Industrial water quality remote measurement method and terminal

Abstract

The present application relates to the technical field of water quality monitoring, in particular to an industrial water quality remote monitoring method and terminal, comprising obtaining multi-modal water quality data in a target monitoring area, performing canonical correlation analysis on the multi-modal water quality data to obtain a projection matrix of the multi-modal water quality data; mapping the multi-modal water quality data to a common subspace according to the projection matrix to obtain a fusion feature vector; pre-training the fusion feature vector, obtaining a dynamic change value through time series prediction, and outputting target water quality data based on the dynamic change value, effectively solving the problem that the existing industrial water quality remote monitoring architecture cannot provide a high-reliability evaluation basis for water quality dynamic changes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water quality monitoring, in particular to an industrial water quality remote monitoring method and terminal. BACKGROUND

[0002] Industrial water quality remote monitoring is an important means to realize real-time and continuous monitoring of water quality in industrial environment, and has important significance for protecting water environment safety and early warning of pollution events.

[0003] At present, the water quality remote monitoring for industrial environment usually adopts the method of deploying multi-parameter sensing devices to synchronously collect pH value, conductivity, dissolved oxygen and biochemical oxygen demand and other water quality data, and based on remote transmission technology, the water quality data is directly sent to the backend device, and then the backend device needs to fuse and analyze these water quality data with different characteristics to form a unified water quality state evaluation.

[0004] However, due to the inherent differences in working principle, response characteristics and data format of different parameter sensing devices, there is a serious heterogeneity problem between the water quality data received by the backend system, which makes it difficult to effectively mine the internal correlation between these water quality data and form an accurate water quality state representation, resulting in a decline in the final analysis accuracy and inability to provide a high-reliability evaluation basis for water quality dynamic changes. SUMMARY

[0005] In order to solve the technical problem that the existing industrial water quality remote monitoring architecture cannot provide a high-reliability evaluation basis for water quality dynamic changes, the present application provides an industrial water quality remote monitoring method and terminal.

[0006] The industrial water quality remote monitoring method and terminal provided by the present application adopt the following technical solutions:

[0007] An industrial water quality remote monitoring method, comprising:

[0008] Obtaining multi-modal water quality data in a target monitoring area, performing canonical correlation analysis on the multi-modal water quality data to obtain a projection matrix of the multi-modal water quality data;

[0009] Mapping the multi-modal water quality data to a common subspace according to the projection matrix to obtain a fusion feature vector;

[0010] Pre-training the fusion feature vector, obtaining a dynamic change value through time series prediction, and outputting target water quality data based on the dynamic change value.

[0011] Further, the step of obtaining multi-modal water quality data in a target monitoring area comprises:

[0012] According to a preset collection time period, obtaining an initial water quality data sequence corresponding to each target parameter through a multi-parameter sensing device deployed in the target monitoring area;

[0013] Based on the initial water quality data sequence, the spatio-temporal alignment error between each target parameter is calculated;

[0014] According to the spatio-temporal alignment error, the initial water quality data sequence is dynamically time-warped and calibrated to generate multi-modal water quality data.

[0015] Further, the step of performing canonical correlation analysis on the multi-modal water quality data to obtain a projection matrix of the multi-modal water quality data includes:

[0016] A real-time observation data matrix containing each target parameter is constructed through a sliding time window mechanism;

[0017] Canonical correlation analysis is performed on the multi-modal water quality data to obtain a benchmark projection matrix;

[0018] Canonical correlation analysis is performed on the real-time observation data matrix and the benchmark projection matrix to obtain a projection matrix.

[0019] Further, the step of mapping the multi-modal water quality data to a common subspace according to the projection matrix to obtain a fusion feature vector includes:

[0020] The multi-modal water quality data is mapped to a common subspace according to the projection matrix to generate an initial fusion feature vector;

[0021] An environmental interference factor analysis is performed on the initial fusion feature vector to calculate a real-time signal-to-noise ratio compensation coefficient;

[0022] The initial fusion feature vector is optimized by the real-time signal-to-noise ratio compensation coefficient to obtain a fusion feature vector.

[0023] Further, the step of pre-training the fusion feature vector to obtain a dynamic change value through time series prediction includes:

[0024] Based on the fusion feature vector, time series prediction is performed on the multi-modal water quality data through pre-training to obtain an initial prediction sequence;

[0025] A water quality parameter measurement sequence is obtained, and a prediction bias between the initial prediction sequence and the water quality parameter measurement sequence is calculated through residual analysis;

[0026] The initial prediction sequence is compensated according to the prediction bias to obtain a dynamic change value.

[0027] Further, the step of outputting target water quality data based on the dynamic change value includes:

[0028] The dynamic change value is confidence evaluated to calculate a real-time confidence;

[0029] The real-time confidence is combined with a multi-level early warning threshold to make a state decision, and a water quality early warning level is output.

