Pond dissolved oxygen content prediction method and system based on cross-modal information fusion

Abstract

The invention relates to the field of water environment monitoring, and provides a pond dissolved oxygen content prediction method and system based on cross-modal information fusion. The method comprises the steps that multi-source time sequence data of a pond is collected and preprocessed, and a structured time sequence data set is obtained; constructing a fusion model, and training the fusion model through the structured time sequence data set to obtain a prediction model; and predicting time series data to be predicted through the prediction model to obtain a dissolved oxygen content prediction sequence. According to the method, the reliability and robustness of the prediction result of the cooperative influence of the dissolved oxygen content under different environmental factors are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water environment monitoring, and in particular to a pond dissolved oxygen content prediction method and system based on cross-modal information fusion. BACKGROUND

[0002] Dissolved oxygen content is one of the most important water quality parameters in aquaculture water bodies, directly affecting the growth, health and survival rate of cultured organisms. The dissolved oxygen content is influenced by water quality parameters (such as water temperature, pH value, historical dissolved oxygen value) and meteorological environmental parameters (such as air temperature, wind speed, air pressure, humidity, rainfall, etc.), showing complex nonlinear dynamic characteristics. Accurate prediction of pond dissolved oxygen content is of great significance for preventing hypoxia death of cultured organisms, optimizing operation strategies of oxygenation equipment, and reducing energy consumption of aquaculture. With the development of Internet of Things technology and artificial intelligence technology, deep learning-based dissolved oxygen prediction methods have gradually become a research hotspot in the field of smart fisheries, providing new technical means for precise management of aquaculture.

[0003] Existing dissolved oxygen prediction methods mostly use a single deep learning model, such as long short-term memory network (LSTM), gated recurrent unit (GRU) or Transformer model. LSTM model performs well in capturing short-term time series dynamics, but its memory capacity for long time series data decays with increasing sequence length, making it difficult to effectively model long-term dependencies; while long sequence prediction models such as Informer can handle long time span data, but they perform poorly in capturing local mutations and short-term subtle changes. In addition, existing methods usually simply concatenate water quality modal data and meteorological modal data as model input when dealing with multi-source heterogeneous data, failing to fully exploit the deep interaction and synergistic mechanism between different modal data, resulting in limited prediction accuracy and robustness of the model for dissolved oxygen content changes, especially when facing sudden environmental changes such as temperature fluctuations, air pressure drops, heavy rainfall, etc., the prediction effect is often not ideal, which is difficult to meet the actual needs of fine management of aquaculture. SUMMARY

[0004] The present application provides a pond dissolved oxygen content prediction method and system based on cross-modal information fusion to solve the defects of the prior art.

[0005] The present application provides a pond dissolved oxygen content prediction method based on cross-modal information fusion, comprising:

[0006] S1: Collecting multi-source time series data of the pond and preprocessing to obtain a structured time series dataset;

[0007] S2: Constructing a fusion model, training the fusion model with the structured time series dataset to obtain a prediction model;

[0008] S3: Using the prediction model, predict the time series data to be predicted to obtain the dissolved oxygen content prediction sequence.

[0009] According to the present invention, a method for predicting dissolved oxygen content in ponds based on cross-modal information fusion is provided, wherein the multi-source time-series data in step S1 includes:

[0010] Water quality modal data, including water temperature data, historical dissolved oxygen data, and pH data;

[0011] Meteorological modal data, which includes temperature data, wind force data, wind speed data, wind direction data, humidity data, atmospheric pressure data, and current rainfall data.

[0012] According to the method for predicting dissolved oxygen content in ponds based on cross-modal information fusion provided by the present invention, step S1 further includes:

[0013] S11: Collect multi-source time-series data of the pond;

[0014] S12: Calculate the statistical distribution parameters of multiple feature dimensions in the multimodal time series data, identify abnormal data points through an anomaly detection algorithm based on standard deviation threshold, and obtain an anomaly label sequence;

[0015] S13: Correct the abnormal data points according to the abnormal labeling sequence, replace minor abnormal values ​​with the time neighborhood mean, and mark extreme abnormal values ​​as null values ​​to obtain preliminary cleaned data;

[0016] S14: Apply a sliding window smoothing algorithm to the preliminary cleaned data to reduce noise and obtain smooth time series data;

[0017] S15: For missing values ​​in the smoothed time series data, an interpolation strategy is selected based on the missing length to fill in the missing values, thereby obtaining continuous and complete data;

[0018] S16: Perform interval normalization transformation on multiple feature dimensions of the continuous complete data to obtain a structured time series dataset.

