Environmental protection illegal behavior intelligent identification system based on multi-mode AI

Abstract

The invention discloses an intelligent recognition system for environmental protection violation behaviors based on multi-modal AI, belongs to the technical field of environmental protection monitoring and video recognition, and aims to solve the problems of data fragmentation, incomplete violation recognition, low precision and insufficient supervision efficiency in traditional environmental protection monitoring. The system comprises a data analysis module, a dynamic sampling module, a data preprocessing module and an intelligent identification module. Analyzing the production log data, and generating a sewage discharge sequence of a sewage discharge port; generating sampling parameters in combination with the emission sequence and historical violation characteristics, and extracting video frames to form a monitoring frame sequence; positioning a sewage draining exit to obtain a dynamic monitoring area, and identifying emission frames to construct an emission frame data set; processing key evaluation indexes of emission frames, identifying excessive emission and abnormal frames, marking violation behavior types, and generating a violation identification report; according to the invention, full-dimension and high-precision recognition of environmental protection violation behaviors is realized, a traceable and efficient decision support is provided for environmental protection supervision, and the supervision efficiency is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of environmental protection monitoring and video recognition technology, more particularly to an intelligent identification system for environmental protection violations based on multi-modal AI. BACKGROUND

[0002] In the field of environmental protection supervision, there are many problems in traditional monitoring and violation identification methods. On the one hand, data acquisition and integration are difficult. Existing monitoring technologies mainly rely on traditional sensors and laboratory analysis, which have low sampling frequency and insufficient real-time performance, making it difficult to meet the demand for rapid grasp of pollution conditions. Even if remote real-time monitoring is achieved by Internet of Things devices, data transmission and privacy protection still need to be strengthened. Different sources of monitoring data, such as production log data and environmental monitoring data, lack effective fusion means, making it difficult to establish a comprehensive and dynamic pollution correlation analysis system, and making it difficult to accurately understand the internal relationship between production activities and sewage discharge. On the other hand, the accuracy and efficiency of violation identification need to be improved. Manual patrol methods are not only inefficient, but also difficult to cover a wide area. For hidden violations such as non-working hours, holidays or night-time illegal discharges, it is difficult to detect. When relying on traditional algorithms to analyze video images and other data, misjudgment and omission are likely to occur. In the face of complex environmental scenes and massive data, it is difficult to timely and accurately identify abnormal discharge behavior and classify and mark violations, making it difficult to effectively carry out environmental protection supervision and unable to meet the increasingly stringent environmental protection supervision requirements. Therefore, in order to overcome these limitations, the present application proposes an intelligent identification system for environmental protection violations based on multi-modal AI. SUMMARY

[0003] In view of the deficiencies in the prior art, the present application aims to provide an intelligent identification system for environmental protection violations based on multi-modal AI, which solves the problem of how to comprehensively and accurately identify environmental protection violations, while efficiently integrating multi-source data, scientifically planning sampling strategies, clearly marking violation types and generating traceable judgment basis to improve environmental protection supervision efficiency.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0005] The intelligent identification system for environmental protection violations based on multi-modal AI comprises a data analysis module, a dynamic sampling module, a data preprocessing module and an intelligent identification module:

[0006] The data analysis module is used to receive and analyze production log data of the monitoring target, extract multi-modal features of production parameters, construct a time series feature matrix, and establish a dynamic mapping relationship between production parameters and sewage discharge, construct a time series correlation learning model, and generate a sewage discharge sequence for each sewage outlet;

[0007] The dynamic sampling module is configured to construct a sampling weight evaluation model based on the sewage discharge sequence of each sewage outlet and a historical illegal characteristic vector, generate a dynamic weight parameter for each sewage discharge period, construct a periodic sampling sequence, extract video frames, and construct a monitoring frame sequence.

[0008] The data preprocessing module is configured to locate the sewage outlet in each video frame in the monitoring frame sequence, obtain a dynamic monitoring area of each video frame through a spatial projection matrix, divide a dynamic foreground area through foreground extraction, identify an emission frame through multi-modal emission judgment, and construct a sewage outlet emission frame data set.

[0009] The intelligent identification module is configured to perform time standardization processing on key evaluation indexes of the emission frame, identify over-discharge anomalies, extract sewage feature parameters of the dynamic foreground area, identify abnormal illegal frames and mark illegal behavior types, and generate an illegal identification report.

[0010] Specifically, the step of obtaining the dynamic monitoring area of each video frame includes:

[0011] Feature extraction is performed on each video frame of the monitoring frame sequence, feature marker points of the sewage outlet are obtained, similarity calculation is performed on the feature marker points and feature marker points in a reference feature library of the sewage outlet, and matching feature pairs are screened.

[0012] A perspective transformation matrix of the video frame and a standard perspective image in the reference feature library is calculated according to the matching feature pairs, perspective correction is performed on the feature marker point coordinates of the sewage outlet, and a center position of the sewage outlet is fitted as a real-time coordinate of the sewage outlet in the current video frame.

[0013] Camera intrinsic calibration data of the monitoring device and shooting angle parameters of the current video frame are called, and a spatial projection matrix of a mapping relationship between three-dimensional space coordinates in a standard perspective and two-dimensional image coordinates of the current video frame is established.

[0014] According to the spatial projection matrix, the real-time coordinates of the sewage outlet and the relative coordinate position of the emission area, the two-dimensional image coordinates of the emission area in the current video frame are calculated to generate an emission area boundary; and an expansion coefficient is preset to expand the emission area boundary range to form a dynamic monitoring area of the monitoring frame.

[0015] Specifically, the step of constructing the sewage outlet emission frame data set includes:

[0016] A monitoring evaluation range is set, and a historical video frame set within the monitoring evaluation range of the current video frame is extracted.

[0017] The dynamic monitoring area of the current video frame and the corresponding dynamic monitoring area of the historical video frame set are compared in terms of gray value, and the gray value change difference of each pixel point in the time dimension is calculated.

[0018] A gray scale change threshold is set to identify dynamic foreground pixels, and a connected region analysis is used to mark foreground pixel clusters for the dynamic foreground pixels. The area of each foreground pixel cluster is calculated to screen out the dynamic foreground region of the current video frame.

[0019] For the extracted dynamic foreground region, a multi-modal discharge judgment is performed to calculate a sewage discharge score, which includes a color matching score, a motion matching score, and a diffusion matching score.

[0020] A qualified threshold for the sewage discharge score is set to identify discharge frames in the monitoring sequence, which are sorted in chronological order. The time interval between adjacent discharge frames is calculated to divide discharge events and mark the timestamps of the start and end frames of each discharge event.

[0021] Key evaluation indicators are extracted for each discharge frame and bound to the outfall identifier and the discharge event identifier to construct an outfall discharge frame data set.

[0022] Specifically, the multi-modal discharge judgment includes:

[0023] Color space transformation is performed on the dynamic foreground region to extract color features of the dynamic foreground region. Similarity comparison is performed between the color features and a pre-stored typical sewage color feature library to generate a color matching score.

[0024] The motion trajectories of pixels in the dynamic foreground region are tracked to calculate the proportion of trajectory-interrupted pixels. The angle between the motion direction of pixels in the dynamic foreground region and the pre-set discharge direction of sewage from the outfall is calculated. The proportion of direction-matched pixels with an angle less than a pre-set angle threshold is calculated. The trajectory-interrupted pixel proportion and the direction-matched pixel proportion are combined to generate a motion matching score through weighted summation with a pre-set weight.

