A fish growth rate prediction method and system based on environment modeling

Abstract

The application provides a fish growth rate prediction method and system based on environment modeling, comprising: acquiring environment data; based on the environment data, constructing an adaptive time mode network to obtain a water temperature prediction reasoning model; inputting fish experience data to be predicted into the water temperature prediction reasoning model to obtain fish growth rate prediction results to be predicted. The application constructs a water temperature prediction reasoning model by collecting comprehensive environment data, combines different fish growth experience values, obtains prediction results of different fish growth rates, has the advantages of wide adaptation range, comprehensive environmental factors, and no need to repeatedly train the prediction model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, and in particular to a fish growth rate prediction method and system based on environment modeling. BACKGROUND

[0002] With the change of population's dietary structure, more and more people take fish meat as the main source of protein, which promotes the development of aquaculture. Aquaculture systems mainly have open, semi-intensive or intensive, which depends on the breeding density of fish population. The recirculating aquaculture system (RAS) as a controllable environment breeding method can realize the control of growth rate and harvest cycle. In addition, the recirculating aquaculture system realizes water reuse through biological filtration, effectively reduces the loss of heat energy and water resources, and the required water volume is reduced by 90%-99% compared with traditional methods, and the waste treatment is pollution-free, which is an environmentally friendly and efficient breeding mode. In the recirculating aquaculture system, accurately grasping and perceiving the growth rate of fish is very important for effective production, and the growth rate of fish determines the required feeding amount, and the feeding amount determines the waste production load and oxygen consumption. Therefore, the prediction of the growth rate can help the practitioners to effectively prepare the required feeding bait in advance and predict the fish pond waste and other environmental factor states.

[0003] At present, there are related schemes using Logistic, Gompertz, Bertallanffy and other statistical models to directly model and analyze fish growth. However, it can only obtain the fish growth rate at a certain moment in an ideal scenario, and cannot be applied to the recirculating aquaculture system with a large number of human interference factors. In order to adapt to the production scene according to individualization, some researchers use deep learning to develop a growth model based on fish growth history data. However, it is difficult to collect fish growth history data, and it is difficult to popularize. In the recirculating aquaculture, the normal behavior and growth of fish are affected by water quality environmental factors. For example, insufficient dissolved oxygen will cause the death of fish, excessive ammonia content will increase the toxicity of water body, and extreme pH has stress and lethality. Among them, temperature directly affects the food acquisition, digestive enzyme activity and food conversion rate of fish, and has important significance for fish behavior, growth and reproduction. Therefore, there is a strong correlation between water temperature and fish growth rate. SUMMARY

[0004] The present application provides a fish growth rate prediction method and system based on environment modeling, which solves the defect that there is no system to predict fish growth rate through environmental water temperature in the prior art.

[0005] In a first aspect, the present application provides a fish growth rate prediction method based on environment modeling, comprising:

[0006] acquire environment data;

[0007] construct an adaptive time pattern network based on the environment data to obtain a water temperature prediction inference model;

[0008] input fish experience data to be predicted into the water temperature prediction inference model to obtain a fish growth rate prediction result to be predicted.

[0009] According to the fish growth rate prediction method based on environment modeling provided by the application, the environment data is acquired, which comprises:

[0010] The sensor in LoRa mode star network collects environment monitoring data at a preset time interval, and the environment monitoring data comprises water temperature, ambient temperature and humidity around the water body, wall temperature, illumination, CO2 concentration and water dissolved oxygen.

[0011] The sensor transmits the environment monitoring data to the cloud through a preset communication network, the cloud integrates the moving average autoregressive model ARIMA algorithm to interpolate missing values in the environment monitoring data, and adopts wavelet transform to smooth and denoise the interpolated environment monitoring data, to obtain the environment data stored in the database in the form of a message queue.

[0012] According to the fish growth rate prediction method based on environment modeling provided by the application, the adaptive time pattern network is constructed based on the environment data to obtain a water temperature prediction inference model, which comprises:

[0013] Determine the exogenous environmental factor, target environmental factor and historical time stamp of the environment data, and obtain the preset deep feature of multiple time patterns of different environmental factors through deep time sequence feature DFT;

[0014] Based on the multi-time pattern convolution method, the preset deep feature and the preset shallow feature of the environment data are spliced to obtain complete environment features;

[0015] Based on the spatial attention mechanism, the preset deep feature and the preset shallow feature are learned end to end to obtain attention weights, and the complete environment features are processed based on the attention weights to obtain the water temperature prediction inference model.

[0016] According to the fish growth rate prediction method based on environment modeling provided by the application, the exogenous environmental factor, target environmental factor and historical time stamp of the environment data are determined, and the preset deep feature of multiple time patterns of different environmental factors is obtained through deep time sequence feature DFT, which comprises:

[0017] Splicing the exogenous environmental factor and the target environmental factor obtains a first channel of a DFT input three-dimensional data structure;

[0018] The historical time timestamp is repeated several times per time step to obtain a second channel, a third channel and a fourth channel of the DFT input three-dimensional data structure;

[0019] A multi-layer time pattern MTP convolution module is superimposed on the DFT to adjust the intermediate layer network channel dimension through one-dimensional convolution and add a residual convolution to realize control of the network layer number;

[0020] For each MTP convolution module, a fusion multiple different dilation rate dilation convolution kernel MD convolution, a rectified linear unit ReLU activation function and a Dropout layer are added to obtain the preset deep-level feature.

