Percolate treatment process anomaly detection method and system based on data fusion
By constructing operating condition vectors and feature importance assessment, combined with the isolated forest algorithm, the accuracy problem of anomaly detection in leachate treatment process was solved, enabling accurate differentiation between operating condition changes and equipment failures, and improving detection accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU XINQI ENVIRONMENTAL PROTECTION CO LTD
- Filing Date
- 2026-03-24
- Publication Date
- 2026-04-21
AI Technical Summary
In the existing technology, the abnormal detection methods of reverse osmosis units in the leachate treatment process are difficult to effectively distinguish between normal parameter drift driven by operating conditions and abnormal deviations induced by faults, resulting in low accuracy of detection results.
By collecting and processing influent flow rate, chemical oxygen demand, and performance parameter sequences, operating condition vectors are constructed for clustering. The importance of features is evaluated by combining sensitivity and independence indices, and anomaly detection is performed using isolated forests to distinguish between normal fluctuations and faults.
It enables high-precision anomaly detection in the leachate treatment process, reduces the probability of normal fluctuations being misjudged as faults, and improves the accuracy of detection results.
Smart Images
Figure CN121891940A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing. In particular, it relates to a method and system for detecting anomalies in leachate treatment processes based on data fusion. Background Technology
[0002] Leachate is a highly polluting wastewater generated during the landfill process. It has a complex composition and high concentration, containing large amounts of organic and inorganic matter, as well as various toxic and harmful substances. With the increasing number of landfills, the amount of leachate generated is increasing year by year. How to effectively and quickly treat leachate has become an important issue in the environmental protection field.
[0003] In the process of landfill leachate treatment, the reverse osmosis (RO) unit, as the core deep treatment link, is affected by multiple factors such as the effect of front-end pretreatment, fluctuations in influent water quality, and operation scheduling strategies. It exhibits significant time-varying and drastic fluctuation characteristics, which leads to frequent switching of system operating conditions between high and low loads and different pollution load ranges. Under this dynamic operating environment, the key performance parameters of the RO membrane will fluctuate normally in accordance with physical laws as the influent conditions and operating conditions change. This fluctuation is essentially an adaptive response of the system to operating condition disturbances rather than a fault manifestation.
[0004] However, existing traditional anomaly detection methods mostly rely on preset fixed threshold criteria or single statistical models based on steady-state assumptions. They lack the ability to model the dynamic evolution of operating conditions and the coupling relationship between parameters, making it difficult to effectively distinguish between normal parameter drift driven by operating conditions and real abnormal deviations induced by faults. They are very likely to misjudge operating condition fluctuations as equipment failures (such as membrane fouling), resulting in low accuracy of anomaly detection results for leachate treatment. Summary of the Invention
[0005] To address the aforementioned technical problems, the present invention provides solutions in the following aspects.
[0006] In a first aspect, the present invention provides a method for detecting anomalies in leachate treatment processes based on data fusion, comprising: collecting a set of influent flow rate sequences, influent chemical oxygen demand (COD) sequences, and performance parameter sequences for any process in history, wherein the performance parameter sequence set includes a pre-membrane pressure sequence, a post-membrane pressure sequence, a permeate flow rate sequence, and a permeate conductivity sequence; calculating a first co-current shock load index for the influent flow rate sequence, and similarly obtaining a second co-current shock load index for the influent COD sequence; constructing a condition vector by combining the first and second co-current shock load indices; clustering all processes in history based on the condition vector to obtain several condition clusters; taking any condition cluster as a target cluster, and taking any process in the target cluster as a target process, calculating the target... The mean and variance of each sequence in the performance parameter sequence set of the target process are used to construct the performance feature vector of the target process. Any element in the performance feature vector is used as the target element, and the sensitivity and independence index of the target element are calculated. The feature importance of the target element is calculated based on the sensitivity and independence index. The feature importance of each element is obtained by traversing the process. The tree weight is calculated based on the feature importance in the historical process. The initial anomaly score of any tree is obtained by using an isolated forest for the process to be detected. The first product of the initial anomaly score of any tree and the tree weight is calculated. The first product of each tree is obtained by traversing the process. The sum of all first products is used as the weighted anomaly score of the process to be detected. The process anomaly detection is completed based on the weighted anomaly score.
