Industrial water quality prediction method and system based on hybrid neural network

Abstract

The invention discloses an industrial water quality prediction method and system based on a hybrid neural network, and particularly relates to the field of industrial water quality prediction.The method comprises the steps that real-time water quality data in the industrial water treatment process is collected at an edge computing node, and a continuous water quality data sequence is constructed; performing time window slicing operation on the continuous water quality data sequence according to a preset time window length to generate an edge time window fragment sequence composed of a plurality of edge time window fragments; and inputting each edge time window segment in the edge time window segment sequence into a deep learning working condition identification model deployed at an edge computing node, and executing feature extraction operation on time distribution features and associated features of pollutant concentration changes in the edge time window segments. A working condition feature vector is constructed on an edge side, and a working condition dependent adaptive feature updating operation is executed in a cloud hybrid neural network, so that each input sequence is definitely bound to a real working condition state before entering a prediction link.

Description

Technical Field

[0001] This invention relates to the field of industrial water quality prediction technology, and more specifically, to an industrial water quality prediction method and system based on a hybrid neural network. Background Technology

[0002] In current deep learning-based industrial water quality prediction practices, data is typically collected in real time by edge computing nodes and uploaded to the cloud for prediction in the form of short time windows. However, industrial water systems go through different stages in actual operation, such as shutdown, load increase, shock disturbance and steady-state operation. These stages not only change the numerical level of pollutants, but also change the correlation and lag pattern between pollutants, so that the same indicator has completely different predictive meanings at different stages. Existing deep learning models are typically designed with the assumption that all input data comes from the same operating state and extracts features and constructs time dependencies under a unified structure. However, when the short window data at edge nodes is constantly changing, this setting will cause the model to receive sequences from different operating states at the same time. The convolutional layers and temporal layers inside the model cannot determine which operating state these sequences belong to, thus learning contradictory feature patterns at different operating stages. As runtime progresses, this state mixture will continue to accumulate, eventually causing the deep learning model in the cloud to struggle to maintain stable predictions under multiple operating conditions; Based on the above, the core problem that can be extracted is that deep learning models cannot distinguish the operating state of the input sequence under the edge computing architecture, which leads to confusion in the prediction logic between different operating states, thereby affecting the overall prediction accuracy and reliability. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an industrial water quality prediction method and system based on a hybrid neural network. By constructing a condition feature vector at the edge and performing condition-dependent adaptive feature update operations in the cloud-based hybrid neural network, each input sequence is explicitly bound to its actual condition state before entering the prediction link, thereby solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an industrial water quality prediction method based on a hybrid neural network, comprising: S1. Collect real-time water quality data of industrial water treatment process at edge computing nodes, construct a continuous water quality data sequence, and perform time window slicing operation on the continuous water quality data sequence according to the preset time window length to generate an edge time window segment sequence composed of multiple edge time window segments. S2. Input each edge time window segment in the edge time window segment sequence into the deep learning operating condition recognition model deployed on the edge computing node, perform feature extraction operation on the temporal distribution characteristics and correlation characteristics of pollutant concentration changes within the edge time window segment, and construct the operating condition feature vector corresponding to the edge time window segment based on the feature extraction operation results. S3. The edge computing node combines each edge time window segment with its corresponding working condition feature vector to form a synchronous input data pair to be predicted. After attaching the timestamp information corresponding to the collection time, it is sent to the cloud. The cloud performs sorting and time alignment operations on the synchronous input data pair to be predicted based on the timestamp information to obtain the cloud synchronous prediction input sequence. S4. The edge time window segments and the operating condition feature vector in the cloud synchronous prediction input sequence are simultaneously input into the hybrid neural network model from the cloud. Local convolution operation is performed on the edge time window segments in the convolutional network substructure, and temporal state update operation is performed in the bidirectional recursive network substructure. Adaptive adjustment operation is performed on the convolution kernel selection parameters, temporal weight allocation parameters and hidden state update parameters according to the operating condition feature vector to generate an industrial water quality prediction feature representation consistent with the operating condition. S5. The cloud performs regression inference calculation based on the industrial water quality prediction feature representation, outputs the effluent water quality prediction result at the target time, and writes the effluent water quality prediction result into the monitoring and control process of the industrial water system.

[0005] In a preferred embodiment, S1 includes: S1-1. Read the real-time water quality data of the industrial water treatment process point by point in the edge computing node, and write each real-time water quality data point into a continuous water quality data sequence according to the collection timestamp order to form a continuous water quality data sequence arranged in time order. S1-2. Calculate the start and end times of each time window in the continuous water quality data sequence in the edge computing node according to the preset time window length, and extract the corresponding data segments in the continuous water quality data sequence according to the calculated start and end times to generate a candidate time window segment set composed of multiple candidate time window segments. S1-3. In the edge computing node, perform a length verification judgment on each candidate time window segment in the candidate time window segment set. When the number of data points contained in the candidate time window segment is equal to the number of data points corresponding to the preset time window length, mark the candidate time window segment as an edge time window segment. When the number of data points contained in the candidate time window segment is less than the number of data points corresponding to the preset time window length, discard the candidate time window segment.

[0006] In a preferred embodiment, S1 further includes: S1-4. In the edge computing node, according to the time order of the candidate time window segments marked as edge time window segments, each edge time window segment is written into the edge time window segment sequence in sequence to construct an edge time window segment sequence composed of multiple edge time window segments.

[0007] In a preferred embodiment, S2 includes: S2-1. Construct a deep learning working condition recognition model in the edge computing node. Set the input layer of the model to receive a two-dimensional input matrix consisting of all pollutant concentration values ​​corresponding to each time point in the edge time window segment. Set the convolutional layer of the model to a convolution operation unit consisting of several fixed-size convolution kernels, so that each convolution kernel can perform covering operation, element-wise multiplication operation and element-wise summation operation along the time axis of the two-dimensional input matrix point by point.

[0008] In a preferred embodiment, S2 further includes: S2-2. In the edge computing node, the two-dimensional input matrix constructed by the edge time window segment is input into the deep learning working condition recognition model. In the convolutional layer of the model, a time-by-time point-by-time movement operation is performed on each convolution kernel. In each movement operation, the following is performed: the element-by-element product of all pollutant concentration values ​​within the coverage area of ​​the convolution kernel and the corresponding weights inside the convolution kernel is performed, and then the element-by-element summation is performed on all product results, thereby calculating the local time change feature vector arranged in time order. S2-3. In the edge computing node, the local time change feature vector is input into the recursive layer of the deep learning working condition recognition model, and the hidden state update operation is performed for each time point in the recursive layer. That is, the local time change feature of the current time point and the hidden state vector of the previous time point are linearly combined according to the state update formula of the recursive layer, and then a non-linear activation operation is performed on the linear combination result to generate a hidden state sequence updated in time order.

[0009] In a preferred embodiment, S2 further includes: S2-4. In the edge computing node, the hidden state vector corresponding to the final time point of the hidden state sequence is extracted as the working condition basic feature vector, and the working condition basic feature vector is input to the feature mapping unit. The feature mapping unit generates the working condition feature vector by performing a fully connected operation and an activation operation on the working condition basic feature vector, and writes the working condition feature vector into the working condition feature sequence corresponding to the current edge time window segment.

[0010] In a preferred embodiment, S3 includes: S3-1. In the edge computing node, perform a one-to-one pairing operation between each edge time window segment and its corresponding working condition feature vector, so as to sequentially write the edge time window segment and the working condition feature vector into the same data structure, and construct a synchronous input data pair to be predicted that contains the correspondence between the two. S3-2. Read the acquisition timestamp corresponding to each synchronous input data pair to be predicted in the edge computing node, and append the acquisition timestamp to the synchronous input data pair to be predicted, so that each synchronous input data pair to be predicted contains three data fields: edge time window segment, working condition feature vector and acquisition timestamp. S3-3. In the edge computing node, all synchronous input data pairs to be predicted, including timestamp fields, are sent to the cloud in the original collection order. The cloud then performs a sorting operation based on the collection timestamps in the synchronous input data pairs to construct a sorted input sequence to be aligned in chronological order. S3-4. Perform time alignment operation on the sorted input sequence to be aligned in the cloud. That is, detect the collection timestamp interval between adjacent synchronous input data pairs to be predicted according to the continuous time step requirement. When the timestamp interval meets the preset continuity condition, write it into the cloud synchronous prediction input sequence in order. When the timestamp interval does not meet the preset continuity condition, perform gap filling operation or fragment removal operation to form a cloud synchronous prediction input sequence with continuous time sequence and complete fragment correspondence.