[0019] According to the method for predicting dissolved oxygen content in ponds based on cross-modal information fusion provided by the present invention, step S2 further includes:

[0020] S21: Construct a fusion model;

[0021] S22: Divide the structured time series dataset to obtain a model development dataset including a training set, a validation set, and a test set;

[0022] S23: Develop training samples based on the model dataset and input them into the fusion model;

[0023] S24: Based on the preset loss function, the fusion model is optimized and trained using the Adam optimizer to obtain the prediction model.

[0024] According to the present invention, a method for predicting dissolved oxygen content in ponds based on cross-modal information fusion is provided. The fusion model in step S21 is a cascaded model composed of an Informer modal feature extractor and an LSTM time-series refiner connected in series. The Informer modal feature extractor is used to extract cross-modal global dependency features, and the LSTM time-series refiner is used to capture local short-term dynamic change features.

[0025] According to the method for predicting dissolved oxygen content in ponds based on cross-modal information fusion provided by the present invention, in step S24, the expression of the loss function is:

[0026] ;

[0027] in, For loss function, The total number of training samples, For training sample index values, To predict the total number of time steps, For the current time step, For time step index, For the first training samples Predicted dissolved oxygen content at any given time. For the first training samples The actual dissolved oxygen content at any given time.

[0028] According to the method for predicting dissolved oxygen content in ponds based on cross-modal information fusion provided by the present invention, in step S3, the expression for predicting the time series data to be predicted using the prediction model is as follows:

[0029] ;

[0030] in, This indicates that the Informer modal feature extractor performs data processing. This indicates that the LSTM timing refiner performs data processing. Input window length, To predict the step size, For the past A multimodal input sequence at time steps. For the predicted future The predicted dissolved oxygen content sequence at each time step.

[0031] A second aspect of the present invention provides a pond dissolved oxygen content prediction system based on cross-modal information fusion, comprising:

[0032] Data processing module: used to collect multi-source time-series data from the pond and perform preprocessing to obtain a structured time-series dataset;

[0033] The training module is used to train a pre-built fusion model using the structured time-series dataset to obtain a prediction model.

[0034] A prediction output module is configured to use the prediction model trained by the construction training module to predict the time series data to be predicted and obtain a dissolved oxygen content prediction sequence.

[0035] A third aspect of the present invention provides a pond dissolved oxygen content prediction device based on cross-modal information fusion, comprising:

[0036] A memory and at least one processor, wherein the memory stores instructions;

[0037] At least one of the processors invokes the instructions in the memory to cause a pond dissolved oxygen content prediction device based on cross-modal information fusion to perform a pond dissolved oxygen content prediction method based on cross-modal information fusion as described in any of the preceding claims.

[0038] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement a method for predicting dissolved oxygen content in a pond based on cross-modal information fusion as described in any of the preceding claims.

[0039] This invention provides a method, system, device, and storage medium for predicting dissolved oxygen content in ponds based on cross-modal information fusion. By constructing a prediction model that cascades and fuses an Informer modal feature extractor and an LSTM time-series refiner, it can simultaneously leverage the advantages of Informer in processing long-term series and capturing cross-modal global dependencies, as well as the advantages of LSTM in extracting local dynamic changes and capturing short-term subtle fluctuations. This significantly improves the prediction accuracy and stability of pond dissolved oxygen content trends, enabling the model to accurately grasp the long-term evolution of dissolved oxygen content while effectively identifying short-term fluctuations caused by environmental factors such as sudden temperature changes, sudden drops in air pressure, and heavy rainfall. This effectively overcomes the limitations of existing single models in terms of long-term memory decay or insufficient capture of local details. Furthermore, the systematic data preprocessing workflow employed in this invention, including outlier removal, denoising, missing value imputation, and normalization, effectively ensures the quality and continuity of input data, providing a high-quality data foundation for model training and further enhancing the generalization ability and practical application value of the prediction model. Meanwhile, this invention utilizes the self-attention mechanism of Informer to deeply mine the complex nonlinear interaction relationships between water quality modal data and meteorological modal data, avoiding the problem of insufficient information utilization caused by simple data splicing in traditional methods. This allows the model to more accurately understand the synergistic influence mechanism of different environmental factors on dissolved oxygen content, thereby improving the reliability and robustness of the prediction results. Especially when facing complex and variable aquaculture environments and sudden environmental changes, it can still maintain high prediction accuracy. Overall, this invention can provide more reliable technical support for precise oxygenation control, energy consumption optimization management, and aquaculture risk early warning in aquaculture, and has good industrial application prospects and economic benefits. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0041] Figure 1 A schematic diagram of a pond dissolved oxygen content prediction method based on cross-modal information fusion provided by the present invention;