[0025] A set of previous video frames of the current video frame is retrieved. The boundary coordinates and pixel number of the dynamic foreground region in each previous video frame are extracted. The area of the circumscribed rectangle of the dynamic foreground region in each previous video frame is calculated to obtain the area of the dynamic foreground region in each previous video frame. The area difference between the dynamic foreground region of the current video frame and that of the previous video frames is calculated. The area difference is divided by the time interval to obtain the area change rate of the dynamic foreground region. The expansion distance of the dynamic foreground region boundary in each direction is calculated. The expansion distance is divided by the time interval to obtain the expansion speed of the dynamic foreground region boundary. The area change rate of the dynamic foreground region and the expansion speed of the dynamic foreground region boundary are combined to generate a diffusion matching score through level mapping scoring.

[0026] Specifically, the steps of identifying over-discharge anomalies include:

[0027] According to the outfall identifier and the discharge event identifier, the discharge frame subsets of each discharge event are extracted from the outfall discharge frame data set.

[0028] According to the sampling time stamp of each discharge frame, the sampling time interval of adjacent discharge frames in the discharge frame subset is calculated, and the time standardization processing is performed on the diffusion characteristic parameters and the motion characteristic parameters in the key evaluation indicators;

[0029] Based on the time-standardized key evaluation indicators, the standardized average time flow rate and the dynamic foreground area are used to calculate the sewage volume in the sampling time of a single discharge frame in combination with the sampling duration of the single discharge frame; the sewage volumes of all single discharge frames in the discharge frame subset are accumulated in the order of the sampling time stamps to obtain the total discharge amount of the entire discharge event;

[0030] Based on the association mapping relationship between the discharge event and the sewage discharge period, the expected sewage discharge amount of the discharge event is extracted from the sewage discharge sequence, the discharge difference between the total discharge amount of the discharge event and the corresponding expected sewage discharge amount is calculated, and if it is greater than a preset discharge threshold, it is determined that the current discharge event has a potential illegal behavior and is marked as an over-discharge anomaly.

[0031] Specifically, the step of identifying abnormal illegal frames and marking the illegal behavior type comprises:

[0032] The sewage feature parameters of the dynamic foreground area of the discharge frames in the discharge frame subset of the discharge event are extracted, including the foam morphology parameter and the sewage transparency parameter, and after standardization, the compliance judgment vector of each discharge frame is constructed;

[0033] A standard judgment vector is set to quantify the deviation degree of the discharge frame compliance judgment vector and the compliance state, and the cosine similarity algorithm is used to calculate the similarity value of the compliance judgment vector of each discharge frame with the standard judgment vector, and a violation similarity threshold is set to identify abnormal illegal frames;

[0034] If there are abnormal illegal frames in the discharge frame subset of the discharge event, the proportion of the abnormal illegal frames in the discharge frame subset is counted, and the abnormal illegal frames are classified in combination with the parameter deviation of each dimension of the abnormal illegal frame compliance judgment vector, and the illegal behavior type of the abnormal illegal frame is marked, and the proportion of each type of abnormal illegal frame in the discharge frame subset is counted.

[0035] Specifically, the step of constructing the time sequence feature matrix comprises:

[0036] The production log data is received, and the production log data is divided into data formats according to the data format characteristics, so as to perform data preprocessing on the production log data;

[0037] The core elements related to sewage discharge in the production log data are located, and the associated information in different data formats is aggregated through semantic fusion to form a sewage discharge information set of each sewage outlet;

[0038] Based on the sewage discharge information set and the production log data, multi-modal features reflecting the production state are extracted with production parameters as the core. The multi-modal features include quantitative features, derived features and qualitative features.

[0039] The multi-modal features are converted into a unified data dimension and sorted in chronological order to construct a time sequence feature matrix for each sewage outlet.

[0040] Specifically, the step of generating the sewage discharge sequence of each sewage outlet includes:

[0041] Based on the time sequence feature matrix of each sewage outlet and the corresponding historical discharge data, a time sequence correlation learning model is constructed to establish a dynamic mapping relationship between the production parameters and the sewage discharge, and to predict the sewage discharge of each production period;

[0042] Divide each sewage discharge monitoring period into time units, and extract the sewage discharge rule from the historical discharge data to set the sewage discharge period in each time unit.

[0043] Based on the set sewage discharge period, the predicted sewage discharge of each sewage discharge period is calculated by calling the time sequence correlation learning model, and the discharge time, production link identifier and time sequence identifier are associated to generate a structured sewage discharge sequence for each sewage outlet.

[0044] Specifically, the step of generating the dynamic weight parameter of each sewage discharge period includes:

[0045] Retrieve the historical violation records of the monitoring target, convert the historical violation records into quantifiable violation feature data, and form a historical violation feature vector for each sewage outlet.

[0046] Based on the sewage discharge sequence and the historical violation feature vector of each sewage outlet, a sampling weight evaluation model is constructed to calculate the stability score, the risk score and the discharge amount score, and to generate the dynamic weight parameter of each sewage discharge period by weighted fusion.

[0047] Obtain the monitoring resource capacity data and convert it into quantifiable sampling constraint indicators. Based on the dynamic weight parameter and the sampling constraint indicator, a weight sampling frequency mapping table is used to assign a sampling frequency to each sewage discharge period of each sewage outlet, and a periodic sampling sequence is generated.

[0048] According to the sampling time points in the periodic sampling sequence, frame extraction instructions are sent to the corresponding equipment to extract the monitoring frame sequence of each sewage outlet.

[0049] Specifically, the sampling weight evaluation model includes an emission stability evaluation unit, a historical violation risk evaluation unit and a discharge amount weight evaluation unit.

[0050] The discharge stability evaluation unit is used for calculating the standard deviation of the expected discharge amount in the sewage discharge period, and calculating the deviation proportion with the preset deviation threshold, and mapping the stability score in the 0 to 1 interval according to the deviation proportion;

[0051] The historical violation risk degree evaluation unit is used for calculating the similarity value of the multi-dimensional production feature matrix of the current sewage discharge period and the multi-dimensional production feature matrix of the historical violation period by the cosine similarity algorithm, and combining the ratio of the historical violation frequency of the sewage discharge period and the average violation frequency in the whole cycle to generate a risk degree score;

[0052] The discharge amount weight evaluation unit converts the expected sewage discharge amount of each sewage discharge period into a discharge amount score in the 0 to 1 interval by a normalization algorithm.

[0053] The beneficial effects of the present application are:

[0054] The present application effectively solves the problem of scattered multi-source data and difficulty in tracing the discharge source in the traditional technology by integrating production log data and establishing a dynamic correlation between production parameters and sewage discharge amount, and generating a structured sewage discharge sequence; dynamic weight parameters are generated by combining the sewage discharge sequence and historical violation characteristics to develop a targeted periodic sampling strategy, which avoids the defects of resource waste or omission of high-risk discharge periods in the traditional fixed sampling mode; the sewage outlet is accurately positioned and a dynamic monitoring area is constructed, and the emission frame is identified by combining multi-modal emission judgment, which overcomes the shortcomings of fixed monitoring area and inaccurate emission behavior identification in the traditional monitoring area; the key evaluation indicators are time standardized to identify over-discharge, and the sewage characteristic parameters are extracted to construct a compliance judgment vector, and the violation behavior type is classified and marked, which solves the problems of multiple misjudgments and inability to clearly identify the violation type in the traditional violation identification, and finally realizes the whole-process efficient supervision of environmental protection violation behavior from data integration, accurate sampling to classification identification, significantly improves the supervision accuracy and efficiency, and provides comprehensive and traceable basis for environmental protection decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 It is a structural diagram of the environmental protection violation behavior intelligent identification system based on multi-modal AI of the present application;

[0056] Figure 2 It is a flowchart for generating a structured sewage discharge sequence of the present application;

[0057] Figure 3 It is a flowchart for constructing a structured monitoring frame sequence of the present application;

[0058] Figure 4 It is a flowchart for constructing a sewage outlet emission frame data set of the present application;

[0059] Figure 5 It is a flowchart for generating a violation identification report of the present application. DETAILED DESCRIPTION

[0060] Referring to Figure 1 The embodiment introduces an environmental violation behavior intelligent identification system based on multi-modal AI, including a data analysis module, a dynamic sampling module, a data preprocessing module, and an intelligent identification module.