[0011] In a preferred embodiment, S4-1, each edge time window segment in the cloud synchronous prediction input sequence is constructed into a two-dimensional convolutional input matrix in the cloud, the corresponding working condition feature vector is constructed into a one-dimensional auxiliary input vector, and the two-dimensional convolutional input matrix and the one-dimensional auxiliary input vector are respectively written into the input end of the convolutional network substructure and the input end of the bidirectional recursive network substructure of the hybrid neural network model. S4-2. In the convolutional network substructure of the hybrid neural network model, a local convolution operation is performed on the two-dimensional convolutional input matrix. That is, the convolution kernel is slid along the time dimension with a preset stride. At each position covered by the convolution kernel, the element-wise multiplication operation is performed on all pollutant concentration values ​​in the two-dimensional convolutional input matrix and the weight values ​​inside the convolution kernel, and the summation operation is performed to generate a convolutional local response sequence arranged in time order. S4-3. In the bidirectional recursive network substructure of the hybrid neural network model, the temporal state update operation is performed on the convolutional local response sequence in both forward and reverse temporal order. That is, at each time point, the convolutional local response value at the current time point is combined with the forward or reverse hidden state vector at the previous time point and a nonlinear activation operation is performed to generate the forward and reverse hidden state sequences. The two sequences are then concatenated according to the corresponding time points. S4-4. In the hybrid neural network model, based on the working condition feature vector as a one-dimensional auxiliary input vector, adaptive adjustment operations are performed on the convolution kernel selection parameters of the convolutional network substructure, the temporal weight allocation parameters of the bidirectional recursive network substructure, and the hidden state update parameters. That is, element-wise weighting, element-wise scaling, or element-wise offset operations are performed on the above parameters and the working condition feature vector to make the adjusted parameters consistent with the current working condition state. The spliced ​​hidden state vector calculated with the adjusted parameters is used as the industrial water quality prediction feature representation.

[0012] In a preferred embodiment, S5 includes: S5-1. Write the industrial water quality prediction feature representation into the first regression calculation layer of the multi-level regression inference unit in the cloud. In the first regression calculation layer, perform element-wise multiplication and summation operations on each feature element of the industrial water quality prediction feature representation and the first regression weight element to generate the first regression response vector. S5-2. Input the first regression response vector into the multi-threshold judgment unit in the cloud. In the multi-threshold judgment unit, perform the first threshold judgment, the second threshold judgment, and the cross threshold judgment on each element of the first regression response vector in sequence: when any element in the first regression response vector is greater than the first threshold, perform the first perturbation adjustment operation; otherwise, perform the first perturbation hold operation; after the first perturbation adjustment operation, when any element in the first regression response vector is less than the second threshold, perform the second perturbation reverse operation; otherwise, perform the second perturbation hold operation; after the second perturbation reverse operation, when there are elements in the first regression response vector that are greater than the first threshold and elements that are less than the second threshold at the same time, perform the cross perturbation fusion operation; otherwise, perform the cross perturbation termination operation. S5-3. In the cloud, the second regression weight element after processing by the first disturbance adjustment operation, the first disturbance hold operation, the second disturbance reverse operation, the second disturbance hold operation, the cross disturbance fusion operation, or the cross disturbance termination operation is written into the second regression calculation layer. In the second regression calculation layer, the first regression response vector and the second regression weight element are multiplied element by element and summed. After the summing operation is completed, the summing result is normalized for three consecutive segments to generate the predicted value of the effluent water quality at the target time. S5-4. Input the predicted effluent water quality value into the result writing unit in the cloud. Perform a control threshold judgment operation on the predicted effluent water quality value in the result writing unit: when the predicted effluent water quality value meets the control threshold requirement, execute the control flow writing operation and terminate the monitoring value update operation; otherwise, execute the monitoring value update operation and terminate the control flow writing operation.

[0013] An industrial water quality prediction system based on a hybrid neural network includes a windowing timing module, an operating condition analysis module, a calibration timing module, a feature generation module, and an output prediction module. The windowing timing module is used to collect real-time water quality data of industrial water treatment process at edge computing nodes, construct a continuous water quality data sequence, and perform time window slicing operation on the continuous water quality data sequence according to the preset time window length to generate an edge time window segment sequence composed of multiple edge time window segments. The condition analysis module is used to input each edge time window segment in the edge time window segment sequence into the deep learning condition recognition model deployed on the edge computing node, perform feature extraction operation on the temporal distribution characteristics and correlation characteristics of pollutant concentration changes within the edge time window segment, and construct the condition feature vector corresponding to the edge time window segment based on the feature extraction operation results. The calibration timing module is used by the edge computing node to combine each edge time window segment with its corresponding working condition feature vector to form a synchronous input data pair to be predicted. After attaching the timestamp information corresponding to the acquisition time, it is sent to the cloud. The cloud performs sorting and time alignment operations on the synchronous input data pair to be predicted based on the timestamp information to obtain the cloud synchronous prediction input sequence. The feature generation module is used to simultaneously input the edge time window segments and the operating condition feature vector from the cloud synchronous prediction input sequence into the hybrid neural network model. In the convolutional network substructure, local convolution operation is performed on the edge time window segments, and temporal state update operation is performed in the bidirectional recursive network substructure. Based on the operating condition feature vector, adaptive adjustment operation is performed on the convolution kernel selection parameters, temporal weight allocation parameters, and hidden state update parameters to generate an industrial water quality prediction feature representation consistent with the operating condition. The output prediction module is used to perform regression inference calculations based on the industrial water quality prediction feature representation by the cloud, output the effluent water quality prediction result at the target time, and write the effluent water quality prediction result into the monitoring and control process of the industrial water system.

[0014] The technical effects and advantages of this invention are as follows: In this invention, by performing working condition identification on short-window data at edge nodes and constructing working condition feature vectors, the cloud-based deep learning model can distinguish different operating states at the input stage, fundamentally eliminating the feature conflict problem caused by multiple working conditions being mixed into the same prediction link, thereby solving the problem of prediction instability of existing models under mixed state conditions. This invention constructs continuous time window segments at edge nodes and adds timestamps to perform cloud sorting and alignment, so that the predicted input maintains strict consistency in time structure, avoids temporal dependency chaos caused by different collection rhythms and segment misalignment, and improves the interpretability and maintainability of deep models under long-term operation. This invention extracts local change patterns from convolutional substructures and extracts front-to-back lag relationships from bidirectional recursive substructures in a cloud-based hybrid neural network structure, and dynamically adjusts the convolutional kernel selection parameters and temporal weights using operating condition feature vectors, so that the prediction link has adaptive feature expression capabilities under different operating conditions. This invention introduces progressive operators such as multi-threshold judgment, perturbation adjustment, perturbation reversal, and cross-fusion in the regression inference stage, so that the second regression weight forms a dynamic response structure under different working condition combinations, avoiding the accumulation of distortion caused by fixed parameters of the regression layer, thereby enhancing the sensitivity and stability of the prediction results to working condition switching. This invention performs two-path selection based on control thresholds in the result writing unit to ensure that the predicted value has a unique and effective landing point in the monitoring and control processes. This avoids the repeated use or erroneous triggering of the predicted value in different processes and ensures that the predicted output has decision-making significance that can be directly applied in the industrial field. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method steps of the present invention.