[0042] Figure 2 This is a schematic diagram of the fusion model in a pond dissolved oxygen content prediction method based on cross-modal information fusion provided by the present invention.

[0043] Figure 3This is a schematic diagram of a pond dissolved oxygen content prediction system based on cross-modal information fusion, provided by the present invention.

[0044] Figure labeling: 100, data processing module; 200, training module; 300, prediction output module. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0046] The embodiments of the present invention are described below with reference to the figures.

[0047] like Figure 1 As shown, this invention provides a method for predicting dissolved oxygen content in ponds based on cross-modal information fusion, comprising:

[0048] S1: Collect multi-source time-series data from the pond and preprocess it to obtain a structured time-series dataset.

[0049] The multi-source time-series data in step S1 includes: water quality modal data, which includes water temperature data, historical dissolved oxygen data, and pH data; and meteorological modal data, which includes air temperature data, wind force data, wind speed data, wind direction data, humidity data, atmospheric pressure data, and current rainfall data.

[0050] Step S1 further includes:

[0051] S11: Collect multi-source time-series data of the pond.

[0052] In step S11, this invention utilizes water quality sensors to collect various water quality parameters, and simultaneously utilizes meteorological sensors to collect various meteorological parameters. Each sensor records data at fixed intervals, forming a continuous time series. After collection, this invention aligns the collected raw data according to timestamps to ensure that the water quality modal data and meteorological modal data remain synchronized in the time dimension.

[0053] S12: Calculate the statistical distribution parameters of multiple feature dimensions in the multimodal time series data, identify abnormal data points through an anomaly detection algorithm based on standard deviation threshold, and obtain an anomaly label sequence.

[0054] This invention calculates the mean and standard deviation for each feature dimension (such as water temperature, pH value, air pressure, etc.). Specifically, for a feature sequence containing N data points, the mean of the sequence is calculated, then the sum of squares of the deviations of each data point from the mean is calculated, divided by the total number of data points, and the square root is taken to obtain the standard deviation. Subsequently, this invention performs anomaly detection based on the standard deviation threshold; that is, when the deviation of a data point from the mean at a certain time exceeds three times the standard deviation, the data point is marked as an outlier. After this invention traverses all time steps and all feature dimensions, the positions of data points that meet the anomaly conditions are recorded, forming an anomaly marking sequence. This sequence identifies which features at which time steps are abnormal.

[0055] Furthermore, during the sensor acquisition process, "jump values" or "extreme values" may occur due to equipment failure or external interference. For example, the water temperature may suddenly rise to 80°C or the dissolved oxygen may suddenly drop to 0. Therefore, in order to avoid outliers affecting the training accuracy of the model, the present invention performs outlier removal in step S12.

[0056] The expression determines whether data at any given time deviates from the mean by more than three standard deviations, thus classifying it as an anomaly.

[0057] ;

[0058] in, for Data values ​​at any given time This is the calculated average value.

[0059] S13: Correct the abnormal data points according to the abnormal labeling sequence, replace minor outliers with the time neighborhood mean, and mark extreme outliers as null values ​​to obtain preliminary cleaned data.