[0007] Preferably, the calculation of the first coordinated shock load index of the influent flow sequence includes: obtaining the extreme points of the influent flow sequence, including maximum and minimum values, and constructing an extreme value sequence by arranging all extreme points in chronological order; using the mean of the extreme value sequence as the average shock intensity feature; using the ratio of the standard deviation of the extreme value sequence to the mean of the extreme value sequence as the coefficient of variation, calculating the first ratio of the total number of extreme points to the number of occurrences of a preset event, and calculating the first negative exponent of the first ratio, and using the product of the coefficient of variation and the first negative exponent as the confidence-weighted fluctuation feature; using the least squares method to obtain the slope value for all extreme points in the extreme value sequence, and using the slope value as the trend feature; calculating the second product of a preset first adjustment coefficient and the confidence-weighted fluctuation feature, and calculating the first sum of the second product and 1; calculating the third product of a preset second adjustment coefficient and the trend feature, and calculating the second sum of the third product and 1; and using the product of the average shock intensity feature, the first sum, and the second sum as the first coordinated shock load index of the influent flow sequence.
[0008] Preferably, the performance feature vector includes: taking the mean of the pre-membrane pressure sequence as the first mean and the standard deviation of the pre-membrane pressure sequence as the first standard deviation; similarly obtaining the second mean and second standard deviation of the post-membrane pressure sequence, the third mean and third standard deviation of the permeate flow rate sequence, and the fourth mean and fourth standard deviation of the permeate conductivity sequence; and constructing the first mean, first standard deviation, second mean, second standard deviation, third mean, third standard deviation, fourth mean, and fourth standard deviation as the performance feature vector.
[0009] Preferably, the calculation of the sensitivity of the target element includes: calculating the mean of the performance feature vector of each process in the target cluster as the center vector of the target cluster, and calculating the Euclidean distance between the performance feature vector of the target process and the center vector; establishing a univariate linear regression model with the Euclidean distance from the target process to the center of the target cluster as the independent variable and the target element as the dependent variable, and obtaining the regression coefficients using the least squares method; calculating the mean square error of the univariate linear regression model corresponding to the target element, and calculating the second negative exponent of the mean square error, and normalizing the product of the absolute value of the regression coefficient and the second negative exponent as the sensitivity of the target element.
[0010] Preferably, the calculation of the independence index of the target element includes: calculating the mean of the performance feature vector of each process in the target cluster as the center vector of the target cluster, and calculating the Euclidean distance between the performance feature vector of the target process and the center vector; traversing to obtain the Euclidean distance between the performance feature vector of each process and the center vector, and calculating the mean of all Euclidean distances in the target cluster; calculating the ratio of the Euclidean distance of the target process to the mean distance, and using the negative exponent of the ratio as the representative weight of the target process; multiplying the representative weight of the target process by the target element to obtain the weighted target element, traversing to obtain the weighted target element of each process, and constructing a weighted element sequence; taking any element other than the target element in the performance feature vector as the control element, and similarly obtaining the weighted element sequence of the control element according to the weighted element sequence of the target element; calculating the correlation coefficient between the weighted element sequence of the target element and the weighted element sequence of the control element, and calculating the absolute value of the correlation coefficient; traversing to obtain the absolute value of the correlation coefficient between the target element and each control element, taking the mean of the absolute values of all correlation coefficients as the fifth mean, and using the difference between 1 and the fifth mean as the independence index of the target element.
[0011] Preferably, the feature importance of the target element includes: calculating the fourth product of the sensitivity of the target element and the independence index of the target element, traversing to obtain the fourth product of each element, calculating the cumulative sum of the fourth products; and using the ratio of the fourth product of the target element to the cumulative sum of the fourth products as the feature importance of the target element.
[0012] Preferably, the tree weight calculation for feature importance in the historical process includes: obtaining several trees using an isolated forest for the target cluster; obtaining the frequency of the target element in any tree; calculating the fifth product of the frequency of the target element and the feature importance of the target element; traversing to obtain the fifth product of all elements; calculating the cumulative sum of the fifth products; calculating the cumulative sum of the frequencies of all elements; using the ratio of the cumulative sum of the fifth products to the cumulative sum of the frequencies as the importance score of any tree; and using the normalized importance score as the tree weight of any tree.
[0013] Secondly, the present invention also provides a data fusion-based leachate treatment process anomaly detection system, comprising: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned data fusion-based leachate treatment process anomaly detection method is implemented.