[0016] Figure 2 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Refer to the instruction manual appendix Figure 1-2 An embodiment of the present invention provides an industrial water quality prediction method based on a hybrid neural network, comprising: S1. Collect real-time water quality data of industrial water treatment process at edge computing nodes, construct a continuous water quality data sequence, and perform time window slicing operation on the continuous water quality data sequence according to the preset time window length to generate an edge time window segment sequence composed of multiple edge time window segments. S2. Input each edge time window segment in the edge time window segment sequence into the deep learning operating condition recognition model deployed on the edge computing node, perform feature extraction operation on the temporal distribution characteristics and correlation characteristics of pollutant concentration changes within the edge time window segment, and construct the operating condition feature vector corresponding to the edge time window segment based on the feature extraction operation results. S3. The edge computing node combines each edge time window segment with its corresponding working condition feature vector to form a synchronous input data pair to be predicted. After attaching the timestamp information corresponding to the collection time, it is sent to the cloud. The cloud performs sorting and time alignment operations on the synchronous input data pair to be predicted based on the timestamp information to obtain the cloud synchronous prediction input sequence. S4. The edge time window segments and the operating condition feature vector in the cloud synchronous prediction input sequence are simultaneously input into the hybrid neural network model from the cloud. Local convolution operation is performed on the edge time window segments in the convolutional network substructure, and temporal state update operation is performed in the bidirectional recursive network substructure. Adaptive adjustment operation is performed on the convolution kernel selection parameters, temporal weight allocation parameters and hidden state update parameters according to the operating condition feature vector to generate an industrial water quality prediction feature representation consistent with the operating condition. S5. The cloud performs regression inference calculation based on the industrial water quality prediction feature representation, outputs the effluent water quality prediction result at the target time, and writes the effluent water quality prediction result into the monitoring and control process of the industrial water system.

[0019] S1 includes: S1-1. Read the real-time water quality data of the industrial water treatment process point by point in the edge computing node, and write each real-time water quality data point into a continuous water quality data sequence according to the collection timestamp order to form a continuous water quality data sequence arranged in time order. S1-2. Calculate the start and end times of each time window in the continuous water quality data sequence in the edge computing node according to the preset time window length, and extract the corresponding data segments in the continuous water quality data sequence according to the calculated start and end times to generate a candidate time window segment set composed of multiple candidate time window segments. S1-3. In the edge computing node, perform a length verification judgment on each candidate time window segment in the candidate time window segment set. When the number of data points contained in the candidate time window segment is equal to the number of data points corresponding to the preset time window length, mark the candidate time window segment as an edge time window segment. When the number of data points contained in the candidate time window segment is less than the number of data points corresponding to the preset time window length, discard the candidate time window segment. In S1-1, it should be noted that the construction of the "continuous water quality data sequence" involves using the acquisition timestamp as the sole sorting criterion within the edge computing node. For each real-time water quality data point in the industrial water treatment process, a sequential "read-verify-write" operation is performed point-by-point. Specifically, after receiving a piece of real-time water quality data, the edge computing node first reads the acquisition timestamp and the values ​​of each monitoring indicator carried by the real-time water quality data. Then, it appends the real-time water quality data to the same storage structure in order of timestamp from earliest to latest. This storage structure is defined as the continuous water quality data sequence. This point-by-point writing method ensures that there are no time reversals or time gaps in the continuous water quality data sequence, thus providing a continuous and ordered data foundation for subsequent slicing operations according to time windows. In S1-2, it should be noted that the generation of the "candidate time window segment set" involves performing pairwise calculations of the start and end times on the constructed continuous water quality data sequence, using a preset time window length as the sole dividing metric. Based on this, an interval truncation operation is performed on the continuous water quality data sequence. Specifically, the edge computing node first determines the start time of the first time window based on the first timestamp of the continuous water quality data sequence. Then, it calculates the corresponding end time by superimposing the preset time window length on the start time. Subsequently, it truncates all data points in the continuous water quality data sequence whose timestamps fall within the interval between the start and end times, and saves the data points within this interval as a candidate time window segment. Then, based on the end time of the previous time window, it repeats the calculation of the start and end times and the data truncation operation until the entire continuous water quality data sequence has been traversed, thereby constructing a candidate time window segment set composed of multiple candidate time window segments. In S1-3, it should be noted that the length verification of the "candidate time window segment" involves performing a comparison operation between the number of data points in each candidate time window segment and the number of data points corresponding to the preset time window length within the candidate time window segment set. Specifically, for each candidate time window segment, the edge computing node first counts the actual number of real-time water quality data points contained in the candidate time window segment, and then compares the counted number of data points with the theoretical number of data points determined by the sampling frequency and the preset time window length. When the actual number of data points equals the theoretical number of data points, the candidate time window segment is marked as an edge time window segment, and its data is considered complete within its time range and can be used for subsequent calculations. When the actual number of data points is less than the theoretical number of data points, the candidate time window segment is considered to have missing boundaries or incomplete data collection, and the candidate time window segment is directly discarded to ensure that all edge time window segments with consistent length and continuous data are used in subsequent processing.

[0020] S1 also includes: S1-4. In the edge computing node, according to the time order of the candidate time window segments marked as edge time window segments, each edge time window segment is written into the edge time window segment sequence in sequence to construct an edge time window segment sequence composed of multiple edge time window segments. In S1-4, it should be noted that the construction of the "edge time window segment sequence" involves re-sorting and sequentially writing all candidate time window segments that have been marked as edge time window segments in chronological order. Specifically, the edge computing node first sorts all edge time window segments from earliest to latest according to the start timestamp of each edge time window segment. Then, according to the sorting results, each edge time window segment is written sequentially into the same sequential storage structure, which is defined as the edge time window segment sequence. Through this construction process, the multiple edge time window segments that were originally discrete are reorganized in the time dimension into a temporally continuous edge time window segment sequence with consistent window length and complete data. This provides a directly callable input sequence for the water quality data of the industrial water treatment process that is input into the hybrid neural network according to the time window.

[0021] S2 includes: S2-1. Construct a deep learning working condition recognition model in the edge computing node. Set the input layer of the model to receive a two-dimensional input matrix consisting of all pollutant concentration values ​​corresponding to each time point in the edge time window segment. Set the convolutional layer of the model to a convolution operation unit consisting of several fixed-size convolutional kernels, so that each convolutional kernel can perform covering operation, element-wise multiplication operation and element-wise summation operation along the time axis of the two-dimensional input matrix point by point. In S2-1, it should be noted that the construction of the "deep learning working condition recognition model" involves first defining the input within the edge computing nodes, and then building the computational structure layer by layer. Specifically, each time point in the edge time window segment collects a set of pollutant concentration values. For example, a certain time point may contain multi-dimensional values ​​such as pH, ammonia nitrogen, chemical oxygen demand, sulfide, and chloride ions. This set of multi-dimensional values ​​occupies one row in the two-dimensional input matrix. As time progresses, the pollutant concentrations at different time points are sequentially written into the next row. The final two-dimensional input matrix represents the time axis by rows and the pollutant dimensions by columns, ensuring that each matrix element corresponds to a specific pollutant. The text describes a process where a convolutional kernel is used to define the concentration of a specific pollutant at a given time point. Each kernel in the convolutional layer is set as a small matrix with a time dimension shorter than the total time length of the edge time window segment and a pollutant dimension consistent with the pollutant dimension of the two-dimensional input matrix. This ensures that each covering operation of the kernel completely covers the concentration values ​​of all pollutants within a continuous time range. Subsequently, by moving the kernel along the time axis point by point and performing covering operations, element-wise multiplication operations, and element-wise summation operations at each covering position, a structural foundation is laid for generating local temporal variation features.