[0060] In step S13, after locating specific abnormal data points based on the anomaly marker sequence, the present invention determines the degree of anomaly. For minor anomalies, i.e., data points deviating from the average by three to five standard deviations, the present invention extracts the normal values ​​of the adjacent time steps before and after the data point, calculates the arithmetic mean of the adjacent values, and uses this mean to replace the original anomaly value. For extreme anomalies, i.e., data points deviating from the average by more than five standard deviations or showing obviously physically impossible values ​​(such as negative water temperature, dissolved oxygen of 0, etc.), the present invention marks them as null values. After correction processing, the jump points and extreme error values ​​in the original data sequence are corrected or marked, forming preliminary cleaned data.

[0061] S14: Apply a sliding window smoothing algorithm to the preliminary cleaned data to reduce noise and obtain smooth time series data.

[0062] In step S14, the present invention sets the sliding window width to 3 to 5 time steps. For each time point in the initially cleaned data, the data values ​​of that time point and several time steps prior are extracted. Then, the arithmetic mean of all data values ​​within the window is calculated, and this average is used as the smoothed value for the current time point. Subsequently, the sliding window moves forward step by step from the beginning of the sequence, repeating the above averaging process for each time point. Through the sliding window smoothing algorithm, high-frequency noise and short-term random fluctuations are effectively suppressed, the data curve becomes smoother, and the overall trend is preserved. After processing, smoothed time-series data is obtained, and the numerical continuity and stability are significantly improved.

[0063] S15: For missing values ​​in the smoothed time series data, an interpolation strategy is selected based on the missing length to fill in the missing values, thereby obtaining continuous and complete data.

[0064] In step S15, this invention statistically smooths the time series data, identifying the positions marked as null values ​​and determining the number of consecutive missing time steps for each segment. When the number of consecutive missing time steps does not exceed three, this invention employs a linear interpolation strategy. Specifically, the timestamps and values ​​of the two valid data points before and after the missing segment are extracted, and the values ​​at the missing positions are calculated proportionally based on the time interval and filled into the corresponding positions. When the number of consecutive missing time steps exceeds three, this invention uses spline interpolation or extracts the mean of similar historical time periods for filling. After interpolation completion, all null value positions are assigned reasonable estimated values, forming continuous and complete data, and the data sequence no longer has interruptions in the time dimension.

[0065] S16: Perform interval normalization transformation on multiple feature dimensions of the continuous complete data to obtain a structured time series dataset.

[0066] This invention scans all values ​​for each feature dimension—water temperature, pH, air pressure, and humidity—throughout the entire time series, identifying the minimum and maximum values. Then, for each data point in that dimension, the minimum value is subtracted from its original value, and the result is divided by the difference between the maximum and minimum values, yielding a normalized value. This normalization transformation maps features with different physical dimensions and numerical ranges to a uniform range of 0 to 1, eliminating the impact of dimensional differences on model training. Finally, after processing, a structured time-series dataset is obtained. This dataset is continuous in time, numerically stable, with anomaly removal, noise smoothing, missing value imputation, and consistent value ranges for all feature dimensions.

[0067] S2: Construct a fusion model by training the fusion model using the structured time-series dataset to obtain a prediction model.

[0068] Step S2 further includes:

[0069] S21: Construct a fusion model.

[0070] The fusion model in step S21 is a cascaded model composed of an Informer modal feature extractor and an LSTM temporal refiner connected in series. The Informer modal feature extractor is used to extract cross-modal global dependency features, and the LSTM temporal refiner is used to capture local short-term dynamic change features.

[0071] like Figure 2 As shown, the fusion model constructed in this invention adopts a cascaded structure, consisting of an Informer modal feature extractor and an LSTM temporal refiner connected in series. The Informer modal feature extractor is located in the first stage of the cascaded model, and its function is to extract cross-modal global dependency features. This invention first establishes an input embedding layer, which receives the input sequence from the structured temporal dataset. The shape of the input sequence is the number of time steps multiplied by the feature dimension. The embedding layer performs a linear transformation on the input sequence through a weight matrix, mapping the multidimensional feature vector of each time step to a higher-dimensional embedding space, resulting in an embedding feature matrix. Subsequently, this invention constructs a ProbSparse self-attention mechanism, which first generates a query matrix, a key matrix, and a value matrix. Specifically, this invention multiplies the embedding feature matrix by three different weight matrices to obtain the query matrix, key matrix, and value matrix, where the dimension of each matrix is ​​the number of time steps multiplied by the feature dimension. This invention then calculates the matrix product of the query matrix and the transpose of the key matrix to obtain an attention score matrix, where each element represents the attention intensity of one time step to another. After scaling and normalizing the attention score matrix, this invention retains only the top u% of the largest values ​​in each row, setting the remaining elements to zero to achieve sparse attention. The sparse attention matrix is ​​then multiplied by the value matrix to obtain the attention output feature, which contains the global dependencies between different time steps in the sequence. This invention further constructs a distillation encoder structure, which consists of multiple layers of one-dimensional convolutions and downsampling operations. Each convolution operation performs convolution calculations on the attention output feature, extracting local patterns, and then performs downsampling through pooling or stride convolution to progressively compress the sequence length. After multi-layer distillation encoding, the sequence length is compressed from the original hundreds of time steps to tens of time steps, while the feature dimension increases and the level of abstraction is improved. The final output of the distillation encoder is a global feature representation vector, which contains cross-modal interaction information and global dependency features over a long time span.