[0030] The dynamic change value is packaged according to the water quality early warning level, and target water quality data is output.

[0031] Further, before the step of combining the real-time confidence with the multi-level early warning threshold to make a state decision and output a water quality early warning level, the method further includes:

[0032] The historical water quality data is obtained, the periodic change rule of the historical water quality data is analyzed through a time series clustering algorithm, and a threshold cluster is obtained.

[0033] The redundant threshold points in the threshold cluster are removed to obtain an initial multi-level early warning threshold.

[0034] The initial multi-level early warning threshold is calibrated in real time according to the water quality data fluctuation characteristics on the preset time collection section through a sliding window mechanism to generate a multi-level early warning threshold.

[0035] The application provides an industrial-grade water quality remote sensing terminal, which comprises:

[0036] A data acquisition module is configured to obtain multi-modal water quality data in a target monitoring area, perform a canonical correlation analysis operation on the multi-modal water quality data, and obtain a projection matrix of the multi-modal water quality data.

[0037] A feature fusion module is configured to map the multi-modal water quality data to a common subspace according to the projection matrix and obtain a fusion feature vector.

[0038] A prediction output module is configured to pre-train the fusion feature vector, obtain a dynamic change value through time series prediction, and output target water quality data based on the dynamic change value.

[0039] The application has the following beneficial effects:

[0040] The application provides an industrial-grade water quality remote sensing method, which comprises the following steps: obtaining multi-modal water quality data in a target monitoring area, performing a canonical correlation analysis operation on the multi-modal water quality data, obtaining a projection matrix of the multi-modal water quality data, mapping the multi-modal water quality data to a common subspace according to the projection matrix, obtaining a fusion feature vector, pre-training the fusion feature vector, obtaining a dynamic change value through time series prediction, and outputting target water quality data based on the dynamic change value.

[0041] That is, in the present application, by directly performing a canonical correlation analysis operation on the acquired multi-modal water quality data at the front end, the problem of decreased analysis accuracy caused by data heterogeneity when processing centrally at the back end is effectively solved; the canonical correlation analysis can mine the internal correlation between multi-modal water quality data with different characteristics, map the multi-modal water quality data to a common subspace through the projection matrix obtained by calculation, thereby eliminating the data heterogeneity between multi-parameter sensing devices and generating a fusion feature vector that uniformly represents the water quality state; the fusion feature vector retains the key correlation information of the multi-modal water quality data, providing a highly consistent input basis for subsequent analysis, and then through pre-training of the fusion feature vector, the system can output accurate target water quality data, thereby significantly improving the reliability of water quality dynamic change evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is the overall flowchart of the industrial water quality telemetry method of the present application;

[0043] Figure 2 is the step flowchart of step S10 in the industrial water quality telemetry method of the present application;

[0044] Figure 3 is the step flowchart of step S20 in the industrial water quality telemetry method of the present application;

[0045] Figure 4 is the step flowchart of step S30 in the industrial water quality telemetry method of the present application;

[0046] Figure 5 is the module diagram of the industrial water quality telemetry terminal of the present application.

[0047] BRIEF DESCRIPTION OF DRAWINGS

[0048] 10, data acquisition module; 20, feature fusion module; 30, prediction output module; 40, early warning threshold generation module. DETAILED DESCRIPTION

[0049] The following will be described in detail in combination with the accompanying Figures 1-5 The present application will be further described in detail.

[0050] In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, the terms "first", "second", "third" are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0051] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0052] The embodiment of the present application discloses an industrial water quality remote measurement method and terminal.

[0053] Please refer to Figure 1 The embodiment proposes an industrial water quality remote measurement method, which includes steps S10-S30:

[0054] Step S10, obtaining multi-modal water quality data in the target monitoring area, performing a typical correlation analysis operation on the multi-modal water quality data to obtain a projection matrix of the multi-modal water quality data.

[0055] It should be noted that compared with the conventional industrial water quality remote measurement scheme, the embodiment proposes to process the multi-modal water quality data at the front end of the water quality remote measurement, so as to maximize the preservation and utilization of the spatial and temporal consistency and environmental context when the multi-modal water quality data is generated.

[0056] The conventional industrial water quality remote measurement scheme is to collect the water quality data corresponding to each target parameter at the front end, and then perform encoding, transmission, decoding and other operations to transmit to the back end, and then output the target water quality data on the back end. However, because various sensing devices are deployed in the industrial field, the data formats, response characteristics and sampling frequencies of these sensing devices are different, and at the same time, the network delay and jitter existing in the transmission process will cause the water quality data to arrive at the back end out of synchronization, destroying its simultaneity, so that the finally calculated target water quality data has deviation, and cannot provide a high reliability evaluation basis for water quality dynamic changes.