[0061] The data analysis module receives production log data of the monitoring target. The production log data is various types of original records formed by the monitoring target in the production process, covering unstructured text such as production team operation records, equipment maintenance logs, process adjustment instructions, semi-structured forms such as raw material taking account books and energy consumption reports, and structured data such as real-time temperature of reaction kettles, water pump running power, and production line start-stop time. A multi-modal AI technology combining natural language processing and structured data analysis is adopted to locate the core elements related to emission in the production log through entity recognition, and to establish the association logic between the core elements through relationship extraction, from which a sewage discharge information set of each sewage outlet is extracted. The sewage discharge information set refers to key evaluation indicators that can reflect the sewage discharge state of the monitoring target, including the corresponding emission amount of the production link, the emission start and end period, the sewage source, and the associated emission outlet number. A multi-dimensional production feature matrix is constructed through feature engineering technology and encoded in time sequence; the dynamic mapping relationship between production parameters and sewage discharge amount is captured, and the expected sewage discharge amount of each time unit is calculated, wherein the time unit refers to a continuous time segment divided according to environmental regulations or the production cycle of the monitoring target, such as each production shift or each daily fixed time period. Combined with the sewage discharge period set by the monitoring target according to environmental regulations or its own production rhythm, i.e., the specific time point at which the emission amount needs to be recorded in advance, the expected sewage discharge amount, the emission time length, the associated production link, and the time sequence marker of each sewage discharge period are integrated to generate a structured sewage discharge sequence of each sewage outlet.

[0062] In this embodiment, by fusing the multi-modal AI technology of natural language processing and structured data analysis, the adaptation limitations of traditional single data processing mode to heterogeneous production log data are broken, full-type coverage analysis of unstructured text, semi-structured form and structured data is realized, the omission of key elements related to emissions due to data format differences is effectively avoided, the completeness and accuracy of the collection and extraction of sewage discharge information are significantly improved, and key evaluation indicators such as discharge amount, discharge period and sewage source are efficiently obtained from production logs; the multi-dimensional production feature matrix constructed by feature engineering technology, combined with time sequence coding and dynamic mapping relationship capture method, strengthens the correlation between production parameters and sewage discharge, and makes the calculation of predicted sewage discharge in each time unit deeply fit the actual production law of the monitoring target, greatly reduces the prediction deviation caused by parameter fragmentation or static correlation, and improves the reliability of the discharge prediction; by combining the requirements of environmental protection regulations and the production rhythm of the monitoring target to set the sewage discharge period, integrating multi-dimensional information to generate a structured sewage discharge sequence, it not only ensures the dual adaptation of sequence time dimension and compliance requirements, production rhythm, but also provides structured and highly available basic data support for subsequent dynamic sampling planning and violation identification by clearly marking the time sequence and associated production links, avoiding the efficiency loss of subsequent processes caused by data fragmentation.

[0063] Please refer to Figure 2 , preferably, the specific steps of generating a structured sewage discharge sequence include:

[0064] The production management system of the monitoring target is connected through a standardized interface, and production log data is received in batches; the production log data is divided according to data format characteristics and pre-processed, and the data format includes unstructured text, semi-structured form and structured data. Data preprocessing refers to special processing operations designed according to different data formats. For example, for unstructured text, remove redundant symbols and duplicate records through text segmentation and format recognition technology; for semi-structured forms, use table structure analysis algorithms to extract the correspondence between table headers and data fields, and fill in missing field identifiers; for structured data, filter abnormal format data through data format verification rules to ensure that the three types of data meet the subsequent analysis requirements and avoid processing interruptions caused by data format chaos.

[0065] Based on the production log data after data preprocessing, a multi-modal AI analysis framework is used to generate a set of sewage discharge information for each sewage outlet. The multi-modal AI analysis framework is a technical architecture designed for heterogeneous production log data that can cooperatively process multiple data formats and extract emission-related information. It includes unified entity recognition to locate core elements related to sewage discharge in production log data, such as production stages, discharge volume, time nodes, and discharge outlet identifiers. At the same time, a cross-modal relationship extraction mechanism is started to establish the association information of core elements in different types of data formats, such as the causal relationship between production parameters and discharge volume, and the time sequence relationship between production operations and discharge behavior. Through semantic fusion, the associated information scattered in different data formats is aggregated to form a set of sewage discharge information for each sewage outlet, achieving overall extraction and correlation of emission-related information in production log data.

[0066] Based on the extracted set of sewage discharge information and production log data, multi-modal features reflecting production status are extracted based on production parameters, including: quantitative features, which are used to represent quantifiable resource consumption and equipment operating status in the production process; such as resource consumption intensity and equipment operating efficiency calculated based on production parameters; derived features, which are used to reveal the internal correlation between production parameters and emission behavior; such as the ratio of production load to emission intensity, and the correlation index of production parameter fluctuation and emission volume change; qualitative features, which are used to describe the process stability and operation standardization that are difficult to quantify in the production process; such as process stability description based on production parameter stability analysis. Through standardization processing, various multi-modal features are converted into a unified data dimension. Time sequence coding technology is used to arrange all multi-modal features in chronological order, constructing a time sequence feature matrix for each sewage outlet covering the entire production process, providing a structured feature basis for subsequent correlation analysis.

[0067] The deep learning architecture of fusing the recurrent neural network and the attention mechanism is used to construct a time series correlation learning model. In the training process of the time series correlation learning model, the time series feature matrix and the corresponding historical emission data are input, the time series evolution law of the multi-modal features is captured through the nonlinear transformation of multiple neurons, the feature influence related to the emission is strengthened by means of the attention weight distribution, the adaptive gradient descent algorithm is used to optimize the model parameters to minimize the prediction error; the time series correlation learning model is used to establish the dynamic mapping relationship between the production parameters and the sewage emission, the correlation mode of the multi-modal feature change and the emission fluctuation under different production states is learned, the sewage emission of each production period of each sewage outlet is estimated, the ability of identifying abnormal fluctuations such as process adjustment and equipment state change in the production process is possessed, and the mapping parameters can be dynamically adjusted according to the fluctuation characteristics, so that the estimation accuracy can be ensured when the production conditions change. The historical emission data refers to the sewage emission data of each time period recorded by the monitoring target in the past production cycle through the actual measurement method, including the emission records under the normal production and abnormal production states, and is used to provide a true emission benchmark for model training.

[0068] According to the cycle requirement of sewage emission monitoring, the adaptive time series segmentation algorithm is used to divide the time unit of each sewage emission monitoring cycle according to the production law characteristics of the monitoring target obtained by analyzing the historical production data, so that the time unit meets the compliance requirement and the production rhythm matching. Based on the divided time unit, the sewage emission period is set in each time unit according to the sewage emission law extracted from the historical emission data, so as to ensure that the time period can accurately cover the potential emission period.