[0072] The LSTM temporal refiner is located in the second stage of the cascaded model, and its function is to capture local short-term dynamic changes. In the LSTM refiner, this invention receives the global feature representation vector output by the Informer, expands it step-by-step, and uses it as the input sequence for the LSTM. The core of the LSTM is the gated recurrent unit, including the forget gate, input gate, and output gate. At each time step, this invention first calculates the output of the forget gate. The forget gate receives the hidden state from the previous time step and the input from the current time step, concatenates them, performs a linear transformation through the weight matrix, and then passes through the sigmoid activation function, outputting a forgetting coefficient between 0 and 1. The forgetting coefficient is multiplied element-wise by the memory cell state from the previous time step to determine how much historical information is retained. Subsequently, this invention calculates the output of the input gate. The input gate also receives the previous hidden state and the current input, performs a weight matrix transformation and sigmoid activation, and outputs an input coefficient between 0 and 1. This invention simultaneously calculates candidate memory cells, which are generated through the tanh activation function, with values ​​ranging from -1 to 1. The input coefficient is multiplied element-wise by the candidate memory cells to obtain the amount of new information written. This invention adds the output of the forget gate to the output of the input gate to update the memory cell state at the current time step. The memory cell state is then activated by tanh and multiplied element-wise by the output gate to obtain the hidden state output at the current time step. The calculation method of the output gate is similar to that of the forget gate and the input gate, obtaining the output coefficients through weight matrix transformation and sigmoid activation. This invention processes each element in the global feature representation vector sequentially according to the time step order, updating the memory cell state and the hidden state. After processing all time steps, this invention extracts the hidden state of the last time step as the refined feature representation.

[0073] The final stage of the cascaded model is the output layer. This invention constructs a linear mapping layer that receives the refined feature representation output by the LSTM. Through the linear transformation of the weight matrix and the bias vector, the refined features are mapped to the dissolved oxygen numerical space of prediction step size H time steps, and the predicted dissolved oxygen content values ​​for the next H time steps are output.

[0074] The data flow of the entire fusion model is as follows: the structured time series dataset is input into the Informer modality feature extractor, which is then processed by embedding, self-attention, and distillation encoding to output a global feature representation vector; the global feature representation vector is input into the LSTM time series refiner, processed by a gated recurrent unit to output a refined feature representation; the refined feature representation is input into the output layer, and after linear transformation, a dissolved oxygen content prediction sequence is generated.

[0075] S22: Divide the structured time series dataset to obtain a model development dataset including a training set, a validation set, and a test set.

[0076] In step S22, this invention partitions the structured time-series dataset according to chronological order, strictly adhering to the temporal sequence to avoid data leakage. Assuming the structured time-series dataset contains data over T time steps, this invention first calculates the time step number corresponding to 70%, i.e., 0.7T, and partitions all data from time step 1 to time step 0.7T into the training set. Subsequently, this invention calculates the time step number corresponding to 85%, i.e., 0.85T, and partitions all data from time step 0.7T+1 to time step 0.85T into the validation set. Finally, this invention partitions all data from time step 0.85T+1 to time step T into the test set. During the partitioning process, this invention maintains the temporal continuity of the data within each subset: the training set contains the earliest historical data, the validation set contains data from intermediate time periods, and the test set contains data from the most recent time period. After partitioning, this invention obtains a model development dataset, which includes three independent subsets: a training set for model parameter learning, a validation set for hyperparameter tuning and training monitoring, and a test set for final performance evaluation.