[0022] S2 also includes: S2-2. In the edge computing node, the two-dimensional input matrix constructed by the edge time window segment is input into the deep learning working condition recognition model. In the convolutional layer of the model, a time-by-time point-by-time movement operation is performed on each convolution kernel. In each movement operation, the following is performed: the element-by-element product of all pollutant concentration values ​​within the coverage area of ​​the convolution kernel and the corresponding weights inside the convolution kernel is performed, and then the element-by-element summation is performed on all product results, thereby calculating the local time change feature vector arranged in time order. S2-3. In the edge computing node, the local time change feature vector is input into the recursive layer of the deep learning working condition recognition model, and the hidden state update operation is performed for each time point in the recursive layer. That is, the local time change feature of the current time point and the hidden state vector of the previous time point are linearly combined according to the state update formula of the recursive layer, and then a non-linear activation operation is performed on the linear combination result to generate a hidden state sequence updated in time order. In S2-2, it should be noted that the calculation of the "local temporal change feature vector" involves sliding each convolutional kernel along the time axis with a fixed stride within the convolutional layer, precisely calculating the comprehensive change within that time period at each sliding position. Specifically, in one movement operation, the convolutional kernel covers several rows within a continuous time range of the two-dimensional input matrix, with each row corresponding to a time point and each column corresponding to a pollutant concentration. The convolutional kernel also stores weight values ​​corresponding one-to-one with each pollutant and each time position, arranged by row and column. At the coverage position, the edge computing node first calculates the weight of each pollutant within the coverage area. The concentration values ​​are multiplied element-wise with the weight values ​​at the same location in the convolution kernel to obtain a set of product results. Then, all product results are summed one by one to form a convolution response value for that coverage location. After that, the convolution kernel moves backward along the time axis by one time step and repeats the above coverage, element-wise multiplication and summation operations until the convolution kernel has traversed the entire time axis of the two-dimensional input matrix. The convolution response values ​​formed at all coverage locations are arranged in chronological order to form the local temporal change feature vector corresponding to the convolution kernel, which is used to represent the local pattern of pollutant concentration changes in this edge time window segment in different time periods. In S2-3, it should be noted that the calculation of the "hidden state sequence" involves processing the local time-varying feature vectors point by point in the recursive layer in chronological order, and updating the hidden state vector at each time point. Specifically, an initial hidden state vector is first set for the recursive layer in the edge computing nodes; this can typically be a zero vector or a fixed vector obtained through training. When the recursive layer processes the first time point, the elements of the local time-varying feature vector at that time point are used as the input for the current time point, and the initial hidden state vector is used as the hidden state input for the previous time point. Both are then subjected to linear operations with the preset weight coefficients in the recursive layer. Specifically, the features at the current time point are multiplied element-wise with their corresponding weights and summed, and the hidden state vector at the previous time point is multiplied element-wise with its corresponding weights and summed. The two sums are then superimposed to form a linear combination. Subsequently, a non-linear activation operation (e.g., compression and transformation using a preset activation function) is performed on the linear combination result element-wise to obtain the hidden state vector for the first time point. When processing the second and subsequent time points, the recursive layer repeats the above steps each time: read the local time change features of the current time point, read the hidden state vector of the previous time point, perform linear combination operation, and perform nonlinear activation operation to obtain a new hidden state vector. As the time point advances from the start time to the end time of the edge time window segment, the hidden state vectors corresponding to all time points are arranged in chronological order to form a complete hidden state sequence, which is used to centrally represent the working condition evolution trajectory of this edge time window segment over the entire time range.

[0023] S2 also includes: S2-4. In the edge computing node, the hidden state vector corresponding to the final time point of the hidden state sequence is extracted as the working condition basic feature vector, and the working condition basic feature vector is input to the feature mapping unit. The feature mapping unit generates the working condition feature vector by performing a fully connected operation and an activation operation on the working condition basic feature vector, and writes the working condition feature vector into the working condition feature sequence corresponding to the current edge time window segment. In S2-4, it should be noted that the generation of the "operating condition feature vector" is performed after the hidden state sequence calculation is completed. The hidden state vector at the end of the sequence is extracted from the time dimension as the basic operating condition feature vector. Then, a fully connected operation and an activation operation are performed in the feature mapping unit to further compress and rearrange it. The specific process is as follows: The edge computing node first reads the hidden state vector corresponding to the final time point from the hidden state sequence and inputs the hidden state vector as the basic operating condition feature vector into the feature mapping unit. The feature mapping unit pre-stores a set of fully connected operation weights and bias parameters. In the fully connected operation stage, each element in the basic operating condition feature vector is multiplied with the corresponding weight parameter, and all product results and bias parameters are summed to obtain a set of linearly transformed intermediate feature values. Subsequently, the feature mapping unit performs an activation operation on the intermediate feature values ​​element by element, mapping the intermediate feature values ​​to a preset numerical range and introducing nonlinear expressive power to obtain a fixed-length operating condition feature vector. Finally, the edge computing node establishes a one-to-one correspondence between the working condition feature vector and the currently processed edge time window segment, and writes the working condition feature vector into the working condition feature sequence in chronological order, so that each working condition feature vector in the working condition feature sequence can be consistent with an edge time window segment in terms of time and working condition status, providing a structurally complete and semantically clear working condition representation for subsequent cloud prediction input.

[0024] S3 includes: S3-1. In the edge computing node, perform a one-to-one pairing operation between each edge time window segment and its corresponding working condition feature vector, so as to sequentially write the edge time window segment and the working condition feature vector into the same data structure, and construct a synchronous input data pair to be predicted that contains the correspondence between the two. S3-2. Read the acquisition timestamp corresponding to each synchronous input data pair to be predicted in the edge computing node, and append the acquisition timestamp to the synchronous input data pair to be predicted, so that each synchronous input data pair to be predicted contains three data fields: edge time window segment, working condition feature vector and acquisition timestamp. S3-3. In the edge computing node, all synchronous input data pairs to be predicted, including timestamp fields, are sent to the cloud in the original collection order. The cloud then performs a sorting operation based on the collection timestamps in the synchronous input data pairs to construct a sorted input sequence to be aligned in chronological order. S3-4. Perform time alignment operation on the sorted input sequence to be aligned in the cloud. That is, detect the collection timestamp interval between adjacent synchronous input data pairs to be predicted according to the continuous time step requirement. When the timestamp interval meets the preset continuity condition, write it into the cloud synchronous prediction input sequence in order. When the timestamp interval does not meet the preset continuity condition, perform gap filling operation or fragment removal operation to form a cloud synchronous prediction input sequence with continuous time sequence and complete fragment correspondence. In S3-1, it should be noted that the construction process of synchronous input data pairs to be predicted involves combining edge time window segments with operating condition feature vectors in a one-to-one correspondence within the edge computing node. Specifically, the edge computing node sequentially reads each edge time window segment from the edge time window segment sequence and simultaneously reads the operating condition feature vector corresponding to that edge time window segment in time order from the operating condition feature sequence. Subsequently, the edge computing node writes these two objects into a unified data structure in a fixed order, where the first part stores the edge time window segment and the second part stores the operating condition feature vector, thus forming a data pair containing two fields. This operation is repeated for all edge time window segments to ensure that each edge time window segment is precisely paired with an operating condition feature vector, ultimately forming a data set composed of multiple synchronous input data pairs to be predicted. In S3-2, it should be noted that the timestamp appending process is performed after the synchronous input data pair to be predicted has been written into the edge time window segment and the operating condition feature vector. After generating each synchronous input data pair to be predicted, the edge computing node reads the actual acquisition timestamp of the edge time window segment from the acquisition device or data buffer. This timestamp corresponds to the last or most representative acquisition moment in the time window. Subsequently, the edge computing node writes this timestamp as the third field into the corresponding synchronous input data pair to be predicted, so that the data pair includes three parts: the edge time window segment field, the operating condition feature vector field, and the acquisition timestamp field. In this way, each synchronous input data pair to be predicted has clear time positioning information, providing a reliable time reference for subsequent sorting and time alignment operations performed in the cloud. In S3-3, it should be noted that the sending and sorting process of the synchronous input data pairs to be predicted is completed by the collaborative actions of the edge computing nodes and the cloud. The edge computing nodes first send all synchronous input data pairs to be predicted, which have been timestamped, to the cloud one by one in the order of generation, so that the cloud can obtain the complete data set in the order of receipt. Subsequently, the cloud reads the collection timestamp field from each synchronous input data pair to be predicted and performs sorting operations based on these timestamps. The sorting operation arranges all synchronous input data pairs to be predicted in order of timestamp from earliest to latest, so that the sorted sequence can truly reflect the actual collection order on the edge side and prevent time chaos caused by network latency or changes in transmission order. This sorting result constitutes the sorted input sequence to be aligned, laying a strict time order foundation for the next stage of time alignment. In S3-4, it should be noted that the purpose of time alignment is to ensure the continuity of the sorted input sequence to be aligned in the time dimension, thereby constructing a cloud-synchronized prediction input sequence. The specific process is as follows: The cloud reads the collection timestamps of adjacent synchronous prediction input data pairs from the sorted input sequence to be aligned, and calculates the time interval between adjacent timestamps according to the preset continuous time step condition. When the time interval is equal to the preset step size, the cloud directly writes the synchronous prediction input data pair into the cloud-synchronized prediction input sequence. When the time interval is greater than the preset step size, the cloud first determines the length of the missing time period, and then performs gap filling or fragment removal operations according to the set strategy. Gap filling is usually performed by interpolation, copying, or constructing placeholder data fragments to fill in the missing time points, so that the sequence remains continuous. Fragment removal is performed by removing the fragment from the sequence when the missing information is irreparable. Through the above steps, the cloud finally constructs a cloud-synchronized prediction input sequence that is continuous in time order, complete in data structure, and can be directly used as input for hybrid neural network models.