[0057] Therefore based on the above problems, the embodiment proposes to directly process the water quality data at the front end after collecting the water quality data corresponding to each target parameter, because there is no network delay, jitter and other abnormal situations in the transmission process, so that the calculation of the target water quality data can be truly synchronized, thereby effectively eliminating the calculation deviation.

[0058] Specifically, after obtaining the multi-modal water quality data in the target monitoring area in this step, the statistical correlation law between the multi-modal water quality data such as pH value, conductivity, dissolved oxygen, and biochemical oxygen demand is deeply mined through typical correlation analysis operation, and the projection matrix capable of eliminating the heterogeneity between multi-parameter sensing devices is calculated accordingly. This processing can map the multi-modal water quality data with different characteristics and difficult to directly fuse into a common subspace, laying a foundation for subsequent generation of fusion feature vectors with consistent representation ability.

[0059] It should be noted that the multi-modal water quality data refers to a set of multi-source sensor data without fusion.

[0060] Step S20, according to the projection matrix, the multi-modal water quality data is mapped to a common subspace to obtain a fusion feature vector.

[0061] Using the conversion characteristics of the projection matrix, the multi-modal water quality data originally scattered in different characteristic spaces is mapped to a common subspace that can reflect the maximum correlation between target parameters, thereby eliminating the influence of the difference characteristics between multi-parameter sensing devices on data consistency. In this way, the heterogeneous multi-modal water quality data is effectively converted into a fusion feature vector with uniform scale and key correlation information retained, providing a consistent representation data basis for subsequent water quality analysis.

[0062] It should be noted that the target parameter refers to pH value, conductivity, dissolved oxygen, biochemical oxygen demand, etc.

[0063] Step S30, pre-training the fusion feature vector, obtaining the dynamic change value through time series prediction, and outputting the target water quality data based on the dynamic change value.

[0064] The fusion feature vector after fusion processing is used to realize water quality state prediction and decision making. The role is to learn the water quality law reflected by the fusion feature vector through pre-training, and capture the dynamic change trend of water quality data by means of time series prediction, so as to obtain the dynamic change value representing the future water quality state. This step can convert the fusion feature vector into a dynamic change value with time foresight, break through the limitation of conventional monitoring that can only reflect the current water quality state, and finally output target water quality data containing prediction results and decision information based on the dynamic change value, realize the leap from passive monitoring to active early warning, provide high reliability trend evaluation and decision basis for water quality safety management, and significantly improve the reliability of water quality dynamic change evaluation.

[0065] It should be noted that the water quality data refers to the numerical values corresponding to the target parameters such as pH value, conductivity, dissolved oxygen, biochemical oxygen demand, etc.

[0066] Specifically, referring to FIG. 1, in a feasible implementation, step S10 can include steps S11-S16: Figure 2

[0067] Step S11, according to the preset acquisition time period, the initial water quality data sequence corresponding to each target parameter is obtained by the multi-parameter sensing device deployed in the target monitoring area.

[0068] By presetting the preset acquisition time period in the multi-parameter sensing device, the water quality data acquisition process of all sensing devices is started synchronously at the same acquisition time point, ensuring that the water quality data of each target parameter such as pH value, conductivity, dissolved oxygen and biochemical oxygen demand has strict time synchronization.

[0069] Each sensing device deployed in the target monitoring area continuously acquires the original signal according to the unified time reference, and converts the analog signal into digital water quality data value through the built-in analog-to-digital conversion module. Finally, the initial water quality data sequence corresponding to the water quality data is generated in time sequence, which can fundamentally ensure that the time stamps of the water quality data are completely aligned, providing a data basis with strict time sequence consistency for subsequent calculation of space-time alignment error, thereby effectively solving the problem of data time sequence disorder caused by the difference in response speed of sensing devices.

[0070] It should be noted that the preset acquisition time period is the time period for controlling the sensing device to acquire water quality data; the initial water quality data sequence is composed of parameter sequences corresponding to each target parameter, such as pH value sequence, conductivity sequence, dissolved oxygen sequence and biochemical oxygen demand sequence, etc.

[0071] Step S12, based on the initial water quality data sequence, the space-time alignment error between each target parameter is calculated.

[0072] ​From the initial water quality data sequence, select one as the reference parameter sequence, for example, the conductivity sequence, match the parameter sequences corresponding to other target parameters with the reference parameter sequence, find the minimum cumulative distance between the parameter sequences by finding the minimum cumulative distance between the parameter sequences, calculate the time offset and shape distortion of each parameter sequence relative to the reference parameter sequence, and jointly form the space-time alignment error, to accurately quantify the time sequence asynchronization and waveform distortion caused by the inherent characteristics of different sensing devices, such as response delay and sampling jitter, and provide accurate correction basis for subsequent data calibration, thereby effectively eliminating the space-time heterogeneity of multi-modal water quality data.

[0073] Step S13, dynamically time warping and calibrating the initial water quality data sequence according to the space-time alignment error, generating multi-modal water quality data.