[0077] S23: Develop training samples based on the model and input them into the fusion model.

[0078] In step S23, the present invention uses a sliding window method to construct training samples from the training set. First, the present invention sets the input window length L and the prediction step size H. The input window length determines how many historical time steps of data are used for each sample, and the prediction step size determines how many future time steps of dissolved oxygen values ​​are predicted. The present invention sets the sliding step size to 1, meaning that the sliding window moves backward by one time step each time.

[0079] This invention starts from the Lth time step of the training set and establishes the first training sample. The input sequence of this training sample is multimodal data from the 1st time step to the Lth time step, with each time step containing values ​​for all feature dimensions such as water temperature, pH, air pressure, and humidity, forming an L × feature dimension matrix. The target output sequence of this training sample is the true dissolved oxygen values ​​from the (L+1)th time step to the (L+H)th time step, forming a vector of length H. This invention records this training sample as the first sample pair, containing the input matrix and the output vector.

[0080] Subsequently, the present invention moves the sliding window forward one time step to establish a second training sample. The input sequence of this sample is from the second time step to the (L+1)th time step, and the output sequence is from the (L+2)th time step to the (L+H+1)th time step. The present invention continues to slide the window and extract samples until the end of the window's output sequence reaches the last time step of the training set.

[0081] Using a sliding window method, this invention generates thousands or even tens of thousands of training sample pairs from the training set. These samples overlap, with adjacent samples sharing most of their input sequences at the same time step, differing only at the beginning and end. Subsequently, this invention organizes all training samples into batches, each batch containing several samples, with the batch size determined by computational resources. This invention inputs the training samples into the fusion model batch by batch, with each batch's input being a three-dimensional tensor, the dimension of which is batch size × input window length × number of feature dimensions.

[0082] S24: Based on the preset loss function, the fusion model is optimized and trained using the Adam optimizer to obtain the prediction model.

[0083] In step S24, the expression for the loss function is:

[0084] ;

[0085] in, For loss function, The total number of training samples, For training sample index values, To predict the total number of time steps, For the current time step, For time step index, For the first training samples Predicted dissolved oxygen content at any given time. For the first training samples The actual dissolved oxygen content at any given time.

[0086] In step S24, this invention uses mean squared error as the loss function to measure the difference between the model's predicted value and the true value. For each batch input into the fusion model, the model forward propagates to obtain the predicted output, which represents the predicted dissolved oxygen content of a sample at a future time step. This invention also extracts the true value corresponding to that batch, and the dimension of the true value is also batch size × prediction step size.

[0087] This invention calculates the difference between the predicted and actual values ​​element by element, squares the difference, sums the results over all prediction time steps, and then averages the sum over all samples within the batch to obtain the mean squared error loss value for that batch. The loss value is a scalar that reflects the current predictive performance of the model; a smaller loss value indicates a more accurate prediction.

[0088] This invention then performs backpropagation to calculate the gradient of the loss value with respect to each parameter in the fusion model. The gradient represents the direction and magnitude of the influence of a small change in the parameter on the loss value. The Adam optimizer receives the gradient information and maintains the first-moment and second-moment estimates for each parameter. The first-moment estimate is an exponential moving average of the gradient, and the second-moment estimate is an exponential moving average of the squared gradient. The Adam optimizer calculates the adaptive learning rate for each parameter based on the first and second moments; a larger learning rate indicates a larger parameter update step size. Subsequently, this invention multiplies the gradient by the adaptive learning rate to obtain the parameter update amount, which is then added to the original parameter to complete one parameter update.

[0089] During updates, this invention also sets the number of training rounds. Each round iterates through all batches of the training set, and for each batch, it performs a complete process of forward propagation, loss calculation, backpropagation, and parameter update. During training, this invention periodically inputs the validation set into the model, calculates the loss value on the validation set, and monitors the trend of the validation set loss. When the validation set loss stops decreasing or begins to increase for several consecutive rounds (e.g., 5 rounds), this invention determines that the model training has converged or overfitting has occurred, stops the training process, saves all parameters of the current model, and obtains the trained prediction model.