[0069] Based on the set sewage emission period, the predicted sewage emission of each sewage emission period of each sewage outlet is calculated by calling the time series correlation learning model, and the corresponding emission time, associated production link and other auxiliary information are associated. The time series marking rule is used to add a unique time sequence identifier to all sewage emission periods in chronological order. The above information is integrated into a standardized data structure, the integrity and time logic consistency of each field are checked through a logical verification mechanism, abnormal data is corrected or supplemented, and finally a structured sewage emission sequence of each sewage outlet containing multi-dimensional information and clear time sequence is generated.

[0070] The dynamic sampling module fuses the historical violation records of the monitoring target, including the violation period, violation type, and violation frequency, according to the sewage discharge sequence of each sewage outlet, constructs a sampling weight evaluation model, calculates the emission stability and historical violation risk degree of each sewage discharge period in the sewage discharge sequence of each sewage outlet, and generates a dynamic weight parameter for each sewage discharge period in combination with the predicted sewage discharge amount, wherein a period with high predicted discharge amount, poor stability, and high violation risk degree is given a higher weight. Based on the dynamic weight parameter and the monitoring resource capacity constraint, such as the video frame storage upper limit and the device sampling rate, a periodic sampling sequence is generated. According to the sampling sequence, the video frame at the specified time point is extracted from the monitoring device of the corresponding sewage outlet through the video stream control interface, and metadata such as the sewage outlet identifier, sampling period, and weight parameter are added to each frame in synchronization. The structured monitoring frame sequence is integrated and constructed, and analysis data for accurately matching the emission period are provided for the subsequent intelligent recognition module.

[0071] In the present embodiment, by means of the sampling weight evaluation model, the emission stability, historical violation risk degree, and predicted discharge amount are included in the unified weight calculation system, avoiding the problem of low weight of high-risk period or excessive attention to low-risk period caused by traditional single-dimensional evaluation, and ensuring that the dynamic weight parameter can accurately reflect the monitoring value of each emission period. Based on the dynamic weight parameter and the monitoring resource capacity constraint, the periodic sampling sequence is generated, realizing the accurate matching of the sampling frequency and the period risk. In the high-weight period, high-frequency sampling is used to ensure the coverage of potential violation behaviors, and in the low-weight period, low-frequency sampling is used to reduce resource consumption, effectively solving the contradiction between resource waste and violation omission in the traditional fixed-frequency sampling. By adding the sewage outlet identifier, sampling period, and other metadata to the extracted video frame and constructing the structured monitoring frame sequence, the accurate association of the video frame with the emission period and the sewage outlet is ensured, avoiding the analysis deviation caused by the lack of data association in the subsequent intelligent recognition module. At the same time, the standardized data format reduces the complexity of data interaction between modules, improving the overall system processing efficiency. In addition, the evaluation logic that fuses the historical violation records enables the sampling strategy to inherit the past monitoring experience, reducing the adaptation period for new monitoring scenarios, and further improving the rationality of the sampling scheme and the preposition of violation recognition.

[0072] For details, please refer to Figure 3 , preferably, the specific steps of constructing the structured monitoring frame sequence include:

[0073] The historical violation records of the monitoring target are called through the data interface, and non-standardized information such as the violation period, violation type, and violation frequency in the historical violation records is converted into quantifiable violation feature data. For example, for the violation period, the time information in different record formats is converted into a standard timestamp through time format unification processing; for the violation type, qualitative types such as over-standard emission and illegal discharge are converted into quantified labels using label encoding technology; and for the violation frequency, the average violation frequency of each discharge outlet in different time periods is calculated through statistical analysis to form a historical violation feature vector of each discharge outlet, thereby providing standardized violation data support for subsequent weight evaluation.

[0074] Based on the sewage discharge sequence of each discharge outlet and the historical violation feature vector, a sampling weight evaluation model is constructed. The sampling weight evaluation model has three evaluation units built in: the emission stability evaluation unit first calculates the standard deviation of the expected emission amount in the sewage discharge period, then determines a preset deviation threshold based on the emission amount fluctuation mean of the monitoring target in the latest production cycle, calculates the deviation proportion, and inversely maps the deviation proportion to a stability score in the interval of 0 to 1; the historical violation risk degree evaluation unit calculates the similarity value of the multi-dimensional production feature matrix of the current sewage discharge period and the multi-dimensional production feature matrix of the historical violation period through a cosine similarity algorithm, and generates a risk degree score in combination with the ratio of the historical violation frequency of the sewage discharge period to the average violation frequency of the whole cycle, wherein the similarity value and the frequency ratio are summed after being weighted according to preset weights, and are mapped to a risk degree score in the interval of 0 to 1; the emission amount weight evaluation unit converts the expected sewage emission amount of each sewage discharge period into an emission amount score in the interval of 0 to 1 through a normalization algorithm. The scores of the three evaluation units are integrated according to a preset logic using a weighted fusion algorithm to generate a dynamic weight parameter for each sewage discharge period, wherein a period with poor stability, high risk degree, and high emission amount obtains a higher weight value.

[0075] The resource monitoring interface is called to obtain monitoring resource capacity data, including the sampling rate of each discharge outlet monitoring device, the upper limit of video frame storage, data transmission bandwidth, etc. The resource capacity data is converted into quantifiable sampling constraint indicators, such as the maximum number of samples per period and the upper limit of the total number of sampling frames per day. Based on the dynamic weight parameter and the sampling constraint indicator, a weight sampling frequency mapping table is used to assign a sampling frequency to each sewage discharge period of each discharge outlet, and according to the interval matching of the dynamic weight value, the sampling time point, sampling interval, and sampling frame number parameters of each sewage discharge period are integrated in the order of discharge outlet dimension and time sequence to generate a periodic sampling sequence containing the discharge outlet identifier, the sampling time point list, the sampling interval at each time point, and the sampling frame number at each time point.

[0076] The monitoring device of each sewage outlet is communicated through a video stream control interface, frame extraction instructions are sent to the corresponding device at the sampling time point in the periodic sampling sequence, the monitoring device receives the instructions, extracts the video frame at the specified time point and feeds back to the dynamic sampling module, and spatial and temporal alignment verification is performed, the device generation timestamp of the video frame is calculated with the target sampling timestamp of the sampling time point list in the periodic sampling sequence, if the time difference value exceeds the preset deviation threshold, the time difference value generates a correction coefficient, and the video frame timestamp is corrected; the video frame that passes the spatial and temporal alignment verification is added with multi-dimensional metadata information, including the sewage outlet identifier, the sampling time period, the dynamic weight parameter, and the sampling timestamp. The video frames with metadata are arranged in chronological order, and the frame data and metadata are integrated into a standardized data structure through data format conversion. The structured monitoring frame sequence of each sewage outlet is generated to provide analysis data for the intelligent recognition module to accurately associate the emission information.

[0077] The data preprocessing module takes the monitoring frame sequence output by the dynamic sampling module as input, extracts global features and performs similarity comparison on each video frame by calling the reference feature library, realizes accurate positioning of the real-time coordinates of the sewage outlet through perspective transformation correction and least squares fitting; according to the coordinates and the pre-stored relative coordinates of the standard view angle emission area, the camera intrinsic parameter and the shooting angle parameter are fused to construct a space projection matrix, and the affine transformation correction and the sewage diffusion characteristic expansion coefficient are combined to form a dynamic monitoring area covering the sewage outlet and the diffusion area; the dynamic foreground is separated by limiting the time span of the historical frame gray scale contrast, and the isolated pixel cluster is removed by combining the connected region analysis to purify the foreground area; for the purified foreground area, multi-modal emission judgment is adopted, each score threshold is set based on historical data and scene characteristics to identify emission frames; the emission frames are classified according to the sewage outlet identifier and sorted according to the timestamp, the emission events are divided through time continuity analysis, the key evaluation indicators of each frame are extracted and bound with the sewage outlet and event identifier, and integrated into a structured sewage outlet emission frame data set, realizing accurate, scene-based and structured processing throughout, providing standardized support for the intelligent recognition module, and being different from the traditional single-dimensional recognition and fixed area analysis mode.