[0014] The present invention has the following effects: This invention constructs operating condition vectors for operating condition clustering, dividing the dynamic and continuous operation process into several relatively steady-state operating condition clusters, thus achieving sub-operating condition detection and eliminating interference caused by operating condition drift at the source. Secondly, this invention innovatively introduces a feature importance evaluation mechanism based on sensitivity and independence, which can intelligently identify which parameter fluctuations are normal physical responses to changes in operating conditions (high sensitivity, low independence) and which are local anomalies caused by equipment failures (low sensitivity, high independence). By using a tree weighting mechanism to differentiate between these in isolated forests, it effectively suppresses the probability of normal fluctuations being misjudged as faults, improving the accuracy of anomaly detection results. Attached Figure Description
[0015] Figure 1 This is a flowchart of an anomaly detection method for leachate treatment process based on data fusion, according to an embodiment of the present invention. Detailed Implementation
[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.
[0017] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0018] Reference Figure 1 A method for detecting anomalies in leachate treatment processes based on data fusion includes steps S1-S4, as detailed below: S1: Collect the influent flow rate sequence, influent chemical oxygen demand sequence, and performance parameter sequence set for any process in the history. The performance parameter sequence set includes the pre-membrane pressure sequence, post-membrane pressure sequence, permeate flow rate sequence, and permeate conductivity sequence.
[0019] In one embodiment, during the data acquisition phase for abnormal detection of the leachate treatment process, two types of key process parameter time series sequences for any process need to be obtained from the historical monitoring database of the reverse osmosis (RO) unit: one is the operating condition parameter sequence, including the feed water flow rate sequence and the feed water chemical oxygen demand sequence, which respectively characterize the hydraulic load and fouling load on the system; the other is a set of performance parameter sequences, including the pre-membrane pressure sequence, post-membrane pressure sequence, permeate flow rate sequence, and permeate conductivity sequence, which together reflect the operating response and performance of the RO membrane unit under specific operating conditions; pre-processing operations such as missing value interpolation, median filtering for noise reduction, low-pass filtering for high-frequency interference suppression, and standardization are performed simultaneously on all sequences.
[0020] S2: Calculate the first co-current impact load index of the influent flow rate sequence, and similarly obtain the second co-current impact load index of the influent chemical oxygen demand sequence. Construct the first co-current impact load index and the second co-current impact load index into a working condition vector. Based on the working condition vector, cluster all processes in history to obtain several working condition clusters.
[0021] In one embodiment, extreme points of the influent flow rate sequence are obtained, including maximum and minimum values, and all extreme points are constructed into an extreme value sequence in chronological order.
[0022] The mean of the extreme value sequence is used as the average impact intensity characteristic.
[0023] It should be noted that extreme point identification is a prerequisite step in constructing the collaborative impact load index, used to extract key peak and valley information characterizing hydraulic impact events in the influent flow sequence. However, when no local maxima or minima are identified in the influent flow sequence within a preset sampling time window, it indicates that the hydraulic load on the system is in a relatively stable state during that period, with no significant impact events occurring. Forcibly extracting extreme value features at this time would lead to data loss or statistical distortion. Therefore, this invention uses the arithmetic mean of the influent flow sequence as its average impact intensity feature to characterize the background load level under conditions without impact events. This average impact intensity feature quantifies the average hydraulic energy input received by the system within a unit time window, serving as a fundamental measure of load action and ensuring the completeness and continuity of the feature vector construction.
[0024] The ratio of the standard deviation of the extreme value sequence to the mean of the extreme value sequence is used as the coefficient of variation. The first ratio of the total number of extreme points to the number of occurrences of the preset event is calculated, and the first negative exponent of the first ratio is calculated. The product of the coefficient of variation and the first negative exponent is used as the confidence-weighted fluctuation feature.
[0025] It should be noted that the coefficient of variation reflects the relative fluctuation amplitude of a single impact event, and the first negative exponent value characterizes the frequency confidence of the impact event. A smaller total number of extreme points indicates sparse load impact times and a relatively stable system condition. In this case, even if individual extreme points fluctuate significantly, the overall time-period fluctuation characteristics should be reasonably suppressed; therefore, the weight is reduced by using a negative exponent value. Conversely, a larger total number of extreme points indicates frequent impact times, and the confidence-weighted fluctuation characteristics are statistically significant. The first negative exponent value approaches 1, fully preserving the true fluctuation information. This method effectively distinguishes between high-frequency impact fluctuations and occasional extreme disturbances.