[0090] S3: Using the prediction model, predict the time series data to be predicted to obtain the dissolved oxygen content prediction sequence.

[0091] In step S3, the expression for predicting the time series data to be predicted using the prediction model is as follows:

[0092] ;

[0093] in, This indicates that the Informer modal feature extractor performs data processing. This indicates that the LSTM timing refiner performs data processing. Input window length, To predict the step size, For the past A multimodal input sequence at time steps. For the predicted future The predicted dissolved oxygen content sequence at each time step.

[0094] In practical applications, this invention receives time-series data to be predicted, which includes water quality modal data and meteorological modal data for the current time and L-1 previous time steps, totaling L time steps, forming an input sequence. The dimension of the input sequence is L × the number of feature dimensions, consistent with the sample input format in the training phase.

[0095] After obtaining the data to be predicted, this invention inputs the time-series data to be predicted into a trained prediction model. The data first enters the embedding layer of the Informer modality feature extractor, where it undergoes a linear transformation of the weight matrix to obtain the embedded feature matrix. The embedded feature matrix then enters the ProbSparse self-attention layer, generating a query matrix, a key matrix, and a value matrix. A sparse attention score is calculated and multiplied by the value matrix to obtain the attention output feature. The attention output feature enters the distillation encoder, where it undergoes multiple convolutions and downsampling, gradually compressing the sequence length and gradually increasing the feature dimension, ultimately outputting a global feature representation vector.

[0096] The global feature representation vector is then fed into the LSTM temporal refiner, processed sequentially step by step. At each time step, the states of the forget gate, input gate, and output gate are updated, as are the memory cells and hidden states. After processing all time steps, this invention extracts the hidden state of the last time step as the refined feature representation.

[0097] Subsequently, the refined feature representation enters the output layer. After a linear transformation using a weight matrix and bias vector, H numerical values ​​are generated, each corresponding to a predicted dissolved oxygen content at a future time step. The dissolved oxygen content prediction sequence output by this invention is a vector of length H. The first element of the vector represents the predicted dissolved oxygen content at the next time step (5 minutes after the current time), the second element represents the predicted dissolved oxygen content at the next time step after that (10 minutes after the current time), and so on. After the prediction sequence is output, aquaculture managers can determine whether to activate aeration equipment or adjust stocking density based on the predicted dissolved oxygen change trend, achieving precise management.

[0098] like Figure 3 As shown, the present invention also provides a pond dissolved oxygen content prediction system based on cross-modal information fusion, comprising:

[0099] Data processing module 100: used to collect multi-source time-series data from the pond and perform preprocessing to obtain a structured time-series dataset;

[0100] Construction training module 200: used to train a pre-built fusion model using the structured time series dataset to obtain a prediction model;

[0101] The prediction output module 300 is configured to use the prediction model trained by the construction training module 200 to predict the time series data to be predicted and obtain the dissolved oxygen content prediction sequence.

[0102] The present invention also provides a pond dissolved oxygen content prediction device based on cross-modal information fusion, comprising: a memory and at least one processor, wherein the memory stores instructions; at least one processor invokes the instructions in the memory to cause the pond dissolved oxygen content prediction device based on cross-modal information fusion to perform a pond dissolved oxygen content prediction method based on cross-modal information fusion as described above.

[0103] The present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement a method for predicting pond dissolved oxygen content based on cross-modal information fusion as described in any of the preceding claims.

[0104] This invention proposes a prediction model that combines the advantages of Informer in processing long sequences and capturing global dependencies with the advantages of LSTM in capturing local details and short-term dynamics through a cascaded structure of Informer and LSTM. This results in a collaborative division of labor between capturing macro trends and fitting micro fluctuations, significantly improving the model's overall prediction capability. In addition, this invention utilizes the self-attention mechanism of Informer to deeply mine the complex nonlinear relationships within water quality modes, meteorological modes, and between the two at the front end, achieving true deep cross-modal information fusion.