[0025] S4-1. In the cloud, each edge time window segment in the cloud synchronous prediction input sequence is constructed into a two-dimensional convolutional input matrix, and the corresponding working condition feature vector is constructed into a one-dimensional auxiliary input vector. The two-dimensional convolutional input matrix and the one-dimensional auxiliary input vector are respectively written into the input end of the convolutional network substructure and the input end of the bidirectional recursive network substructure of the hybrid neural network model. S4-2. In the convolutional network substructure of the hybrid neural network model, a local convolution operation is performed on the two-dimensional convolutional input matrix. That is, the convolution kernel is slid along the time dimension with a preset stride. At each position covered by the convolution kernel, the element-wise multiplication operation is performed on all pollutant concentration values ​​in the two-dimensional convolutional input matrix and the weight values ​​inside the convolution kernel, and the summation operation is performed to generate a convolutional local response sequence arranged in time order. S4-3. In the bidirectional recursive network substructure of the hybrid neural network model, the temporal state update operation is performed on the convolutional local response sequence in both forward and reverse temporal order. That is, at each time point, the convolutional local response value at the current time point is combined with the forward or reverse hidden state vector at the previous time point and a nonlinear activation operation is performed to generate the forward and reverse hidden state sequences. The two sequences are then concatenated according to the corresponding time points. S4-4. In the hybrid neural network model, based on the working condition feature vector as a one-dimensional auxiliary input vector, adaptive adjustment operations are performed on the convolution kernel selection parameters of the convolutional network substructure, the temporal weight allocation parameters of the bidirectional recursive network substructure, and the hidden state update parameters. That is, element-wise weighting, element-wise scaling, or element-wise offset operations are performed on the above parameters and the working condition feature vector to make the adjusted parameters consistent with the current working condition state. The spliced ​​hidden state vector calculated with the adjusted parameters is used as the industrial water quality prediction feature representation. In S4-1, it should be noted that the input structure of the hybrid neural network model must receive edge time window segments and operating condition feature vectors separately. Therefore, the cloud first performs an explicit formatting construction process for the two types of inputs. For edge time window segments, the cloud uses the pollutant concentration values ​​at all time points within the segment as a basis, and arranges the multidimensional concentration values ​​at each time point in chronological order into a two-dimensional convolutional input matrix, so that the rows of the matrix correspond to consecutive time points and the columns correspond to pollutant types. For operating condition feature vectors, the cloud keeps their one-dimensional vector form unchanged and treats them as auxiliary input vectors, writing them into the auxiliary input terminal of the model. Finally, the cloud sends the two-dimensional convolutional input matrix into the input terminal of the convolutional network substructure and the operating condition feature vector into the input terminal of the bidirectional recursive network substructure, so that the hybrid neural network model can simultaneously receive time-series pollutant concentration data and operating condition information. In S4-2, it should be noted that the convolutional network substructure processes the two-dimensional convolutional input matrix by sequentially moving the convolutional kernel along the time dimension and performing a complete numerical calculation at each kernel position. Specifically, after the convolutional kernel covers a continuous time range of the two-dimensional convolutional input matrix, the cloud performs element-wise multiplication on each pollutant concentration value in the covered area with the corresponding weight value in the convolutional kernel, and then accumulates all the multiplication results from beginning to end to form a convolutional response value. Subsequently, the convolutional kernel moves backward by one step along the time axis and performs the same element-wise multiplication and accumulation operation on the new covered area again. The convolutional kernel repeats this process along the entire time dimension, ultimately forming a convolutional local response sequence arranged in time order, enabling the convolutional network substructure to record the pollutant change pattern of the entire edge time window segment within the local time range. In S4-3, it should be noted that the bidirectional recursive network substructure processes the convolutional local response sequence by performing state calculations in two time directions separately, and combining the results of the two directions into a complete set of temporal features. Specifically, in the forward direction, starting from the beginning of the time series, the cloud reads the convolutional local response value point by point and performs element-wise linear combination and element-wise activation with the forward hidden state vector of the previous time point according to the state update formula. Similarly, in the reverse direction, starting from the end of the time series, the cloud reads the convolutional local response value point by point and performs the same linear combination and activation operation with the reverse hidden state vector of the previous time point. The forward and reverse hidden state sequences generated independently in the two directions are connected into a new hidden state vector at each corresponding time point through a state concatenation operation, thus forming a temporal feature structure that can simultaneously express forward and backward information. In S4-4, it should be noted that the role of the operating condition feature vector is to directly participate in the adaptive adjustment of model parameters, enabling the convolution kernel selection parameters, temporal weight allocation parameters, and hidden state update parameters to dynamically change with the operating condition. Specifically, the cloud inputs the operating condition feature vector into the parameter adjustment module and performs an element-wise weighted operation on the vector and the convolution kernel selection parameters, so that the effective weights of the convolution kernel change proportionally under different operating conditions. Subsequently, the operating condition feature vector and the temporal weight allocation parameters of the bidirectional recursive network substructure are subjected to an element-wise scaling operation, so that the forward and backward time-dependent weights have differentiated distributions under different operating conditions. Finally, the operating condition feature vector and the hidden state update parameters are subjected to an element-wise offset operation, so that the update amplitude of the hidden state is consistent with the operating condition. After three types of adaptive adjustments, the cloud uses the adjusted parameters to perform the final feature formation on the concatenated hidden state vector, making it an industrial water quality prediction feature representation that can accurately express the current operating condition. It should be further explained that in S4, after the cloud completes the local convolution operation of the convolutional network substructure, the forward and backward hidden state calculations of the bidirectional recursive network substructure, and the adaptive adjustment of the convolution kernel selection parameters, temporal weight allocation parameters, and hidden state update parameters based on the operating condition feature vector, the final result is an industrial water quality prediction feature representation composed of a concatenated hidden state vector. This industrial water quality prediction feature representation uses the convolutional local response sequence as the time input basis, the hidden state vector formed by concatenating the forward and backward hidden state vectors at the same time point as the intermediate structure, and the fixed-length feature vector obtained by performing the final feature formation operation on the hidden state vector with the adaptively adjusted convolution kernel selection parameters, temporal weight allocation parameters, and hidden state update parameters. This fixed-length feature vector completely contains the temporal distribution relationship of all pollutant concentrations in the edge time window segment, the bidirectional time dependence relationship, and the state response relationship formed by the operating condition feature vector after parameter adjustment. It is the unique and explicit input feature vector in the subsequent regression inference operation.