[0026] The slope value is obtained by using the least squares method for all extreme points in the extreme value sequence, and the slope value is used as the trend feature.
[0027] It should be noted that the slope value reflects the overall trend of load impact. The sign of the slope intuitively represents the overall evolution direction of the load impact within the sampling time window: a positive value reflects a cumulative upward trend in load, while a negative value reflects a decreasing trend in load. The absolute value quantifies the instantaneous rate of load change; the larger the absolute value, the more drastic the rise / fall of hydraulic load or pollutant load per unit time, which is a key indicator for characterizing the intensity and direction of dynamic load evolution. At the level of the operating mechanism of the leachate RO system, the load trend characteristics have clear process semantics: a rapid upward trend often indicates an intensified influent impact or insufficient buffering capacity of the equalization tank, which may accelerate concentration polarization and pollutant deposition on the membrane surface, while a rapid downward trend may reflect a sudden change in influent scheduling or the adjustment response of the pretreatment unit. Both of these will significantly change the coupling relationship between the instantaneous operating point and performance parameters of the membrane unit.
[0028] Calculate the second product of the preset first adjustment coefficient and the confidence-weighted fluctuation characteristic, and calculate the first sum of the second product and 1; calculate the third product of the preset second adjustment coefficient and the trend characteristic, and calculate the second sum of the third product and 1; use the product of the average impact intensity characteristic, the first sum and the second sum as the first coordinated impact load index of the influent flow sequence.
[0029] The average impact intensity characteristic represents the baseline load level, directly determining the reference stress the system bears. The confidence-weighted fluctuation characteristic and trend characteristic act as amplification factors through the first and second sums. When fluctuations or trends are significant, the first co-impact load index exhibits non-linear growth. The average impact intensity characteristic, confidence-weighted fluctuation characteristic, and trend characteristic work synergistically to accurately characterize the complex mechanism by which fluctuations exacerbate energy load and trends superimpose fluctuation risks.
[0030] The second synergistic shock load index of the influent chemical oxygen demand sequence is obtained by analogy with the method for obtaining the first synergistic shock load index.
[0031] The first and second coordinated impact load indices are combined to construct a working condition vector, which is a two-dimensional vector. The load condition vector set is standardized to convert all features into a standard normal distribution with a mean of 0 and a standard deviation of 1, ensuring that the contribution weights of the two types of load features to the load condition pattern division are balanced and comparable in the subsequent clustering process. Then, the standardized load condition vector is used as input, and the K-means++ clustering algorithm is used for adaptive and fine division of the load condition patterns. When determining the optimal number of clusters, a dual verification mechanism of elbow rule and silhouette coefficient is adopted: on the one hand, the intra-cluster error sum of squares curve is plotted as the K value changes from 2 to 10, and the elbow inflection point where the rate of decrease of the intra-cluster error sum of squares curve slows down is identified. This point indicates that the marginal benefit of improving intra-cluster compactness begins to decrease. On the other hand, the average silhouette coefficient under each K value is calculated simultaneously to quantify the fit between the sample and its cluster and the separation from other clusters. The optimal K value that achieves the best balance between clustering effect (high silhouette coefficient) and model complexity (low K value) is selected. Finally, K-means++ clustering iteration is performed to assign a unique load condition cluster to each historical normal sample. For each generated operating condition cluster, the arithmetic mean of the operating condition vectors of all processes within it is calculated, which forms the central vector of the cluster.
[0032] S3: Take any operating condition cluster as the target cluster, and any process in the target cluster as the target process. Calculate the mean and variance of each sequence in the set of performance parameter sequences of the target process, construct the performance feature vector of the target process, take any element in the performance feature vector as the target element, calculate the sensitivity of the target element, and calculate the independence index of the target element. Calculate the feature importance of the target element based on the sensitivity and independence index; iterate through the set to obtain the feature importance of each element.