[0074] According to the time offset in the space-time alignment error, perform nonlinear stretching transformation on each parameter sequence, align the time axis of the reference sequence with the time axis of other parameter sequences, and simultaneously compensate and correct the waveform amplitude according to the shape distortion, eliminate the phase deviation and amplitude distortion caused by the difference in response characteristics of sensing devices, then use a linear calibration model based on least squares method, take the reference sequence as a reference, and correct the amplitude value of the aligned other parameter sequences respectively, generate time-synchronized, amplitude-accurate multi-modal water quality data, which can effectively eliminate the space-time representation difference between multi-parameter sensing devices and realize data isomorphism, providing multi-modal water quality data with space-time consistency for subsequent canonical correlation analysis operation.

[0075] Step S14, construct a real-time observation data matrix containing each target parameter through a sliding time window mechanism.

[0076] By setting a sliding time window of fixed length, such as containing the latest 60 sampling points, taking the latest collection time as a reference to intercept the time series data of each water quality data within the sliding time window, these time-ordered parameter sequences are used as column vectors, and a real-time observation data matrix is constructed by combining data types, wherein each row can represent a sampling time, and each column can represent all observation values of a water quality data. The dynamic streaming data is converted into a structured matrix representation, which not only retains the time sequence correlation characteristics between multi-target parameters, but also eliminates random fluctuations during data collection, providing data input that can reflect water quality state changes for subsequent canonical correlation analysis operations.

[0077] Step S15, performing canonical correlation analysis operation on the multi-modal water quality data to obtain a reference projection matrix.

[0078] Multimodal water quality data are divided into two sets of variables based on parameter type. For example, pH and conductivity are used as the first set of variables, and dissolved oxygen and biochemical oxygen demand are used as the second set. By calculating the covariance matrix between the two sets of variables and the cross-set covariance matrix, the characteristic equation is solved to obtain the eigenvector corresponding to the largest eigenvalue. This eigenvector constitutes the initial projection matrix. Then, the initial projection matrix is ​​orthogonalized using the singular value decomposition algorithm to finally obtain the reference projection matrix that satisfies the orthogonality constraint conditions.

[0079] The above operations can extract the most stable statistical correlation patterns among water quality data, establish a benchmark projection matrix to eliminate inherent differences in sensing devices, and provide data support for subsequent mining of essential correlations among multi-source water quality data in the feature space.

[0080] Step S16: Perform canonical correlation analysis on the real-time observation data matrix and the reference projection matrix to obtain the projection matrix.

[0081] The real-time observation data matrix is ​​grouped according to the target parameters to maintain structural consistency with the reference projection matrix. Then, using the reference projection matrix as a constraint, the covariance matrix of the real-time observation data matrix and its cross-group covariance matrix with the reference projection matrix are calculated. By solving the generalized characteristic equation, the feature vector set reflecting the current data characteristics is obtained. Then, the feature vector set is linearly weighted and fused with the reference projection matrix. Finally, the fused matrix is ​​orthogonalized using the singular value decomposition algorithm to obtain the projection matrix.

[0082] It can achieve high-precision spatiotemporal alignment and deep fusion of multi-source water quality data, providing an accurate and consistent data foundation for subsequent water quality status assessment, and effectively improving the reliability and prediction accuracy of water quality telemetry.

[0083] Specifically, refer to Figure 3 As shown, in one feasible implementation, step S20 may include steps S21 to S23:

[0084] Step S21: Based on the projection matrix, the multimodal water quality data is mapped to a common subspace to generate an initial fusion feature vector.

[0085] Multimodal water quality data is used as the input matrix and matrix multiplication is performed with the projection matrix. Through linear transformation, the water quality data are transformed from their independent feature spaces to a common subspace with the greatest correlation. In this process, each weight coefficient of the projection matrix corresponds to the contribution of different water quality data in the fusion process. By weighted combination, the dimensional differences and scale fluctuations between different water quality data are eliminated. Finally, an initial fusion feature vector with unified dimension and consistent representation is output, thus solving the heterogeneity problem of different water quality data.

[0086] Step S22: Perform environmental interference factor analysis on the initial fused feature vector and calculate the real-time signal-to-noise ratio compensation coefficient.

[0087] A baseline eigenvalue distribution model of the initial fused feature vector under a standard environment is established. Then, the Mahalanobis distance between the eigenvalues ​​of each dimension of the initial fused feature vector and the baseline model is calculated in real time. This distance value quantifies the degree of feature shift caused by environmental interference. Simultaneously, based on a sliding window mechanism, short-time fluctuation components of the initial fused feature vector are extracted. These short-time fluctuation components are decomposed into intrinsic mode functions of different frequencies using the Hilbert-Huang transform. High-frequency mode components that match the frequency characteristics of environmental interference are selected, and their high-frequency energy proportion is calculated. Finally, the feature shift degree and the high-frequency energy proportion are weighted and fused to generate a real-time signal-to-noise ratio (SNR) compensation coefficient. This real-time SNR compensation coefficient can accurately characterize the distortion degree caused by the current environmental interference to the initial fused feature vector, providing a precise compensation basis for subsequent optimization of the initial fused feature vector.