[0026] S5 includes: S5-1. Write the industrial water quality prediction feature representation into the first regression calculation layer of the multi-level regression inference unit in the cloud. In the first regression calculation layer, perform element-wise multiplication and summation operations on each feature element of the industrial water quality prediction feature representation and the first regression weight element to generate the first regression response vector. S5-2. Input the first regression response vector into the multi-threshold judgment unit in the cloud. In the multi-threshold judgment unit, perform the first threshold judgment, the second threshold judgment, and the cross threshold judgment on each element of the first regression response vector in sequence: when any element in the first regression response vector is greater than the first threshold, perform the first perturbation adjustment operation; otherwise, perform the first perturbation hold operation; after the first perturbation adjustment operation, when any element in the first regression response vector is less than the second threshold, perform the second perturbation reverse operation; otherwise, perform the second perturbation hold operation; after the second perturbation reverse operation, when there are elements in the first regression response vector that are greater than the first threshold and elements that are less than the second threshold at the same time, perform the cross perturbation fusion operation; otherwise, perform the cross perturbation termination operation. S5-3. In the cloud, the second regression weight element after processing by the first disturbance adjustment operation, the first disturbance hold operation, the second disturbance reverse operation, the second disturbance hold operation, the cross disturbance fusion operation, or the cross disturbance termination operation is written into the second regression calculation layer. In the second regression calculation layer, the first regression response vector and the second regression weight element are multiplied element by element and summed. After the summing operation is completed, the summing result is normalized for three consecutive segments to generate the predicted value of the effluent water quality at the target time. S5-4. Input the predicted effluent water quality value into the result writing unit in the cloud. Perform a control threshold judgment operation on the predicted effluent water quality value in the result writing unit: when the predicted effluent water quality value meets the control threshold requirement, execute the control flow writing operation and terminate the monitoring value update operation; otherwise, execute the monitoring value update operation and terminate the control flow writing operation. In S5-1, it should be noted that for the regression initialization operation of the industrial water quality prediction feature representation, the industrial water quality prediction feature representation is treated as a one-dimensional input vector composed of multiple feature elements arranged in a fixed order in the cloud. The first regression weight element in the first regression calculation layer of the multi-level regression inference unit is treated as a weight matrix arranged in rows and columns for step-by-step calculation. Specifically, in the cloud, a row of first regression weight elements is first selected in the first regression calculation layer. Each weight element in this row is matched one-to-one with the corresponding feature element in the industrial water quality prediction feature representation. Then, a product is performed on each pair of matched feature elements and weight elements. The calculation yields a set of product results. Then, the summation operation is performed on this set of product results from the first to the last to obtain a response value corresponding to the first regression weight element of that row. Subsequently, the cloud sequentially selects the next row of the first regression weight element in the first regression calculation layer and repeats the above matching, product, and summation operation process until all rows of the first regression weight elements in the first regression calculation layer have been traversed. The response values ​​calculated in each row are arranged in the order of row index to form a one-dimensional vector composed of multiple response values. This one-dimensional vector is the first regression response vector, which can be directly used as the input data basis for subsequent multi-threshold judgment unit to perform judgment operations. In S5-2, it should be noted that: for the multi-threshold judgment and perturbation operation of the first regression response vector, the judgment link and the corresponding weight perturbation link are executed in three layers in the cloud with the first regression response vector as the only input basis, and the numerical state of the second regression weight element is dynamically adjusted. Specifically, the cloud first reads each element of the first regression response vector element by element in the multi-threshold judgment unit, and compares the value of each element with the first threshold. When any element is found to be greater than the first threshold in any comparison, the first perturbation adjustment operation is immediately triggered, and the element-by-element offset calculation is performed on the second regression weight element. That is, according to the preset proportional coefficient, the second regression weight element is added to the corresponding index position with an offset proportional to the value of the element, so as to bias the second regression weight element as a whole towards the direction that is more sensitive to high amplitude regression response. If no element greater than the first threshold is found after comparing all elements, the first perturbation hold operation is performed along the otherwise path. In this operation, the cloud directly passes the second regression weight element to the next stage with the current value state without making any changes. Subsequently, the cloud reads the first regression response vector element by element again in the same multi-threshold judgment unit, compares the value of each element with the second threshold, and when any element is found to be less than the second threshold in any comparison, the second perturbation reverse operation is triggered. The second regression weight element that has been processed in the first stage is adjusted element by element, that is, an offset proportional to the value of the element is subtracted at the corresponding index position or the sign direction of the weight at that position is changed to prevent the regression weight from being locked in a state that is insensitive to low amplitude response for a long time. If no element less than the second threshold is found after all comparisons with the second threshold are completed, the second perturbation hold operation is performed along the otherwise path. In this operation, the value state of the second regression weight element after the first stage processing remains unchanged. Finally, the cloud performs a third traversal of the first regression response vector, recording whether there are elements greater than the first threshold and elements less than the second threshold in the same traversal. When the result indicates that there are elements greater than the first threshold and elements less than the second threshold in the first regression response vector, a cross-perturbation fusion operation is triggered. In this operation, the cloud performs an element-wise weighted summation calculation on the offset set formed by the aforementioned first perturbation adjustment operation and the second perturbation inverse operation, fusing the two offset effects at each index position of the second regression weight element, so that the second regression weight element simultaneously reflects the amplification effect on the high-amplitude response and the correction effect on the low-amplitude response. If the above simultaneous existence condition is not met, that is, there is no case where both high-amplitude and low-amplitude extreme responses are contained at the same time, then the cross-perturbation termination operation is performed along the otherwise path. In this operation, the cloud directly retains the numerical state of the second regression weight element after the second stage processing is completed, without adding any new offsets, thus forming the final set of second regression weight elements for the next joint regression calculation. In S5-3, it should be noted that the final regression calculation of the second regression weight element is performed in the cloud using the first regression response vector as the input vector and the second regression weight element obtained from the previous stage as the calculation parameter. Element-by-element multiplication and summation are then performed to construct a regression response structure that can be used to generate the predicted value at the target time. Specifically, the cloud first writes all the second regression weight elements processed by the first perturbation adjustment operation, the first perturbation hold operation, the second perturbation reverse operation, the second perturbation hold operation, the cross perturbation fusion operation, or the cross perturbation termination operation into the second regression calculation layer. This establishes a one-to-one correspondence between each weight element within the second regression calculation layer and its corresponding index position in the first regression response vector. Subsequently, the cloud reads each element of the first regression response vector element by element in the second regression calculation layer and performs a multiplication and summation operation on each element and its corresponding second regression weight element. The cloud performs a series of element-wise multiplication operations to obtain a set of element-wise product results. Then, it performs a summation operation on this set of element-wise product results in indexed order, accumulating from the first product result to the last, aggregating all product results into a single summation value. This summation value serves as the basic output for regression calculations. Immediately after the summation operation, the cloud performs three consecutive normalization operations on the summation value according to a preset three-stage normalization rule: the first normalization operation compresses the overall numerical range of the summation value, keeping it within a controllable range; the second normalization operation proportionally stretches the subtle numerical differences in the summation value, preserving fine variations in the calculation result; the third normalization operation maps the result to a fixed numerical range used for water quality prediction. The final value after these three normalization stages is the predicted effluent water quality value for the target time, which can be directly used for subsequent control or monitoring process calculations. In S5-4, it should be noted that the process of writing the predicted effluent quality values ​​is as follows: The cloud uses the predicted values ​​as input, performs a clear control threshold judgment through the result writing unit, and selects either a control path or a monitoring path based on the judgment result. Specifically, the cloud first inputs the predicted effluent quality values ​​at the target time into the result writing unit, and then performs a comparison judgment between the predicted values ​​and the control threshold within the result writing unit. When the comparison result indicates that the predicted values ​​meet the control threshold requirements, the cloud executes the control process writing operation along the execution path of the judgment chain. In the control process writing operation, the cloud writes the predicted values ​​into the control process structure used to control the operating status of the industrial water system, and because the control process writing... If the input operation has covered the control requirements, the monitoring value update operation is terminated simultaneously to prevent the monitoring process from repeatedly receiving the same predicted value; otherwise, when the predicted value does not meet the control threshold requirements, the cloud performs the monitoring value update operation along the other path of the judgment chain. In this operation, the cloud writes the predicted value into the monitoring process structure of the industrial water system for subsequent observation and recording, and at the same time terminates the control process write operation so that the predicted result that does not meet the control conditions will not be written into the control process. Through this clear judgment chain and the two-way execution path, the cloud can ensure that the predicted value of the effluent water quality has a unique and valid write position in the monitoring process or the control process, avoid data conflicts and ensure that the role path of the prediction result in the system is clear. In the working principle of this method, it is important to note that the entire execution chain revolves around the core objective of "obtaining stable, reliable, and highly targeted water quality predictions under dynamically changing industrial operating conditions." Layer-by-layer calculation and judgment ensure that the prediction results remain consistent with the actual operating conditions. Specifically, the edge computing nodes first construct continuous and uninterrupted time window segments in chronological order, ensuring that all subsequent calculations are based on structurally complete inputs. Next, the edge computing nodes utilize a deep learning operating condition recognition model to perform convolution and temporal operations on each time window segment, obtaining operating condition feature vectors that represent the actual operating state, thus preventing different operating conditions from being mixed into the same prediction space. Subsequently, the edge nodes combine the time window segments and operating condition feature vectors by timestamp and send them to the cloud. The cloud strictly performs sorting and alignment according to the chronological order, ensuring that all inputs enter the prediction model in the order of actual operating conditions. In the cloud-based hybrid neural network model, convolutional network sub-nodes... This method extracts local temporal variations and uses a bidirectional recursive network substructure to extract temporal dependencies. It then adaptively adjusts various parameters using operating condition feature vectors, ensuring that the predicted features accurately reflect pollutant variation patterns under different operating conditions. Finally, in the regression inference stage, the cloud-based system abandons a fixed regression structure and instead constructs a dynamic regression chain through multi-threshold judgment, multi-path perturbation adjustment, and piecewise normalization. This allows the predicted values ​​to automatically adapt to high and low amplitude feature changes during calculation, ultimately outputting a prediction result completely consistent with the current operating conditions. Based on the comparison between the prediction result and the control threshold, the prediction is either written into the monitoring or control process, ensuring that the prediction result directly serves operational decisions. Through this execution chain, the method achieves its objective: to continuously output "true predicted values ​​consistent with the current operating conditions" in industrial scenarios with rapidly changing operating conditions and constantly fluctuating data structures, avoiding the distortion problems caused by traditional predictions due to operating condition mismatch, feature contamination, and parameter fixation.

