of frequency domain features will be emphasized. For another example, the weight of spatial features is non-linearly positively correlated with rainfall intensity. The greater the rainfall intensity, the more fluid the water body becomes, and the more complex and significant the spatial propagation and diffusion of pollution will be. At this time, it is necessary to increase the weight of spatial features in the classification and pay more attention to using the spatial correlation information between sensor networks to analyze the type, scope and propagation direction of pollution events. In addition, the aging factor of the equipment also affects the weighting of the sensor reliability index. As the equipment is used for a long time, the sensor will age, and its measurement reliability will be affected. At this time, it is necessary to reasonably adjust the weight of sensor-related features in the classification according to the aging factor of the equipment, appropriately reduce the weight of features that are more affected by aging, and increase the weight of other relatively stable and reliable features to ensure the quality of the classification results and enable the classification to maintain good accuracy under different aging levels of the equipment.

[0135] (3) Feedback Optimization Mechanism: An online learning framework is established. When classification errors occur, this framework is used to conduct detailed feature backtracking analysis of the misclassification event, tracing back to see which features behaved abnormally or which link had a problem in each stage of feature extraction, clustering, and classification, thereby finding the cause of the misclassification. On this basis, the cluster center vector is dynamically updated, and an exponentially weighted moving average strategy is adopted to continuously adjust the position of the cluster center according to the new data, so that the cluster center can keep up with the changes in data distribution in a timely manner, ensuring the effectiveness and accuracy of clustering. Meanwhile, by using a reinforcement learning framework to optimize the threshold parameter set, the classification results are used as reward signals (e.g., positive rewards for correct classification and negative rewards for misclassification). The agent (analogous to the module responsible for parameter adjustment in the system) automatically adjusts the threshold in the discriminant tree, the density parameter in the cluster, and other key parameters based on these reward feedbacks. This enables the entire system to continuously learn from past classification experience and continuously optimize its parameter configuration, thereby improving long-term adaptability and classification performance. This allows the system to better cope with complex and ever-changing real-world water pollution monitoring scenarios and ensures accurate classification and identification of conductivity pollution events under different environmental conditions and pollution situations.

[0136] As can be seen from the above description, the technical solutions in this application (1) adopt a cascaded architecture for multi-scale anomaly detection: multi-scale collaboration improves the accuracy and robustness of anomaly detection. (2) adopt a density clustering method with hybrid distance metrics: integrates multiple distance metrics and adapts to achieve accurate clustering. (3) adopt a feature engineering system with enhanced conductivity specificity: multi-dimensional feature mining enhances the ability to distinguish contamination. (4) an adaptive correction algorithm for sensor degradation: corrects sensor problems and ensures data quality and system stability.

[0137] The technical solution in this application has the following core advantages: (1) Low-cost alternative: a single conductivity sensor combined with intelligent algorithms can replace multi-parameter detection at a low cost. (2) Real-time advantage: conductivity data has a fast response and can capture sudden pollution events. (3) Technical feasibility: pollution events are accompanied by changes in conductivity characteristics, and machine learning can be used to mine patterns. (4) Advantages of conductivity detection: low price, strong anti-interference, and can comprehensively reflect the pollution status of water bodies. (5) Limitations of conductivity detection: it cannot distinguish the type of pollutants, there is baseline drift, and it is affected by temperature. There are corresponding solutions to these problems.

[0138] Compared to conventional solutions in existing technologies, the technical solution in this application offers improved detection performance: event recognition rate >92%, false alarm rate <8%, representing a significant performance improvement over traditional methods. Economic advantages include: equipment costs reduced by 90%, maintenance cycle extended to 6 months, saving costs and manpower. Adaptability innovation: supports operation from -20℃ to 60℃, tolerates electrode aging deviation <15%, and adapts to various environments.

[0139] The scope of protection of this invention includes, but is not limited to, the above embodiments. The scope of protection of this invention is defined by the claims. Any substitutions, modifications, or improvements to this technology that are easily conceived by those skilled in the art fall within the scope of protection of this invention.