[0088] Step S23: Optimize the initial fused feature vector using the real-time signal-to-noise ratio compensation coefficient to obtain the fused feature vector.

[0089] The real-time signal-to-noise ratio compensation coefficient is used as the weight index of the initial fused feature vector. By establishing the mapping relationship between the real-time signal-to-noise ratio compensation coefficient and the initial fused feature vector, the components of each dimension in the initial fused feature vector are adaptively weighted. When the real-time signal-to-noise ratio compensation coefficient shows that a certain dimension is severely disturbed, that is, the signal-to-noise ratio is low, its weight is reduced. When the disturbance is slight, that is, the signal-to-noise ratio is high, its weight is increased.

[0090] Simultaneously, based on the inverse projection reconstruction algorithm, the weighted components of each dimension are projected onto the noise subspace and the signal subspace. The main water quality features are retained in the signal subspace, while frequency components related to environmental interference are filtered out in the noise subspace. Finally, the optimized fusion feature vector is reconstructed. This operation can dynamically adjust the contribution rate of each dimension component in the initial fusion feature vector according to the real-time environmental interference level, effectively suppressing feature distortion caused by environmental factors such as temperature fluctuations and water turbidity. This makes the output fusion feature vector more essentially represent changes in water quality status, providing a data foundation with strong anti-interference ability and high reliability for subsequent time series prediction.

[0091] Specifically, refer to Figure 4 As shown, in one feasible implementation, step S30 may include steps S31 to S36:

[0092] Step S31: Based on the fused feature vector, perform time series prediction on multimodal water quality data through pre-training to obtain the initial prediction sequence.

[0093] On the historical data set composed of the multi-modal water quality data accumulated by the historical monitoring period, calibrated by the spatio-temporal alignment and the corresponding fusion feature vectors, the LSTM model is trained to learn the nonlinear mapping relationship between the fusion feature vectors and the multi-modal water quality data. The LSTM model captures the time-dependent characteristics of the multi-modal water quality data through its internal gating mechanism and cell state.

[0094] In the actual prediction stage, the fusion feature vector sequence generated in real time is input into the LSTM model in time sequence, and the LSTM model outputs the multi-modal water quality data prediction value of the future multiple time steps in turn through forward calculation, thereby forming an initial prediction sequence.

[0095] By utilizing the multi-parameter collaborative change information contained in the fusion feature vector, the dynamic law of water quality change is extracted through the time series modeling capability of the deep learning model, i.e., the LSTM model, and finally the future change trend of multiple water quality data such as pH value, conductivity, dissolved oxygen and biochemical oxygen demand is accurately predicted, providing a time window ahead of the actual change for water quality warning.

[0096] Step S32, obtaining the water quality parameter measured sequence, calculating the prediction deviation between the initial prediction sequence and the water quality parameter measured sequence through residual analysis.

[0097] After the generation of the initial prediction sequence is completed through step S31, the water quality parameter measured sequence with the same environmental factors such as water temperature, pH basic value, and the same meteorological conditions such as air temperature and rainfall as the current preset time collection segment is extracted from the historical database as a verification benchmark to ensure data comparability.

[0098] The numerical difference between the initial prediction sequence and the water quality parameter measured sequence is calculated by the point-by-point difference algorithm, and after the residual sequence is formed, the root mean square error algorithm is used for statistical analysis to quantify the prediction deviation between the prediction sequence and the measured sequence. The prediction accuracy can be verified in different time dimensions with similar water quality change rules, and objective prediction deviation is provided for subsequent dynamic change value compensation correction, thereby significantly improving the reliability of the prediction result in actual application.

[0099] Step S33, compensating the initial prediction sequence according to the prediction deviation to obtain the dynamic change value.

[0100] The least square method is used to fit the prediction deviation as an error function changing with time, the inherent deviation rule existing is quantified through the error function, a corresponding compensation sequence is generated based on the error function, the compensation sequence and the initial prediction sequence are superimposed point by point to obtain the dynamic change value, the initial prediction sequence is corrected to eliminate systematic errors caused by environmental changes or equipment aging, and the dynamic change value obtained after correction is closer to the real change trend of the water quality data, so that the accuracy and reliability of the time series prediction result are significantly improved.

[0101] In step S34, the confidence of the dynamic change value is evaluated, and the real-time confidence is calculated.