[0105] Thanks to the above two points, the method of this invention has higher prediction accuracy and better robustness for both long-term trends and short-term fluctuations in dissolved oxygen, providing more reliable technical support for the refined management of aquaculture; moreover, the model structure is clear, the process is well-defined, and it is easy to integrate into existing aquaculture IoT platforms, thus possessing good prospects for industrial application.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting dissolved oxygen content in ponds based on cross-modal information fusion, characterized in that, include: S1: Collect multi-source time-series data from the pond and preprocess it to obtain a structured time-series dataset; S2: Construct a fusion model, train the fusion model using the structured time-series dataset, and obtain a prediction model; S3: Using the prediction model, predict the time series data to be predicted to obtain the dissolved oxygen content prediction sequence.

2. The method for predicting pond dissolved oxygen content based on cross-modal information fusion according to claim 1, characterized in that, The multi-source time-series data in step S1 includes: Water quality modal data, including water temperature data, historical dissolved oxygen data, and pH data; Meteorological modal data, which includes temperature data, wind force data, wind speed data, wind direction data, humidity data, atmospheric pressure data, and current rainfall data.

3. The method for predicting dissolved oxygen content in ponds based on cross-modal information fusion according to claim 1, characterized in that, Step S1 further includes: S11: Collect multi-source time-series data of the pond; S12: Calculate the statistical distribution parameters of multiple feature dimensions in the multimodal time series data, identify abnormal data points through an anomaly detection algorithm based on standard deviation threshold, and obtain an anomaly label sequence; S13: Correct the abnormal data points according to the abnormal labeling sequence, replace minor abnormal values ​​with the time neighborhood mean, and mark extreme abnormal values ​​as null values ​​to obtain preliminary cleaned data; S14: Apply a sliding window smoothing algorithm to the preliminary cleaned data to reduce noise and obtain smooth time series data; S15: For missing values ​​in the smoothed time series data, an interpolation strategy is selected based on the missing length to fill in the missing values, thereby obtaining continuous and complete data; S16: Perform interval normalization transformation on multiple feature dimensions of the continuous complete data to obtain a structured time series dataset.

4. The method for predicting pond dissolved oxygen content based on cross-modal information fusion according to claim 1, characterized in that, Step S2 further includes: S21: Construct a fusion model; S22: Divide the structured time series dataset to obtain a model development dataset including a training set, a validation set, and a test set; S23: Develop training samples based on the model dataset and input them into the fusion model; S24: Based on the preset loss function, the fusion model is optimized and trained using the Adam optimizer to obtain the prediction model.

5. The method for predicting pond dissolved oxygen content based on cross-modal information fusion according to claim 4, characterized in that, The fusion model in step S21 is a cascaded model composed of an Informer modal feature extractor and an LSTM temporal refiner connected in series. The Informer modal feature extractor is used to extract cross-modal global dependency features, and the LSTM temporal refiner is used to capture local short-term dynamic change features.

6. The method for predicting pond dissolved oxygen content based on cross-modal information fusion according to claim 4, characterized in that, In step S24, the expression for the loss function is: ； in, For loss function, The total number of training samples, For training sample index values, To predict the total number of time steps, For the current time step, For time step index, For the first training samples Predicted dissolved oxygen content at any given time. For the first training samples The actual dissolved oxygen content at any given time.

7. The method for predicting pond dissolved oxygen content based on cross-modal information fusion according to claim 1, characterized in that, In step S3, the expression for predicting the time series data to be predicted using the prediction model is: ； in, This indicates that the Informer modal feature extractor performs data processing. This indicates that the LSTM timing refiner performs data processing. Input window length, To predict the step size, For the past A multimodal input sequence at time steps. For the predicted future The predicted dissolved oxygen content sequence at each time step.

8. A pond dissolved oxygen content prediction system based on cross-modal information fusion, characterized in that, include: Data processing module: used to collect multi-source time-series data from the pond and perform preprocessing to obtain a structured time-series dataset; The training module is used to train a pre-built fusion model using the structured time-series dataset to obtain a prediction model. A prediction output module is configured to use the prediction model trained by the construction training module to predict the time series data to be predicted and obtain a dissolved oxygen content prediction sequence.

9. A pond dissolved oxygen content prediction device based on cross-modal information fusion, characterized in that, include: A memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause a pond dissolved oxygen content prediction device based on cross-modal information fusion to perform a pond dissolved oxygen content prediction method based on cross-modal information fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a processor, implement a method for predicting pond dissolved oxygen content based on cross-modal information fusion as described in any one of claims 1 to 7.

Images