[0033] In one embodiment, any cluster of operating conditions is taken as the target cluster, and any process within the target cluster is taken as the target process. For the target process, the mean of the pre-membrane pressure sequence is taken as the first mean, and the standard deviation of the pre-membrane pressure sequence is taken as the first standard deviation. Similarly, the second mean and second standard deviation of the post-membrane pressure sequence, the third mean and third standard deviation of the permeate flow rate sequence, and the fourth mean and fourth standard deviation of the permeate conductivity sequence are obtained. The first mean, first standard deviation, second mean, second standard deviation, third mean, third standard deviation, fourth mean, and fourth standard deviation are used to construct a performance feature vector. The performance feature vector is an eight-dimensional vector.
[0034] Calculate the Euclidean distance between the performance feature vector of the target process and its center vector.
[0035] Using any element in the performance feature vector as the target element, a univariate linear regression model is established with the Euclidean distance from the target process to the center of the target cluster as the independent variable and the target element as the dependent variable. The regression coefficients are obtained using the least squares method.
[0036] Calculate the mean squared error of the univariate linear regression model corresponding to the target element, and calculate the second negative exponent of the mean squared error. The result of normalizing the product of the absolute value of the regression coefficient and the second negative exponent is used as the sensitivity of the target element.
[0037] It should be noted that the larger the absolute value of the regression coefficient, the more significant the impact of external load shock on the target element. Its numerical drift mainly originates from the normal physical response caused by the deviation of operating conditions, and it has a high dependence on operating conditions. Conversely, if the absolute value of the regression coefficient is close to 0, it indicates that the target element is not sensitive to the deviation of operating conditions. Its fluctuations reflect more the independent changes in the state of the membrane unit itself or random noise, rather than being driven by external loads.
[0038] The formula for sensitivity is constructed as follows: when the fitting quality is high, i.e., the mean square error is small, the second negative exponent value is close to 1, and the sensitivity is fully preserved, indicating that the target element's response to deviations from the operating condition is highly reliable; when the fitting quality is low, i.e., the mean square error is large, the second negative exponent value decays rapidly and approaches 0, effectively suppressing unreliable sensitivity estimates and avoiding mistaking noise interference for the operating condition response; by fusing fitting error information, the assessment of feature importance changes from simply looking at the magnitude of change to looking at the reliable magnitude of change.
[0039] The mean of the performance feature vector of each process in the target cluster is calculated and used as the center vector of the target cluster. The Euclidean distance between the performance feature vector of the target process and the center vector is calculated. This process is iterated to obtain the Euclidean distance between the performance feature vector of each process and the center vector, and the mean of all Euclidean distances in the target cluster is calculated. The ratio of the Euclidean distance of the target process to the mean distance is calculated, and the negative exponent of the ratio is used as the representative weight of the target process.
[0040] Multiply the representative weight of the target process with the target element to obtain the weighted target element. Iterate through each process to obtain the weighted target element and construct a weighted element sequence. Take any element other than the target element in the performance feature vector as the control element. Similarly, obtain the weighted element sequence of the control element based on the weighted element sequence of the target element. Calculate the correlation coefficient between the weighted element sequence of the target element and the weighted element sequence of the control element, and calculate the absolute value of the correlation coefficient.
[0041] Iterate through the data to obtain the absolute value of the correlation coefficient between the target element and each control element. Take the mean of the absolute values of all correlation coefficients as the fifth mean, and take the difference between 1 and the fifth mean as the independence index of the target element.
[0042] Calculate the fourth product of the sensitivity of the target element and the independence index of the target element, iterate through each element to obtain the fourth product, and calculate the cumulative sum of the fourth products; use the ratio of the fourth product of the target element to the cumulative sum of the fourth products as the feature importance of the target element.
[0043] It should be noted that feature importance differs from traditional methods that rely solely on statistical measures such as the mean and standard deviation of parameters for feature selection. The independence index actively mines the linear dependency structure between the eight elements within the performance feature vector, identifying key indicators that can independently reflect the local state changes of membrane units under specific operating conditions, breaking away from the group fluctuation pattern. When the independence index is integrated with the operating condition response sensitivity, the two form a two-dimensional discrimination framework of "external drive-internal response": if a feature with high sensitivity also has low independence, it indicates that its fluctuation mainly stems from the systemic coupling response caused by operating condition deviation, which falls within the scope of normal physical laws and can be reasonably suppressed in feature importance calculation; conversely, if a feature with high independence has low sensitivity, its abnormal deviation is more likely to point to local faults such as membrane fouling, sealing failure, or sensor drift, and should receive higher attention in the weighted isolation forest tree weight allocation.