[0102] Firstly, the residual between the initial prediction sequence and the water quality parameter measured sequence is calculated to form a historical contemporaneous prediction deviation sequence, then the mean and variance of the historical contemporaneous prediction deviation sequence are calculated by using the sliding window mechanism, the historical deviation range is constructed with the mean as the center and twice the variance as the interval, the deviation between the dynamic change value and the historical change value is calculated, and the probability that the deviation falls within the historical deviation range is judged, the probability is converted into real-time confidence by fuzzy logic algorithm, the real-time confidence can objectively reflect the reliability level of the dynamic change value relative to the historical contemporaneous performance, and provides a quantitative confidence basis for subsequent early warning decision, so that the false alarm risk caused by unreliable prediction results is effectively reduced.

[0103] It should be noted that the historical change value is the change value with the same environmental factors and meteorological conditions as the dynamic change value in the historical database.

[0104] In step S35, the real-time confidence is combined with the multi-level early warning threshold to make a state decision, and the water quality early warning level is output.

[0105] The real-time confidence is used as an adaptive weight factor to dynamically adjust the multi-level early warning threshold obtained by training the historical water quality data in advance, that is, when the confidence is high, the original threshold is maintained to strictly distinguish, and when the confidence is low, the threshold range is relaxed in proportion, thereby forming an adaptive threshold interval.

[0106] Then, a decision model based on state transition logic is used to make a comprehensive state decision according to the distribution of the dynamic change value in different adaptive threshold intervals, combined with the real-time confidence, and output the water quality early warning level from normal, attention, early warning to alarm, so as to directly integrate the reliability index of the prediction result into the decision process, dynamically adjust the early warning strategy according to the data reliability, and finally realize the intelligent early warning effect of timely capturing water quality abnormalities and effectively avoiding false alarms, thereby significantly improving the decision reliability of the water quality telemetry system.

[0107] In step S36, the dynamic change value is packaged according to the water quality early warning level, and the target water quality data is output.

[0108] According to the water quality warning level, the emergency identifier of the data packaging format is determined, the dynamic change value is taken as the core data field, and the warning level, real-time confidence, time stamp, monitoring area coordinates and other metadata are embedded. Then, the data frame structure specified in the HJ212 protocol is used for encoding and assembling. Through the field separator and check code rules specified in the protocol, various data are combined into a transmission message conforming to the industry standard. Finally, the target water quality data containing complete warning decision information and accurate prediction results are output. This packaging method can integrate the multi-dimensional information generated in the analysis process into standardized data products that can be directly recognized by the monitoring system, realizing seamless connection from prediction analysis to decision execution.

[0109] It should be noted that between steps S35, steps S37-S39 are also included:

[0110] In step S37, historical water quality data is obtained, and the periodic change law of the historical water quality data is analyzed by a time series clustering algorithm to obtain threshold clusters.

[0111] By extracting long-term accumulated historical water quality data from the historical database deployed in the target monitoring area, the periodic change law of the historical water quality data is analyzed by a time series clustering algorithm based on dynamic time warping. Specifically, first, the historical water quality data is divided into equal time periods according to the annual period, and the water quality feature vectors of each time period are extracted. After calculating the similarity distance matrix between the water quality feature vectors of different time periods by the time series clustering algorithm, the K-shapes clustering algorithm is used to merge time periods with similar morphological change laws into the same category, forming data clusters representing different hydrological characteristics. Finally, the quantile features of the statistical distribution of the water quality data of each data cluster are calculated, and each quantile point is taken as the warning threshold under the corresponding hydrological condition. Finally, threshold clusters corresponding to different periodic change patterns are formed.

[0112] In step S38, redundant threshold points in the threshold cluster are removed to obtain initial multi-level warning thresholds.

[0113] All threshold points in the threshold cluster are mapped to a numerical space to form a data point set. The density peak search algorithm based on kernel density estimation is used to calculate the local density and relative distance of each threshold point in the numerical space. By finding threshold points with high local density and high relative distance as core threshold points, redundant threshold points with low density and close distance to the peak points are removed. The sorted core threshold points are sorted by numerical size and divided into levels to form the initial multi-level warning thresholds after simplification.

[0114] In step S39, according to the water quality data fluctuation characteristics on the preset time collection segment, the initial multi-level warning thresholds are real-time calibrated through the sliding window mechanism to generate multi-level warning thresholds.

[0115] From the water quality data collected in real time from the multi-parameter sensing device deployed in the target monitoring area, the continuous monitoring values of the water quality data in the preset time collection section are extracted to form a data sequence, the ratio of the standard deviation to the mean of the data sequence is calculated to obtain the water quality data fluctuation characteristic, then the water quality data fluctuation characteristic is taken as an adjustment factor, and the initial multi-level early warning threshold is weighted and fused to calculate, specifically: when the water quality data fluctuation characteristic is large, the threshold interval is correspondingly expanded to adapt to data fluctuation, and when the water quality data fluctuation characteristic is small, the threshold interval is relatively tightened, and finally the multi-level early warning threshold matched with the water quality fluctuation characteristic is generated according to the weighted calculation result, so that the early warning threshold can be adaptively adjusted according to the dynamic change characteristic of the water quality, thereby ensuring the early warning sensitivity and effectively avoiding the false alarm problem caused by normal fluctuation of the water quality.