[0078] In the embodiment, by calling the benchmark feature library, combining global feature extraction and perspective transformation correction, least square fitting, effectively avoiding the influence of monitoring device jitter, shooting angle deviation on the positioning of the sewage outlet, greatly improving the accuracy of the sewage outlet coordinates in each frame of video; Fusing camera intrinsic parameters, shooting angle parameters to construct a spatial projection matrix and cooperating with affine transformation correction and sewage diffusion expansion coefficient, the dynamic monitoring area can adapt to different shooting angles and sewage diffusion scenes, avoiding the problem of incomplete or redundant coverage of traditional fixed area; By limiting the time span of historical frame gray contrast and connected region analysis, the dynamic foreground is accurately separated and purified, and the static background and isolated interference pixels are excluded, ensuring that the foreground data is only associated with potential sewage discharge; Multi-modal discharge judgment combined with scenario-based threshold setting significantly reduces the risk of single feature misjudgment, improves the accuracy of discharge frame recognition; Finally, through time continuity analysis, the discharge event is divided and the structured data set is integrated, providing a clear time sequence, complete correlation and standardized data for the intelligent recognition module, effectively solving the problem of data fragmentation and poor adaptability in traditional processing, supporting the efficient development of subsequent illegal identification.

[0079] Please refer to Figure 4 , preferably, the specific steps of constructing the sewage outlet discharge frame data set include:

[0080] For each video frame of the monitoring frame sequence, the pre-stored sewage outlet benchmark feature library is called, which includes the physical outline data, surface texture features and feature marker points of the surrounding fixed reference objects of each sewage outlet. Global feature extraction is performed on the video frame, and the edge contour points, surface texture feature points of the sewage outlet and the feature marker points of the surrounding reference objects are extracted. The similarity of the extracted feature points and the feature marker points of the corresponding sewage outlet in the benchmark feature library is calculated, and the matching feature pairs with similarity higher than the preset similarity threshold are selected. According to the matching feature pairs, the perspective transformation matrix of the video frame and the standard view angle image in the benchmark feature library is calculated, the feature marker point coordinates of the extracted sewage outlet are corrected by perspective, and then the center position of the sewage outlet is fitted by least square method based on the corrected feature point coordinates, to locate the real-time coordinates of the sewage outlet in the current video frame.

[0081] According to the real-time coordinates of the sewage outlet and the pre-stored relative coordinate position of the discharge area, the relative coordinate is established based on the spatial position relationship of the sewage outlet and the discharge area under the standard shooting angle, the camera intrinsic calibration data of the monitoring device is called, including the lens focal length, the image sensor pixel size, and the principal point coordinates; and the shooting angle parameters of the current video frame are obtained, including the horizontal rotation angle, the vertical pitch angle, and the lens rotation angle, the mapping relationship of the three-dimensional space coordinates and the two-dimensional image coordinates under the standard view angle is established, the space projection matrix is established, the pre-stored relative coordinate of the discharge area under the standard view angle is input into the space projection matrix, and the relative coordinate is converted into the two-dimensional image coordinates of the current video frame through matrix multiplication operation, and the affine transformation algorithm is used to correct the shape of the converted image coordinates, so that the boundary of the mapped discharge area is consistent with the visual form under the current shooting angle. Based on the corrected discharge area boundary, the extension coefficient preset according to the sewage diffusion characteristics is combined to extend the discharge area boundary range, and the dynamic monitoring area covering the sewage outlet and the sewage diffusion area is formed.

[0082] The background is removed and the foreground is extracted for the delineated dynamic monitoring area, the monitoring evaluation range is set, which is used to limit the time span of the historical video frames, the historical video frame set in the monitoring evaluation range of the current video frame is extracted, and the dynamic monitoring area of the current video frame is compared with the corresponding dynamic monitoring area of the historical video frame set in terms of gray value, the gray value change difference of each pixel point in the time dimension is calculated, the gray value change threshold is set, the pixel points with the gray value change difference higher than the gray value change threshold are determined as dynamic foreground pixels, otherwise they are determined as static background pixels, and the preliminary separation of the dynamic foreground and the static background is realized.

[0083] For the preliminarily separated dynamic foreground pixels, the connectivity of the pixels is scanned by using the connected region analysis, all continuous foreground pixel clusters are marked, the area of each foreground pixel cluster is calculated, the isolated pixel clusters with an area smaller than the preset determination area threshold are removed, and the foreground pixel clusters with an area greater than or equal to the preset determination area threshold are reserved as the dynamic foreground area of the current video frame, so as to ensure that the extracted dynamic foreground is only related to the potential sewage discharge behavior.

[0084] For the extracted dynamic foreground region, the sewage discharge score is calculated through multi-modal discharge judgment, including color matching score, motion matching score and diffusion matching score. The multi-modal discharge judgment includes: performing color space transformation on the dynamic foreground region, extracting color features of the dynamic foreground region, including hue mean value, saturation mean value and brightness mean value, and comparing the color features with the pre-stored sewage typical color feature library to generate a color matching score. The sewage typical color feature library includes common dark, unusual and turbid color features of industrial sewage; the motion trajectory of the pixels in the dynamic foreground region is tracked by using an optical flow field calculation algorithm, the motion path of the same pixel in the dynamic foreground region in adjacent video frames is tracked, the proportion of interrupted pixels in the motion trajectory is calculated, the angle between the motion direction of all pixels in the dynamic foreground region and the preset discharge direction of the sewage discharge outlet is calculated, and the proportion of direction matching pixels with an angle less than the preset angle threshold is calculated; the motion matching score is generated by combining the proportion of interrupted pixels and the proportion of direction matching pixels and by using a preset weight to perform weighted summation; the boundary coordinates and the number of pixels of the dynamic foreground region in the preset number of previous video frames of the current video frame are extracted, the area of the circumscribed rectangle of the dynamic foreground region in each previous video frame is calculated through the boundary coordinates, the area of the dynamic foreground region is calibrated combined with the number of pixels to obtain the area of the dynamic foreground region in each previous video frame. The area difference between the current video frame and the previous video frame is calculated, and the area change rate of the dynamic foreground region is obtained by dividing the time interval; at the same time, the expansion distance of the dynamic foreground region boundary in each direction is calculated through boundary coordinate comparison, and the dynamic foreground region boundary expansion speed is obtained by dividing the time interval, and the diffusion matching score is generated by combining the dynamic foreground region area change rate and the dynamic foreground region boundary expansion speed through level mapping scoring, and the higher the score is, the more matched the diffusion characteristics of the dynamic foreground region and the diffusion characteristics of the sewage discharge are.

[0085] Based on historical sewage discharge monitoring data and environmental protection violation judgment standard, the qualified threshold of sewage discharge score is set, including color matching score qualified threshold, motion matching score qualified threshold, diffusion matching score qualified threshold, according to the sewage discharge score and the corresponding qualified threshold, the emission frame in the monitoring sequence is identified. And sort by sampling time stamp order, through time continuity analysis, calculate the time interval of adjacent emission frame, the emission frame with time interval less than the preset continuous threshold is classified into the same emission event, determine the time stamp of the starting frame and the ending frame of each emission event. And extract the key evaluation index of each emission frame, including the color feature parameter of the dynamic foreground area of the emission frame, that is, the average hue, the average saturation, the average brightness, the motion feature parameter, also including the track interruption pixel ratio, the direction matching pixel ratio, the diffusion feature parameter, also including the dynamic foreground area, the dynamic foreground area change rate, the dynamic foreground area boundary expansion speed, binding with the sewage outlet identifier and the emission event identifier, forming the emission frame data record, then integrating all the emission frame data records of the same sewage outlet according to the emission event, constructing the sewage outlet emission frame data set.