[0027] An industrial water quality prediction system based on a hybrid neural network includes a windowing timing module, an operating condition analysis module, a calibration timing module, a feature generation module, and an output prediction module. The windowing timing module is used to collect real-time water quality data of industrial water treatment process at edge computing nodes, construct a continuous water quality data sequence, and perform time window slicing operation on the continuous water quality data sequence according to the preset time window length to generate an edge time window segment sequence composed of multiple edge time window segments. The condition analysis module is used to input each edge time window segment in the edge time window segment sequence into the deep learning condition recognition model deployed on the edge computing node, perform feature extraction operation on the temporal distribution characteristics and correlation characteristics of pollutant concentration changes within the edge time window segment, and construct the condition feature vector corresponding to the edge time window segment based on the feature extraction operation results. The calibration timing module is used by the edge computing node to combine each edge time window segment with its corresponding working condition feature vector to form a synchronous input data pair to be predicted. After attaching the timestamp information corresponding to the acquisition time, it is sent to the cloud. The cloud performs sorting and time alignment operations on the synchronous input data pair to be predicted based on the timestamp information to obtain the cloud synchronous prediction input sequence. The feature generation module is used to simultaneously input the edge time window segments and the operating condition feature vector from the cloud synchronous prediction input sequence into the hybrid neural network model. In the convolutional network substructure, local convolution operation is performed on the edge time window segments, and temporal state update operation is performed in the bidirectional recursive network substructure. Based on the operating condition feature vector, adaptive adjustment operation is performed on the convolution kernel selection parameters, temporal weight allocation parameters, and hidden state update parameters to generate an industrial water quality prediction feature representation consistent with the operating condition. The output prediction module is used to perform regression inference calculations based on the industrial water quality prediction feature representation by the cloud, output the effluent water quality prediction result at the target time, and write the effluent water quality prediction result into the monitoring and control process of the industrial water system.

[0028] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An industrial water quality prediction method based on hybrid neural networks, characterized in that, include: S1. Collect real-time water quality data of industrial water treatment process at edge computing nodes, construct a continuous water quality data sequence, and perform time window slicing operation on the continuous water quality data sequence according to the preset time window length to generate an edge time window segment sequence composed of multiple edge time window segments. S2. Input each edge time window segment in the edge time window segment sequence into the deep learning operating condition recognition model deployed on the edge computing node, perform feature extraction operation on the temporal distribution characteristics and correlation characteristics of pollutant concentration changes within the edge time window segment, and construct the operating condition feature vector corresponding to the edge time window segment based on the feature extraction operation results. S3. The edge computing node combines each edge time window segment with its corresponding working condition feature vector to form a synchronous input data pair to be predicted. After attaching the timestamp information corresponding to the collection time, it is sent to the cloud. The cloud performs sorting and time alignment operations on the synchronous input data pair to be predicted based on the timestamp information to obtain the cloud synchronous prediction input sequence. S4. The edge time window segments and the operating condition feature vector in the cloud synchronous prediction input sequence are simultaneously input into the hybrid neural network model from the cloud. Local convolution operation is performed on the edge time window segments in the convolutional network substructure, and temporal state update operation is performed in the bidirectional recursive network substructure. Adaptive adjustment operation is performed on the convolution kernel selection parameters, temporal weight allocation parameters and hidden state update parameters according to the operating condition feature vector to generate an industrial water quality prediction feature representation consistent with the operating condition. S5. The cloud performs regression inference calculation based on the industrial water quality prediction feature representation, outputs the effluent water quality prediction result at the target time, and writes the effluent water quality prediction result into the monitoring and control process of the industrial water system.

2. The industrial water quality prediction method based on a hybrid neural network according to claim 1, characterized in that: S1 includes: S1-1. Read the real-time water quality data of the industrial water treatment process point by point in the edge computing node, and write each real-time water quality data point into a continuous water quality data sequence according to the collection timestamp order to form a continuous water quality data sequence arranged in time order. S1-2. Calculate the start and end times of each time window in the continuous water quality data sequence in the edge computing node according to the preset time window length, and extract the corresponding data segments in the continuous water quality data sequence according to the calculated start and end times to generate a candidate time window segment set composed of multiple candidate time window segments. S1-3. In the edge computing node, perform a length verification judgment on each candidate time window segment in the candidate time window segment set. When the number of data points contained in the candidate time window segment is equal to the number of data points corresponding to the preset time window length, mark the candidate time window segment as an edge time window segment. When the number of data points contained in the candidate time window segment is less than the number of data points corresponding to the preset time window length, discard the candidate time window segment.

3. The industrial water quality prediction method based on a hybrid neural network according to claim 2, characterized in that: S1 also includes: S1-4. In the edge computing node, according to the time order of the candidate time window segments marked as edge time window segments, each edge time window segment is written into the edge time window segment sequence in sequence to construct an edge time window segment sequence composed of multiple edge time window segments.

4. The industrial water quality prediction method based on a hybrid neural network according to claim 3, characterized in that: S2 includes: S2-1. Construct a deep learning working condition recognition model in the edge computing node. Set the input layer of the model to receive a two-dimensional input matrix consisting of all pollutant concentration values ​​corresponding to each time point in the edge time window segment. Set the convolutional layer of the model to a convolution operation unit consisting of several fixed-size convolution kernels, so that each convolution kernel can perform covering operation, element-wise multiplication operation and element-wise summation operation along the time axis of the two-dimensional input matrix point by point.

5. The industrial water quality prediction method based on a hybrid neural network according to claim 4, characterized in that: S2 also includes: S2-2. In the edge computing node, the two-dimensional input matrix constructed by the edge time window segment is input into the deep learning working condition recognition model. In the convolutional layer of the model, a time-by-time point-by-time movement operation is performed on each convolution kernel. In each movement operation, the following is performed: the element-by-element product of all pollutant concentration values ​​within the coverage area of ​​the convolution kernel and the corresponding weights inside the convolution kernel is performed, and then the element-by-element summation is performed on all product results, thereby calculating the local time change feature vector arranged in time order. S2-3. In the edge computing node, the local time change feature vector is input into the recursive layer of the deep learning working condition recognition model, and the hidden state update operation is performed for each time point in the recursive layer. That is, the local time change feature of the current time point and the hidden state vector of the previous time point are linearly combined according to the state update formula of the recursive layer, and then a non-linear activation operation is performed on the linear combination result to generate a hidden state sequence updated in time order.