Claims

1. A smart diagnostic method for conductivity pollution based on multimodal feature fusion and adaptive density clustering, characterized in that, The method is as follows: (1) Data processing is performed through the intelligent data preprocessing module, including the following steps: S10: Temperature Compensation: Acquire raw conductivity data C raw The original conductivity data C is obtained through a temperature compensation algorithm. raw Converted to compensated conductivity C 25 ; S20: Drift Correction: Based on the ARIMA model, analyze the conductivity time series, compare the residuals of actual and predicted values, and identify and correct sensor drift. (2) Conductivity data is detected using a multi-scale anomaly detection engine, including the following steps: S30: Real-time tracking of conductivity change rate through microscale detection; a change rate exceeding a threshold is considered abnormal. Microscale detection involves: using an improved Kalman filter to track conductivity change rate in real time, setting a dynamic threshold based on historical standard deviation, and considering a change rate exceeding the threshold as abnormal. S40: Anomalies are detected through mesoscale detection; mesoscale detection involves calculating the Robust Z-score using an adaptive sliding window, removing outliers using a Hampel identifier, and detecting anomalies within the window. S50: Anomalies are identified by comparing cumulative deviation with a threshold through macro-scale detection; macro-scale detection involves constructing a 7-day dynamic baseline and using the non-parametric CUSUM algorithm to compare cumulative deviation with a threshold to identify anomalies. (3) System extraction is performed using a multimodal feature extraction system, including the following steps: S60: System extraction is performed using at least one of time-domain features, frequency-domain features, and spatiotemporal features; (4) Optimize the data using an adaptive density clustering algorithm, including the following steps: S70: Data optimization is performed through at least one of parameter adaptive mechanism, hybrid distance metric, and hierarchical clustering optimization. (5) Diagnosis is performed through an intelligent diagnostic decision-making system, including the following steps: S80: Diagnosis is performed using a multi-expert fusion model and dynamic confidence assessment.

2. The intelligent diagnostic method for conductivity contamination based on multimodal feature fusion and adaptive density clustering according to claim 1, characterized in that, In step S10, the temperature compensation algorithm uses a third-order polynomial fitting to correct the temperature drift. The mathematical model is as follows: C 25 =C raw ·[1+a(T-T ref )+b(T-T ref ) 2 +c(T-T ref ) 3 ]; in: C 25 To compensate for the conductivity at 25℃; C raw This is the raw conductivity data; T represents the current water temperature; T ref The reference temperature is 25℃. a, b, and c are coefficients obtained by fitting a large amount of historical data using least squares.

3. The intelligent diagnostic method for conductivity contamination based on multimodal feature fusion and adaptive density clustering according to claim 2, characterized in that, The parameters a, b, and c in step S10 are obtained using the following algorithm: S11: Collect historical data T and C raw For the data, perform multinomial regression on the data to minimize the sum of squared residuals, as shown in the following formula: ， The parameters a, b, and c are obtained.

4. The intelligent diagnostic method for conductivity contamination based on multimodal feature fusion and adaptive density clustering according to claim 3, characterized in that, The drift correction in step S20 is achieved through the following method: S21: Analyze conductivity time series based on ARIMA model, compare actual and predicted residuals, identify and correct sensor drift, and ensure stable and accurate data.

5. The intelligent diagnostic method for conductivity contamination based on multimodal feature fusion and adaptive density clustering according to claim 4, characterized in that, In step S60: the time-domain features are the manifestation of pollution events in conductivity time-series data; the frequency-domain features include EEMD decomposition, intrinsic mode function entropy value, and energy ratio; the spatiotemporal features include network topology modeling and spatial feature extraction.

6. The intelligent diagnostic method for conductivity contamination based on multimodal feature fusion and adaptive density clustering according to claim 5, characterized in that, In step S70, The parameter adaptation mechanism is as follows: parameters are determined based on the data density distribution analysis of the OPTICS algorithm, and the initial centroid selection is improved using DBSCAN++ to enhance clustering accuracy. The hybrid distance metric is a weighted fusion of temporal DTW distance and spatial GraphEMD distance, optimized by gradient boosting tree coefficients, to accurately measure data similarity. The hierarchical clustering is as follows: initial clustering in the original feature space, fine clustering after UMAP dimensionality reduction, and then merging similar clusters according to Wasserstein distance to optimize the clustering results.

7. The intelligent diagnostic method for conductivity contamination based on multimodal feature fusion and adaptive density clustering according to claim 6, characterized in that, In step S80, The multi-expert fusion model is as follows: it adopts a multi-model collaborative expert fusion strategy, designs special sub-models for different feature types, and improves the overall performance through weighted integration; The dynamic confidence level is achieved by introducing a three-level dynamic confidence interval mechanism.

8. The intelligent diagnostic method for conductivity contamination based on multimodal feature fusion and adaptive density clustering according to claim 7, characterized in that, In step S80, the processing methods of the multi-expert fusion model include structured feature processing, temporal feature processing, spatial feature processing, and fusion strategy.

9. The intelligent diagnostic method for conductivity contamination based on multimodal feature fusion and adaptive density clustering according to claim 8, characterized in that, In step S80, The structured feature processing involves classifying structured environmental parameters using random forests. The time-series feature processing involves using a one-dimensional convolutional neural network to mine local patterns in the conductivity time series. The spatial feature processing involves using a graph attention network to process the spatial topology and node conductivity data of the sensor network. The fusion strategy is to fuse the outputs of the three sub-models—structured feature processing, temporal feature processing, and spatial feature processing—through weighted voting.

Images