[0086] The intelligent identification module takes the sewage discharge sequence of each sewage outlet and the sewage outlet emission frame data set as the core input, extracts the emission frame subset according to the sewage outlet and the emission event identifier, retrieves the associated sewage discharge period information, sampling time stamp and key evaluation index and aggregates them according to the emission event, constructs the evaluation basic data; Perform time standardization processing on the diffusion feature parameter and the motion feature parameter to provide a unified benchmark for emission estimation; Calculate the sewage volume of a single emission frame based on the standardized key evaluation index and accumulate the total emission of the emission event to identify over-emission anomalies; Extract the foam shape parameter and the sewage transparency parameter of the dynamic foreground area of the emission frame, construct the compliance judgment vector after standardization combined with the color feature parameter; According to the standard judgment vector set by the environmental protection standard and the historical compliance sample, calculate the vector similarity by using the cosine similarity algorithm, identify abnormal violation frames combined with the threshold; Statistics the proportion of abnormal violation frames and mark them as color anomaly, shape anomaly and comprehensive anomaly according to the parameter deviation classification, statistics the proportion of each type; Finally, sort the abnormal frame proportion, the type of violation proportion and the core information of the emission event, generate a standardized violation identification report with parameter and proportion support, traceable and intuitive, provide judgment basis for environmental protection supervision.

[0087] In this embodiment, by aggregating multi-source data according to the pollution outlet and the emission event identification, the problem of data fragmentation is effectively avoided, ensuring the integrity and relevance of the evaluation base data, providing unified data support for subsequent judgment; the time standardization processing is performed on the diffusion characteristic and motion characteristic parameters, eliminating the parameter deviation caused by the sampling interval difference, significantly improving the accuracy of the emission estimate and the comparability of the data in different time periods; by calculating the total emission and comparing it with the expected emission, the precise identification of the excess emission anomaly is realized, avoiding the missed judgment of the emission dimension violation; the compliance judgment vector is constructed, combined with the standard judgment vector and the cosine similarity algorithm, which can efficiently identify abnormal violation frames, and at the same time, the violation type is classified and marked according to the parameter deviation, which fully covers the non-emission dimension violation scenarios such as color and shape, reducing the omission of implicit violations; the final standardized violation identification report has clear parameter and proportion support, which not only ensures the traceability of the judgment process, but also presents the core information intuitively, providing clear and reliable decision-making basis for environmental protection supervision, greatly improving the accuracy and efficiency of violation identification.

[0088] Please refer to Figure 5 , preferably, the specific steps of generating the violation identification report include:

[0089] According to the pollution outlet identification and the emission event identification, the emission frame subsets of each emission event are extracted from the pollution outlet emission frame data set; at the same time, the information set associated with the sewage discharge period in the sewage discharge sequence of the pollution outlet is called, including the expected sewage discharge of each sewage discharge period, the discharge start and end period, the associated production link, the time sequence label, and the sampling time stamp of each emission frame, the key evaluation indicators of each emission frame; according to the emission event dimension, the evaluation base data is constructed by aggregating, providing comprehensive and relevant base data for subsequent evaluation.

[0090] According to the sampling time stamp of each emission frame, the sampling time interval of adjacent emission frames in the emission frame subset is calculated, and the time standardization processing is performed on the diffusion characteristic parameters and motion characteristic parameters in the key evaluation indicators to eliminate the influence of sampling time interval difference: for example, for the dynamic foreground area change rate in the diffusion characteristic parameters, the unit time dynamic foreground area change rate is calculated by dividing the original dynamic foreground area change rate by the actual sampling time interval of adjacent frames, to ensure the comparability of the area change trend under different sampling intervals; for the sewage flow rate of the motion characteristic parameters, combined with the sampling time span of a single emission frame, the instantaneous flow rate is corrected to the average time flow rate by multiplying the proportion of a single emission frame sampling time, to ensure the consistency of the flow rate data under different sampling intervals.

[0091] Based on the time-standardized key evaluation indicators, the standardized average time flow rate and the dynamic foreground area are used to calculate the sewage volume in the sampling time of a single discharge frame by combining the sampling time length of a single discharge frame; the sewage volumes of all single discharge frames in the subset of discharge frames are added in the order of sampling time stamps to obtain the total discharge of the entire discharge event;

[0092] Based on the association mapping relationship between the discharge event and the sewage discharge period, i.e., by matching the sewage discharge period in the corresponding time interval of the sewage discharge sequence through the start and end time stamps of the discharge event, the expected sewage discharge volume of the discharge event is extracted from the sewage discharge sequence of the discharge outlet, the discharge difference between the total discharge volume of the discharge event and the corresponding expected sewage discharge volume is calculated, and if it is greater than a preset discharge threshold, it indicates that the actual discharge volume of the discharge event exceeds the compliance range estimated by the production end, and it is determined that the current discharge event has a potential illegal behavior, and the potential illegal behavior is marked as an excessive discharge anomaly;

[0093] The sewage feature parameters of the dynamic foreground area of all discharge frames in the subset of discharge frames of the discharge event are extracted, including foam morphology parameters and sewage transparency parameters, wherein the foam morphology parameters include foam contour number, foam average area and foam area ratio, wherein the foam contour number is obtained by first separating the foam area from the dynamic foreground area through threshold segmentation by image morphology algorithm, and then marking independent foam contours and counting the number by contour detection algorithm; the foam average area is obtained by dividing the total area of all detected foam contours by the number of foam contours; the foam area ratio is calculated by dividing the total area of the foam contours by the total area of the dynamic foreground area; the sewage transparency parameters include brightness mean derived transparency value and gray standard deviation transparency coefficient, wherein the brightness mean derived transparency value is derived from the brightness mean in the color feature parameters of the dynamic foreground area by linear mapping to convert the brightness mean to a transparency value in the interval of 0 to 1; the gray standard deviation transparency coefficient is calculated by calculating the standard deviation of the gray image of the dynamic foreground area, combining a preset standard deviation reference value, and calculating by subtracting the ratio of the actual calculated gray standard deviation and the preset standard deviation reference value from 1 and mapping to the interval of 0 to 1. The compliance judgment vector of each discharge frame is constructed after standardizing each sewage feature parameter;

[0094] The standard judgment vector is set according to the environmental compliance standard and the statistical data of historical compliance discharge samples, which is used to quantify the deviation degree of the discharge frame compliance judgment vector and the compliance state; the cosine similarity algorithm is used to calculate the similarity value between the compliance judgment vector of each discharge frame and the standard judgment vector, and a violation similarity threshold is set to identify abnormal violation frames;

[0095] If there is an abnormal violation frame in the emission frame subset of the emission event, the proportion of the abnormal violation frame in the emission frame subset is counted, and the abnormal violation frame is classified by combining the parameter deviation of each dimension of the compliance judgment vector of the abnormal violation frame, for example: for each abnormal violation frame, compare the parameters of each dimension of the compliance judgment vector with the preset compliance threshold: if the color matching score is lower than the qualified threshold of the color matching score, or the hue mean value and the saturation mean value exceed the compliance interval of the typical color feature library of sewage, mark the violation behavior type of the abnormal violation frame as color abnormality; if the number of foam profiles, the average area of foam, and the area proportion of foam are higher than the compliance threshold of foam morphology, or the derived transparency value of the mean brightness and the transparency coefficient of the gray standard deviation are lower than the compliance threshold of sewage transparency, mark as morphological abnormality; if the parameter deviation conditions of color abnormality and morphological abnormality are both met, mark as comprehensive abnormality. Further, the violation behavior type of the abnormal violation frame is marked, and the proportion of each type of abnormal violation frame in the emission frame subset is counted;

[0096] The abnormal violation frame proportion and violation behavior type proportion statistical results of each emission event are sorted, the violation behavior types include color abnormality, morphological abnormality, comprehensive abnormality and excessive emission abnormality, and the pollution outlet identifier, the emission event identifier, the emission start time stamp, the end time stamp, the total emission amount, the expected sewage emission amount, the emission difference, are combined; a standardized violation identification report is generated to ensure that each judgment result in the report corresponds to clear parameter comparison and proportion statistical support, and meet the needs of environmental protection supervision for traceable judgment process and intuitive results.