6. The industrial water quality prediction method based on a hybrid neural network according to claim 5, characterized in that: S2 also includes: S2-4. In the edge computing node, the hidden state vector corresponding to the final time point of the hidden state sequence is extracted as the working condition basic feature vector, and the working condition basic feature vector is input to the feature mapping unit. The feature mapping unit generates the working condition feature vector by performing a fully connected operation and an activation operation on the working condition basic feature vector, and writes the working condition feature vector into the working condition feature sequence corresponding to the current edge time window segment.

7. The industrial water quality prediction method based on a hybrid neural network according to claim 6, characterized in that: S3 includes: S3-1. In the edge computing node, perform a one-to-one pairing operation between each edge time window segment and its corresponding working condition feature vector, so as to sequentially write the edge time window segment and the working condition feature vector into the same data structure, and construct a synchronous input data pair to be predicted that contains the correspondence between the two. S3-2. Read the acquisition timestamp corresponding to each synchronous input data pair to be predicted in the edge computing node, and append the acquisition timestamp to the synchronous input data pair to be predicted, so that each synchronous input data pair to be predicted contains three data fields: edge time window segment, working condition feature vector and acquisition timestamp. S3-3. In the edge computing node, all synchronous input data pairs to be predicted, including timestamp fields, are sent to the cloud in the original collection order. The cloud then performs a sorting operation based on the collection timestamps in the synchronous input data pairs to construct a sorted input sequence to be aligned in chronological order. S3-4. Perform time alignment operation on the sorted input sequence to be aligned in the cloud. That is, detect the collection timestamp interval between adjacent synchronous input data pairs to be predicted according to the continuous time step requirement. When the timestamp interval meets the preset continuity condition, write it into the cloud synchronous prediction input sequence in order. When the timestamp interval does not meet the preset continuity condition, perform gap filling operation or fragment removal operation to form a cloud synchronous prediction input sequence with continuous time sequence and complete fragment correspondence.

8. The industrial water quality prediction method based on a hybrid neural network according to claim 7, characterized in that: S4-1. In the cloud, each edge time window segment in the cloud synchronous prediction input sequence is constructed into a two-dimensional convolutional input matrix, and the corresponding working condition feature vector is constructed into a one-dimensional auxiliary input vector. The two-dimensional convolutional input matrix and the one-dimensional auxiliary input vector are respectively written into the input end of the convolutional network substructure and the input end of the bidirectional recursive network substructure of the hybrid neural network model. S4-2. In the convolutional network substructure of the hybrid neural network model, a local convolution operation is performed on the two-dimensional convolutional input matrix. That is, the convolution kernel is slid along the time dimension with a preset stride. At each position covered by the convolution kernel, the element-wise multiplication operation is performed on all pollutant concentration values ​​in the two-dimensional convolutional input matrix and the weight values ​​inside the convolution kernel, and the summation operation is performed to generate a convolutional local response sequence arranged in time order. S4-3. In the bidirectional recursive network substructure of the hybrid neural network model, the temporal state update operation is performed on the convolutional local response sequence in both forward and reverse temporal order. That is, at each time point, the convolutional local response value at the current time point is combined with the forward or reverse hidden state vector at the previous time point and a nonlinear activation operation is performed to generate the forward and reverse hidden state sequences. The two sequences are then concatenated according to the corresponding time points. S4-4. In the hybrid neural network model, based on the working condition feature vector as a one-dimensional auxiliary input vector, adaptive adjustment operations are performed on the convolution kernel selection parameters of the convolutional network substructure, the temporal weight allocation parameters of the bidirectional recursive network substructure, and the hidden state update parameters. That is, element-wise weighting, element-wise scaling, or element-wise offset operations are performed on the above parameters and the working condition feature vector to make the adjusted parameters consistent with the current working condition state. The spliced ​​hidden state vector calculated with the adjusted parameters is used as the industrial water quality prediction feature representation.

9. The industrial water quality prediction method based on a hybrid neural network according to claim 8, characterized in that: S5 includes: S5-1. Write the industrial water quality prediction feature representation into the first regression calculation layer of the multi-level regression inference unit in the cloud. In the first regression calculation layer, perform element-wise multiplication and summation operations on each feature element of the industrial water quality prediction feature representation and the first regression weight element to generate the first regression response vector. S5-2. Input the first regression response vector into the multi-threshold judgment unit in the cloud. In the multi-threshold judgment unit, perform the first threshold judgment, the second threshold judgment, and the cross threshold judgment on each element of the first regression response vector in sequence: when any element in the first regression response vector is greater than the first threshold, perform the first perturbation adjustment operation; otherwise, perform the first perturbation hold operation; after the first perturbation adjustment operation, when any element in the first regression response vector is less than the second threshold, perform the second perturbation reverse operation; otherwise, perform the second perturbation hold operation; after the second perturbation reverse operation, when there are elements in the first regression response vector that are greater than the first threshold and elements that are less than the second threshold at the same time, perform the cross perturbation fusion operation; otherwise, perform the cross perturbation termination operation. S5-3. In the cloud, the second regression weight element after processing by the first disturbance adjustment operation, the first disturbance hold operation, the second disturbance reverse operation, the second disturbance hold operation, the cross disturbance fusion operation, or the cross disturbance termination operation is written into the second regression calculation layer. In the second regression calculation layer, the first regression response vector and the second regression weight element are multiplied element by element and summed. After the summing operation is completed, the summing result is normalized for three consecutive segments to generate the predicted value of the effluent water quality at the target time. S5-4. Input the predicted effluent water quality value into the result writing unit in the cloud. Perform a control threshold judgment operation on the predicted effluent water quality value in the result writing unit: when the predicted effluent water quality value meets the control threshold requirement, execute the control flow writing operation and terminate the monitoring value update operation; otherwise, execute the monitoring value update operation and terminate the control flow writing operation.

10. An industrial water quality prediction system based on a hybrid neural network, comprising a windowing timing module, an operating condition analysis module, a calibration timing module, a feature generation module, and an output prediction module, characterized in that: The windowing timing module is used to collect real-time water quality data of industrial water treatment process at edge computing nodes, construct a continuous water quality data sequence, and perform time window slicing operation on the continuous water quality data sequence according to the preset time window length to generate an edge time window segment sequence composed of multiple edge time window segments. The condition analysis module is used to input each edge time window segment in the edge time window segment sequence into the deep learning condition recognition model deployed on the edge computing node, perform feature extraction operation on the temporal distribution characteristics and correlation characteristics of pollutant concentration changes within the edge time window segment, and construct the condition feature vector corresponding to the edge time window segment based on the feature extraction operation results. The calibration timing module is used by the edge computing node to combine each edge time window segment with its corresponding working condition feature vector to form a synchronous input data pair to be predicted. After attaching the timestamp information corresponding to the acquisition time, it is sent to the cloud. The cloud performs sorting and time alignment operations on the synchronous input data pair to be predicted based on the timestamp information to obtain the cloud synchronous prediction input sequence. The feature generation module is used to simultaneously input the edge time window segments and the operating condition feature vector from the cloud synchronous prediction input sequence into the hybrid neural network model. In the convolutional network substructure, local convolution operation is performed on the edge time window segments, and temporal state update operation is performed in the bidirectional recursive network substructure. Based on the operating condition feature vector, adaptive adjustment operation is performed on the convolution kernel selection parameters, temporal weight allocation parameters, and hidden state update parameters to generate an industrial water quality prediction feature representation consistent with the operating condition. The output prediction module is used to perform regression inference calculations based on the industrial water quality prediction feature representation by the cloud, output the effluent water quality prediction result at the target time, and write the effluent water quality prediction result into the monitoring and control process of the industrial water system.

Images