[0116] The application also provides an industrial-grade water quality remote sensing terminal, as shown in Figure 5 The industrial-grade water quality remote sensing terminal comprises:

[0117] A data acquisition module 10 is configured to acquire multi-modal water quality data in a target monitoring area, perform a canonical correlation analysis operation on the multi-modal water quality data, and obtain a projection matrix of the multi-modal water quality data.

[0118] A feature fusion module 20 is configured to map the multi-modal water quality data to a common subspace according to the projection matrix, and obtain a fusion feature vector.

[0119] A prediction output module 30 is configured to pre-train the fusion feature vector, obtain a dynamic change value through time series prediction, and output target water quality data based on the dynamic change value.

[0120] Optionally, the data acquisition module 10 is further configured to:

[0121] According to a preset collection time period, initial water quality data sequences corresponding to each target parameter are acquired by a multi-parameter sensing device deployed in the target monitoring area.

[0122] Based on the initial water quality data sequences, a spatio-temporal alignment error between the target parameters is calculated.

[0123] The initial water quality data sequences are dynamically time-warped and calibrated according to the spatio-temporal alignment error to generate multi-modal water quality data.

[0124] Optionally, the data acquisition module 10 is further configured to:

[0125] A real-time observation data matrix containing each target parameter is constructed through a sliding time window mechanism.

[0126] A canonical correlation analysis operation is performed on the multi-modal water quality data to obtain a reference projection matrix.

[0127] Performing canonical correlation analysis operation on the real-time observation data matrix and the benchmark projection matrix, a projection matrix is obtained.

[0128] Optionally, the feature fusion module 20 is further configured to:

[0129] According to the projection matrix, the multi-modal water quality data is mapped to a common subspace to generate an initial fusion feature vector;

[0130] Performing environmental interference factor analysis on the initial fusion feature vector, a real-time signal-to-noise ratio compensation coefficient is calculated;

[0131] Optimizing the initial fusion feature vector by the real-time signal-to-noise ratio compensation coefficient to obtain a fusion feature vector.

[0132] Optionally, the prediction output module 30 is further configured to:

[0133] Based on the fusion feature vector, performing time series prediction on each target parameter through pre-training to obtain an initial prediction sequence;

[0134] Obtaining a water quality parameter measured sequence on a preset time collection segment, and calculating a prediction deviation between the initial prediction sequence and the water quality parameter measured sequence through residual analysis;

[0135] Compensating the initial prediction sequence according to the prediction deviation to obtain a dynamic change value.

[0136] Optionally, the prediction output module 30 is further configured to:

[0137] Performing confidence evaluation on the dynamic change value to obtain a real-time confidence;

[0138] Combining the real-time confidence with the multi-level warning threshold to make a state decision, and outputting a water quality warning level;

[0139] According to the water quality warning level, the dynamic change value is packaged to output target water quality data.

[0140] The industrial-grade water quality telemetry terminal further comprises a warning threshold generation module 10, configured to:

[0141] Obtaining historical water quality data, analyzing the periodic change rule of the historical water quality data through a time series clustering algorithm to obtain a threshold cluster;

[0142] Removing redundant threshold points in the threshold cluster to obtain an initial multi-level warning threshold;

[0143] According to the fluctuation characteristics of the water quality data on the preset time collection segment, the initial multi-level warning threshold is real-time calibrated through a sliding window mechanism to generate a multi-level warning threshold.

[0144] The above are all preferred embodiments of the present application, and do not limit the protection scope of the present application, so that: all equivalent changes made according to the structure, shape, principle of the present application should be covered in the protection scope of the present application.

Claims

1. An industrial-grade water quality remote sensing method, characterized in that, include: The initial water quality data sequence corresponding to each target parameter in the target monitoring area is obtained during the preset collection time period. After calculating the spatiotemporal alignment error between each target parameter, dynamic time warping and calibration are performed based on the initial water quality data sequence and the spatiotemporal alignment error to obtain multimodal water quality data. Based on the sliding time window mechanism and the multimodal water quality data, after obtaining the real-time observation data matrix and the reference projection matrix, canonical correlation analysis is performed on the real-time observation data matrix and the reference projection matrix to obtain the projection matrix of the multimodal water quality data. Based on the projection matrix, the multimodal water quality data is mapped to a common subspace to generate an initial fusion feature vector. Environmental interference factor analysis is performed on the initial fusion feature vector to calculate the real-time signal-to-noise ratio compensation coefficient. The initial fusion feature vector is then optimized based on the real-time signal-to-noise ratio compensation coefficient to obtain the fusion feature vector. The fused feature vector is pre-trained to obtain an initial prediction sequence. After calculating the prediction deviation between the initial prediction sequence and the obtained measured water quality parameter sequence, the initial prediction sequence is compensated based on the prediction deviation to obtain dynamic change values. The confidence of the dynamic change values ​​is evaluated to calculate the real-time confidence. The real-time confidence is combined with multi-level early warning thresholds to make a state decision and output the water quality early warning level. The dynamic change values ​​are encapsulated based on the water quality early warning level to output the target water quality data.