[0097] Working principle and effect:

[0098] Through the whole-process cooperation of data integration analysis, dynamic accurate sampling, frame data preprocessing and violation identification report generation, the present application realizes efficient identification and supervision of environmental protection violation behaviors.

[0099] By analyzing the heterogeneous production log, the key information of the sewage outlet discharge is accurately extracted by multi-modal AI, a time sequence characteristic matrix and a correlation learning model are constructed, and a structured sewage discharge sequence is generated. It not only solves the pain points of data dispersion and discharge correlation in traditional monitoring, but also provides high-precision basic data support for subsequent links, ensuring that the whole process analysis has reliable data support, and making the correlation between production and discharge more in line with the actual law, thus improving the data usability from the source; then, combined with the sewage discharge sequence and historical violation characteristics, a sampling weight evaluation model is constructed, the sampling resources are dynamically allocated according to the time risk, the periodic sampling sequence is generated, and the video frame sequence is constructed by extracting the video frame, which avoids the waste of fixed sampling resources or the problem of missing judgment of violations, significantly improves the sampling pertinence, and makes the monitoring data more in line with the subsequent analysis requirements, realizing the optimization balance of resource utilization and monitoring effectiveness; then, the preprocessing module accurately locates the sewage outlet and constructs a dynamic monitoring area, purifies the foreground through gray comparison and connected region analysis, accurately identifies the discharge frame by combining multi-modal judgment, and integrates it into a structured data set, breaking through the limitations of traditional fixed area analysis and single feature misjudgment, reducing invalid data interference, ensuring that the frame data is fully adapted to subsequent identification, and significantly improving the data preprocessing quality; finally, the intelligent identification module standardizes the frame data index, accurately calculates the discharge amount and compares and identifies the abnormal amount, constructs a compliance judgment vector to identify abnormal violation frames and classify and mark them, generates a traceable standardized report, realizes accurate differentiation of violation types and visualization of supervision basis, makes the violation judgment more convincing, and breaks through the data format barriers and static analysis limitations of traditional monitoring in practical application, significantly reduces the discharge amount estimation deviation, saves monitoring resources while ensuring high-risk period violation coverage, adapts to complex scenarios and improves the accuracy of discharge frame identification, reduces environmental interference misjudgment, accurately covers multiple types of violation behaviors, helps supervisors quickly grasp the details of violations, significantly improves supervision efficiency, and greatly reduces supervision difficulty and labor cost.

[0100] In summary, a complete closed loop from data acquisition to violation supervision is constructed, which not only solves the core problems of data integration difficulty, inaccurate sampling and one-sided violation identification in traditional environmental monitoring, but also improves the operability and accuracy of supervision through structured data and standardized reports, providing efficient and reliable technical support for environmental supervision, and helping to realize more scientific and intelligent environmental violation management.

[0101] The above only describes the preferred embodiments of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solutions falling within the scope of the present application should be considered within the protection scope of the present application. It should be noted that for ordinary technical personnel in the technical field, some improvements and decorations without departing from the principles of the present application should also be considered as the protection scope of the present application.

Claims

1. An environmental protection violation behavior intelligent identification system based on multi-modal AI, characterized in that, The data analysis module is used for receiving and analyzing production log data of a monitoring target, extracting multi-modal features of production parameters, constructing a time sequence feature matrix, establishing a dynamic mapping relationship between production parameters and sewage discharge, constructing a time sequence correlation learning model, and generating a sewage discharge sequence of each sewage outlet. The dynamic sampling module is used for constructing a sampling weight evaluation model based on the sewage discharge sequence of each sewage outlet and a historical illegal feature vector, generating a dynamic weight parameter of each sewage discharge period, constructing a periodic sampling sequence, extracting video frames, and constructing a monitoring frame sequence. The data preprocessing module is used for positioning a sewage outlet of each video frame in the monitoring frame sequence, obtaining a dynamic monitoring area of each video frame through a spatial projection matrix, dividing a dynamic foreground area through foreground extraction, identifying an emission frame through multi-modal emission judgment, and constructing a sewage outlet emission frame data set. The intelligent identification module is used for performing time standardization processing on key evaluation indexes of the emission frame, identifying over-discharge anomalies, extracting sewage feature parameters of the dynamic foreground area, identifying abnormal illegal frames and marking illegal behavior types, and generating an illegal identification report. The step of obtaining the dynamic monitoring area of each video frame includes:

2. The multi-modal AI-based environmental protection violation behavior intelligent identification system of claim 1, wherein, feature extraction is performed on each video frame of the monitoring frame sequence, feature marker points of the sewage outlet are obtained, similarity calculation is performed on the feature marker points and feature marker points in a reference feature library of the sewage outlet, and matching feature pairs are screened; a perspective transformation matrix of the video frame and a standard perspective image in the reference feature library is calculated according to the matching feature pairs, perspective correction is performed on the feature marker point coordinates of the sewage outlet, and the center position of the sewage outlet is fitted as a real-time coordinate of the sewage outlet in the current video frame; camera intrinsic calibration data of the monitoring device and shooting angle parameters of the current video frame are called, a spatial projection matrix of a mapping relationship between three-dimensional space coordinates in a standard perspective and two-dimensional image coordinates of the current video frame is established; the two-dimensional image coordinates of the emission area in the current video frame are calculated according to the spatial projection matrix, combined with the real-time coordinates of the sewage outlet and the relative coordinate position of the emission area, to generate an emission area boundary; and an expansion coefficient is preset to expand the emission area boundary range to form a dynamic monitoring area of the monitoring frame. The step of constructing the sewage outlet emission frame data set includes: 3.The multi-modal AI-based environmental rule violation behavior intelligent identification system of claim 1, wherein, a monitoring evaluation range is set, and a historical video frame set in the monitoring evaluation range of the current video frame is extracted; a gray value comparison is performed on the dynamic monitoring area of the current video frame and the corresponding dynamic monitoring area of the historical video frame set, and a gray value change difference value of each pixel point in the time dimension is calculated; a gray value change threshold is set, dynamic foreground pixels are identified, and the dynamic foreground pixels are marked as foreground pixel clusters through connected region analysis, the area of each foreground pixel cluster is calculated, and the dynamic foreground area of the current video frame is screened out; for the extracted dynamic foreground area, a sewage discharge score is calculated through multi-modal emission judgment, and the sewage discharge score includes a color matching score, a motion matching score, and a diffusion matching score; ​ Setting a qualified threshold value of the sewage discharge score respectively, identifying the discharge frames in the monitoring sequence, sorting the discharge frames in the order of sampling time stamps, calculating the time interval of adjacent discharge frames to divide the discharge events, and marking the time stamps of the starting frame and the ending frame of each discharge event; Extracting the key evaluation indexes of each discharge frame, binding with the sewage outlet identifier and the discharge event identifier, and constructing the sewage outlet discharge frame data set. 4.The multi-modal AI-based environmental rule violation behavior intelligent identification system of claim 3, wherein, The multi-modal discharge determination comprises: Performing color space transformation on the dynamic foreground region, extracting the color features of the dynamic foreground region, and performing similarity comparison with the pre-stored typical sewage color feature library to generate a color matching score; Tracking the motion trajectory of the pixels in the dynamic foreground region, calculating the proportion of the motion trajectory interrupted pixels, and calculating the angle between the motion direction of the pixels in the dynamic foreground region and the preset discharge direction of the sewage outlet, and calculating the proportion of the direction matching pixels with an angle less than a preset angle threshold; combining the proportion of the trajectory interrupted pixels and the proportion of the direction matching pixels, and generating a motion matching score through weighted summation of the preset weight; Retrieving the pre-sequence video frame set of the current video frame, extracting the boundary coordinates and the number of pixels of the dynamic foreground region in each pre-sequence video frame, calculating the area of the circumscribed rectangle of the dynamic foreground region in each pre-sequence video frame to obtain the area of the dynamic foreground region of each pre-sequence video frame; calculating the area difference of the dynamic foreground region of the current video frame and the pre-sequence video frame, and dividing by the time interval to obtain the area change rate of the dynamic foreground region; and calculating the expansion distance of the dynamic foreground region boundary in each direction, dividing by the time interval to obtain the expansion speed of the dynamic foreground region boundary, and combining the area change rate of the dynamic foreground region and the expansion speed of the dynamic foreground region boundary to generate a diffusion matching score through grade mapping scoring. 5.The multi-modal AI-based environmental rule violation behavior intelligent identification system of claim 1, wherein, The step of identifying the over-discharge anomaly comprises: According to the sewage outlet identifier and the discharge event identifier, extracting the discharge frame subset of each discharge event from the sewage outlet discharge frame data set; According to the sampling time stamp of each discharge frame, calculating the sampling time interval of adjacent discharge frames in the discharge frame subset, and performing time standardization processing on the diffusion feature parameters and the motion feature parameters in the key evaluation indexes; Based on the time-standardized key evaluation indexes, using the standardized average time flow rate and the area of the dynamic foreground region, and combining the sampling duration of a single discharge frame, the sewage volume in the sampling time of a single discharge frame is calculated; according to the sequence of the sampling time stamps, the sewage volumes of all single discharge frames in the discharge frame subset are added to obtain the total discharge of the entire discharge event; Based on the association mapping relationship between the discharge event and the sewage discharge period, the expected sewage discharge volume of the discharge event is extracted from the sewage discharge sequence, the discharge difference between the total discharge of the discharge event and the corresponding expected sewage discharge volume is calculated, and if it is greater than a preset discharge threshold, it is determined that the current discharge event has a potential illegal behavior, and is marked as an over-discharge anomaly. 6.The multi-modal AI-based environmental rule violation behavior intelligent identification system of claim 5, wherein, The step of identifying the abnormal violation frame and marking the type of violation behavior comprises: Extracting the sewage feature parameters of the dynamic foreground region of the discharge frame in the discharge frame subset of the discharge event, including the foam morphology parameter and the sewage transparency parameter, and constructing the compliance determination vector of each discharge frame after standardization; A standard judgment vector is set to quantify the deviation degree of the compliance judgment vector of the emission frame from the compliance state, a similarity value between the compliance judgment vector of each emission frame and the standard judgment vector is calculated by using a cosine similarity algorithm, a violation similarity threshold is set, and an abnormal violation frame is identified; If there is an abnormal violation frame in the emission frame subset of the emission event, the proportion of the abnormal violation frame in the emission frame subset is counted, and the abnormal violation frame is classified in combination with the parameter deviation of each dimension of the abnormal violation frame compliance judgment vector, so as to mark the violation behavior type of the abnormal violation frame, and count the proportion of each type of abnormal violation frame in the emission frame subset.

7. The multi-modal AI-based environmental rule violation behavior intelligent identification system of claim 1, wherein The step of constructing the time sequence feature matrix comprises: Receiving production log data, and performing data format division on the production log data according to data format characteristics to perform data preprocessing on the production log data; Positioning the core elements related to sewage discharge in the production log data, and aggregating the associated information in different data formats through semantic fusion to form a sewage discharge information set of each sewage outlet; Based on the sewage discharge information set and the production log data, a plurality of modal features reflecting the production state are extracted with the production parameters as the core; the plurality of modal features include quantitative features, derived features and qualitative features; The plurality of modal features are converted into a unified data dimension and sorted in time sequence to construct a time sequence feature matrix of each sewage outlet. 8.The multi-modal AI-based environmental rule violation behavior intelligent identification system of claim 7, wherein, The step of generating the sewage discharge sequence of each sewage outlet comprises: Based on the time sequence feature matrix of each sewage outlet and the corresponding historical emission data, a time sequence correlation learning model is constructed to establish a dynamic mapping relationship between the production parameters and the sewage emission, and to predict the sewage emission of each production period; Dividing each time unit of the sewage discharge monitoring period, and extracting the sewage discharge rule from the historical emission data to set the sewage discharge period in each time unit; Based on the set sewage discharge period, the predicted sewage emission of each sewage discharge period is calculated by calling the time sequence correlation learning model, and the emission time length, the production link identifier and the time sequence identifier are associated to generate a structured sewage discharge sequence of each sewage outlet. 9.The multi-modal AI-based environmental rule violation behavior intelligent identification system of claim 1, wherein, The step of generating the dynamic weight parameter of each sewage discharge period comprises: The historical violation record of the monitoring target is called, and the historical violation record is converted into quantifiable violation feature data to form a historical violation feature vector of each sewage outlet; Based on the sewage discharge sequence of each sewage outlet and the historical violation feature vector, a sampling weight evaluation model is constructed to calculate a stability score, a risk score and an emission weight score, and a dynamic weight parameter of each sewage discharge period is generated by weighted fusion; The monitoring resource capacity data is obtained and converted into quantifiable sampling constraint indicators, and based on the dynamic weight parameter and the sampling constraint indicators, a sampling frequency is allocated to each sewage discharge period of each sewage outlet through a weight sampling frequency mapping table to generate a periodic sampling sequence; Frame extraction instructions are sent to the corresponding equipment at the sampling time points in the periodic sampling sequence to extract a monitoring frame sequence of each sewage outlet. 10.The multi-modal AI-based environmental rule violation behavior intelligent identification system of claim 1, wherein, The sampling weight evaluation model comprises an emission stability evaluation unit, a historical violation risk degree evaluation unit and an emission weight evaluation unit; The discharge stability evaluation unit is configured to calculate a standard deviation of the predicted discharge amount in the sewage discharge period, and calculate a deviation proportion with respect to a preset deviation threshold, and map the deviation proportion to a stability score in the interval of 0 to 1; The historical violation risk degree evaluation unit is configured to calculate a similarity value between a multi-dimensional production feature matrix of the current sewage discharge period and a multi-dimensional production feature matrix of a historical violation period by using a cosine similarity algorithm, and generate a risk degree score in combination with a ratio of a historical violation frequency of the sewage discharge period to an average violation frequency in a whole cycle. The discharge amount weight evaluation unit is configured to convert the predicted sewage discharge amount of each sewage discharge period into a discharge amount score in the interval of 0 to 1 by using a normalization algorithm.

Images