[0044] S4: Calculate tree weights based on the importance of features in the historical process. Use an isolated forest to obtain the initial anomaly score of any tree in the process to be detected. Calculate the first product of the initial anomaly score of any tree and the tree weight. Traverse the process to obtain the first product of each tree. Sum all the first products as the weighted anomaly score of the process to be detected. Complete the process anomaly detection based on the weighted anomaly score.
[0045] In one embodiment, an isolated forest is used to obtain several trees for the target cluster. The frequency of the target element in any tree is obtained. The fifth product of the frequency of the target element and the feature importance of the target element is calculated. The fifth product of all elements is obtained by traversing the tree. The cumulative sum of the fifth products is calculated. The cumulative sum of the frequencies of all elements is calculated. The ratio of the cumulative sum of the fifth products to the cumulative sum of the frequencies is used as the importance score of any tree. The normalized importance score is used as the tree weight of any tree.
[0046] It should be noted that the tree weight is obtained by statistically analyzing the frequency of each target element in any tree as a split point, multiplying and summing it with the feature importance of the corresponding target element, and then normalizing it. The tree weight quantifies the degree of fit between the tree splitting path and the operating mechanism of the leachate RO system. The larger the tree weight value, the more the tree adopts key features that are sensitive to deviations from the operating conditions and have independent information content during the random splitting process, and the more its decision-making process conforms to the process logic. Conversely, it means that the tree may over-rely on noise features or redundant indicators.
[0047] According to steps S1 to S3, the operating condition vector of the process to be detected is obtained, and the operating condition cluster to which the process to be detected belongs is obtained based on the operating condition vector. An initial anomaly score for any tree is obtained using an isolated forest algorithm for the process to be detected. The first product of the initial anomaly score and the tree weight of any tree is calculated. The first product of each tree is obtained by traversing the forest. The sum of all first products is used as the weighted anomaly score of the process to be detected. If the weighted anomaly score is greater than a preset anomaly threshold, an alarm signal is generated and sent. For example, the preset anomaly threshold is 0.6.
[0048] The system includes a processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement a data fusion-based method for detecting anomalies in leachate treatment processes according to the first aspect of the present invention.
[0049] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.
[0050] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.
Claims
1. A method for detecting anomalies in leachate treatment processes based on data fusion, characterized in that, include: Collect the influent flow rate sequence, influent chemical oxygen demand sequence, and performance parameter sequence set for any process in history. The performance parameter sequence set includes the pre-membrane pressure sequence, post-membrane pressure sequence, permeate flow rate sequence, and permeate conductivity sequence. The first co-current impact load index of the influent flow rate sequence is calculated, and the second co-current impact load index of the influent chemical oxygen demand sequence is obtained in the same way. The first co-current impact load index and the second co-current impact load index are used to construct a working condition vector. Based on the working condition vector, all processes in history are clustered to obtain several working condition clusters. Take any set of operating conditions as the target set, take any process in the target set as the target process, calculate the mean and variance of each sequence in the set of performance parameter sequences of the target process, construct the performance feature vector of the target process, take any element in the performance feature vector as the target element, calculate the sensitivity of the target element, calculate the independence index of the target element, and calculate the feature importance of the target element based on the sensitivity and independence index. Iterate through the elements to obtain the feature importance of each element; Based on the historical process feature importance calculation tree weights, an isolated forest is used to obtain the initial anomaly score of any tree in the process to be detected. The first product of the initial anomaly score of any tree and the tree weight is calculated. The first product of each tree is obtained by traversing the process. The sum of all first products is used as the weighted anomaly score of the process to be detected. The process anomaly detection is completed based on the weighted anomaly score.
2. The method for detecting anomalies in leachate treatment processes based on data fusion according to claim 1, characterized in that, The first coordinated shock load index for calculating the influent flow sequence includes: Obtain the extreme points of the influent flow rate sequence, including maximum and minimum values, and construct an extreme value sequence by arranging all extreme points in chronological order. The mean of the extreme value sequence is used as the average impact intensity characteristic; The ratio of the standard deviation of the extreme value sequence to the mean of the extreme value sequence is used as the coefficient of variation. The first ratio of the total number of extreme points to the number of occurrences of the preset event is calculated, and the first negative exponent of the first ratio is calculated. The product of the coefficient of variation and the first negative exponent is used as the confidence-weighted fluctuation feature. The slope value is obtained by using the least squares method for all extreme points in the extreme value sequence, and the slope value is used as a trend feature. Calculate the second product of the preset first adjustment coefficient and the confidence-weighted volatility characteristic, and calculate the first sum of the second product and 1; calculate the third product of the preset second adjustment coefficient and the trend characteristic, and calculate the second sum of the third product and 1. The product of the average impact intensity characteristic, the first sum, and the second sum is used as the first co-impact load index of the influent flow sequence.