2. The industrial-grade water quality remote sensing method according to claim 1, characterized in that, The steps of acquiring the initial water quality data sequence corresponding to each target parameter within the target monitoring area during the preset collection time period, calculating the spatiotemporal alignment error between each target parameter, and then performing dynamic time warping and calibration based on the initial water quality data sequence and the spatiotemporal alignment error to obtain multimodal water quality data include: According to the preset collection time period, the initial water quality data sequence corresponding to each of the target parameters is obtained by multi-parameter sensing devices deployed on the target monitoring area; Based on the initial water quality data sequence, the spatiotemporal alignment error between each of the target parameters is calculated; The initial water quality data sequence is dynamically time-normalized and calibrated based on the spatiotemporal alignment error to generate the multimodal water quality data.

3. The industrial-grade water quality remote sensing method according to claim 2, characterized in that, The steps for obtaining the projection matrix of the multimodal water quality data by performing canonical correlation analysis on the real-time observation data matrix and the reference projection matrix based on the sliding time window mechanism and the multimodal water quality data include: The real-time observation data matrix containing the target parameters is constructed using the sliding time window mechanism. Perform canonical correlation analysis on the multimodal water quality data to obtain the baseline projection matrix; The projection matrix is ​​obtained by performing canonical correlation analysis on the real-time observation data matrix and the reference projection matrix.

4. The industrial-grade water quality remote sensing method according to claim 2, characterized in that, The steps of pre-training the fused feature vector to obtain an initial prediction sequence, calculating the prediction deviation between the initial prediction sequence and the obtained measured water quality parameter sequence, and compensating the initial prediction sequence based on the prediction deviation to obtain the dynamic change value include: Based on the fused feature vector, the multimodal water quality data is used for time series prediction through pre-training to obtain the initial prediction sequence; The measured sequence of water quality parameters is obtained, and the prediction deviation between the initial prediction sequence and the measured sequence of water quality parameters is calculated through residual analysis. The initial prediction sequence is compensated based on the prediction deviation to obtain the dynamic change value.

5. The industrial-grade water quality remote sensing method according to claim 1, characterized in that, Before the step of combining the real-time confidence level with the multi-level early warning threshold to make a state decision and output the water quality early warning level, the following steps are also included: Historical water quality data is acquired, and the periodic variation patterns of the historical water quality data are analyzed using a time-series clustering algorithm to obtain threshold clusters; Redundant threshold points are removed from the threshold cluster to obtain the initial multi-level early warning thresholds; Based on the water quality data fluctuation characteristics during the preset collection period, the initial multi-level early warning threshold is calibrated in real time using a sliding window mechanism to generate the multi-level early warning threshold.

6. An industrial-grade water quality remote sensing terminal, characterized in that, The industrial-grade water quality remote sensing terminal is applied to the industrial-grade water quality remote sensing method described above, and the industrial-grade water quality remote sensing terminal includes: The data acquisition and analysis module is used to acquire the initial water quality data sequence corresponding to each target parameter in the target monitoring area during a preset acquisition time period, and calculate the spatiotemporal alignment error between each target parameter. Then, it performs dynamic time warping and calibration based on the initial water quality data sequence and the spatiotemporal alignment error to obtain multimodal water quality data. The data acquisition and analysis module is also used to obtain a real-time observation data matrix and a reference projection matrix based on the sliding time window mechanism and the multimodal water quality data, and then perform canonical correlation analysis on the real-time observation data matrix and the reference projection matrix to obtain the projection matrix of the multimodal water quality data. The feature fusion module is used to map the multimodal water quality data to a common subspace based on the projection matrix, generate an initial fused feature vector, perform environmental interference factor analysis on the initial fused feature vector, calculate the real-time signal-to-noise ratio compensation coefficient, and optimize the initial fused feature vector based on the real-time signal-to-noise ratio compensation coefficient to obtain a fused feature vector. The prediction output module is used to pre-train the fused feature vector to obtain an initial prediction sequence, calculate the prediction deviation between the initial prediction sequence and the obtained measured water quality parameter sequence, compensate the initial prediction sequence based on the prediction deviation to obtain dynamic change values, evaluate the confidence of the dynamic change values, calculate the real-time confidence, combine the real-time confidence with multi-level warning thresholds to make state decisions, output the water quality warning level, encapsulate the dynamic change values ​​based on the water quality warning level, and output the target water quality data.

Images