3. The method for detecting anomalies in leachate treatment processes based on data fusion according to claim 1, characterized in that, The performance feature vector includes: The mean of the pre-membrane pressure sequence is used as the first mean, and the standard deviation of the pre-membrane pressure sequence is used as the first standard deviation; similarly, the second mean and second standard deviation of the post-membrane pressure sequence, the third mean and third standard deviation of the permeate flow rate sequence, and the fourth mean and fourth standard deviation of the permeate conductivity sequence are obtained. The first mean, first standard deviation, second mean, second standard deviation, third mean, third standard deviation, fourth mean, and fourth standard deviation are used to construct a performance feature vector.
4. The method for detecting anomalies in leachate treatment processes based on data fusion according to claim 1, characterized in that, The sensitivity of the target element to be calculated includes: The mean of the performance feature vector of each process in the target cluster is calculated as the center vector of the target cluster, and the Euclidean distance between the performance feature vector of the target process and the center vector is calculated. A univariate linear regression model is established with the Euclidean distance from the target process to the center of the target cluster as the independent variable and the target element as the dependent variable. The regression coefficients are obtained by using the least squares method. Calculate the mean squared error of the univariate linear regression model corresponding to the target element, and calculate the second negative exponent of the mean squared error. The result of normalizing the product of the absolute value of the regression coefficient and the second negative exponent is used as the sensitivity of the target element.
5. The method for detecting anomalies in leachate treatment processes based on data fusion according to claim 1, characterized in that, The independence index of the target element includes: The mean of the performance feature vector of each process in the target cluster is calculated as the center vector of the target cluster, and the Euclidean distance between the performance feature vector of the target process and the center vector is calculated. The Euclidean distance between the performance feature vector and the center vector of each process is obtained by traversing the process, and the mean distance of all Euclidean distances in the target cluster is calculated. Calculate the ratio of the Euclidean distance of the target process to the mean distance, and use the negative exponent of the ratio as the representative weight of the target process; Multiply the representative weight of the target process by the target element to obtain the weighted target element, traverse the process to obtain the weighted target element of each process, and construct a weighted element sequence. Take any element other than the target element in the performance feature vector as the control element, and similarly obtain the weighted element sequence of the control element based on the weighted element sequence of the target element. Calculate the correlation coefficient between the weighted element sequence of the target element and the weighted element sequence of the control element, and calculate the absolute value of the correlation coefficient; Iterate through the data to obtain the absolute value of the correlation coefficient between the target element and each control element. Take the mean of the absolute values of all correlation coefficients as the fifth mean, and take the difference between 1 and the fifth mean as the independence index of the target element.
6. The method for detecting anomalies in leachate treatment processes based on data fusion according to claim 1, characterized in that, The importance of the features of the target element includes: Calculate the fourth product of the sensitivity of the target element and the independence index of the target element, iterate through each element to obtain the fourth product, and calculate the cumulative sum of the fourth products. The ratio of the fourth product of the target element to the sum of the fourth product is used as the feature importance of the target element.
7. The method for detecting anomalies in leachate treatment processes based on data fusion according to claim 1, characterized in that, The weights of the feature importance calculation tree in the historical process include: Use an isolated forest to obtain several trees for the target cluster, obtain the frequency of the target element in any tree, calculate the fifth product of the frequency of the target element and the feature importance of the target element, traverse to obtain the fifth product of all elements, and calculate the cumulative sum of the fifth products. Calculate the cumulative frequency value of all elements, and use the ratio of the fifth-times cumulative value to the cumulative frequency value as the importance score of any tree. Use the normalized importance score as the tree weight of any tree.
8. A data fusion-based leachate treatment process anomaly detection system, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement a data fusion-based method for detecting anomalies in leachate treatment processes according to any one of claims 1-7.