[0021] According to the fish growth rate prediction method based on environment modeling provided by the application, the preset deep-level feature is obtained by adding a fusion multiple different dilation rate dilation convolution kernel MD convolution, a rectified linear unit ReLU activation function and a Dropout layer to each MTP convolution module, including:

[0022] The output channel dimension and the feature map time sequence length of any layer model MD convolution module are obtained;

[0023] Based on the output channel dimension and the feature map time sequence length, the feature map and the hidden layer feature output by any layer model are obtained;

[0024] The hidden layer feature of the current layer is connected in residual with the feature map of the previous layer through the one-dimensional convolution to obtain the feature map of the current layer;

[0025] The multi-layer feature map is output and stretched to obtain the preset deep-level feature.

[0026] According to the fish growth rate prediction method based on environment modeling provided by the application, the preset deep-level feature is obtained by adding a fusion multiple different dilation rate dilation convolution kernel MD convolution, a rectified linear unit ReLU activation function and a Dropout layer to each MTP convolution module, including:

[0027] A plurality of different expansion factor convolution kernels are added to each layer in the time convolution network TCN to obtain a multi-time pattern convolution network;

[0028] The preset deep-level feature and the preset shallow-level feature are respectively processed through the multi-time pattern convolution network to obtain a multi-time pattern preset deep-level feature and a multi-time pattern preset shallow-level feature;

[0029] The multi-time mode preset deep-level feature and the multi-time mode preset shallow-level feature are fused and spliced to obtain the complete environment feature.

[0030] According to the fish growth rate prediction method based on environment modeling provided by the application, the preset deep-level feature and the preset shallow-level feature are learned end to end based on the spatial attention mechanism, attention weights are obtained, the complete environment feature is processed based on the attention weights, and the water temperature prediction reasoning model is obtained, including:

[0031] Based on the preset deep-level feature and the preset shallow-level feature, any environment factor context information and a plurality of attention mechanism learning parameters are obtained.

[0032] Based on the any environment factor context information and the plurality of attention mechanism learning parameters, initial attention weights of each environment factor are obtained.

[0033] The initial attention weights of each environment factor are processed based on a Softmax function to obtain attention weights of each environment factor.

[0034] A plurality of reasoning model learning parameters are obtained according to a prediction future water temperature step number, the complete environment feature is trained based on the attention weights of each environment factor and the plurality of reasoning model learning parameters, and a preset loss function is used to obtain the water temperature prediction reasoning model.

[0035] According to the fish growth rate prediction method based on environment modeling provided by the application, the fish experience data to be predicted is input into the water temperature prediction reasoning model to obtain a fish growth rate prediction result to be predicted, including:

[0036] A plurality of temperature experience data of the fish to be predicted are determined, and a prediction temperature value of the water temperature prediction reasoning model is obtained.

[0037] The prediction temperature value and a maximum temperature value in the plurality of temperature experience data are compared, and a smaller temperature value is used for calculation to obtain the fish growth rate prediction result to be predicted.

[0038] In a second aspect, the application further provides a fish growth rate prediction system based on environment modeling, including:

[0039] An acquisition module is configured to acquire environment data.

[0040] A construction module is configured to construct an adaptive time mode network based on the environment data to obtain a water temperature prediction reasoning model.

[0041] A prediction module is configured to input fish experience data to be predicted into the water temperature prediction reasoning model to obtain a fish growth rate prediction result to be predicted.

[0042] In a third aspect, the present application also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the fish growth rate prediction method based on environment modeling according to any one of the above aspects when executing the program.

[0043] In a fourth aspect, the present application also provides a non-transitory computer readable storage medium, which stores a computer program, wherein the computer program is executable on a processor to implement the steps of the fish growth rate prediction method based on environment modeling according to any one of the above aspects.

[0044] In a fifth aspect, the present application also provides a computer program product, comprising a computer program, wherein the computer program is executable on a processor to implement the steps of the fish growth rate prediction method based on environment modeling according to any one of the above aspects.

[0045] The fish growth rate prediction method and system based on environment modeling provided by the present application have the advantages of wide adaptability, comprehensive coverage of environmental factors, and no need for repeated training of the prediction model, by constructing a water temperature prediction inference model through the collection of comprehensive environmental data, and combining different fish growth experience values to obtain prediction results of different fish growth rates. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0047] Figure 1 is a flowchart of the fish growth rate prediction method based on environment modeling provided by the present application;

[0048] Figure 2 is a whole technical block diagram provided by the present application;

[0049] Figure 3 is a data collection and processing flowchart provided by the present application;

[0050] Figure 4 is an ATPNet model architecture diagram provided by the present application;

[0051] Figure 5 is a deep-level time sequence feature module architecture diagram provided by the present application;

[0052] Figure 6 is a multi-dilation factor convolution structure diagram provided by the present application;

[0053] Figure 7 is a structural schematic diagram of a fish growth rate prediction system based on environment modeling provided by the present application;

[0054] Figure 8 is a structural schematic diagram of an electronic device provided by the present application. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0056] In view of the demand for fish growth rate prediction in a recirculating aquaculture system, considering that the collection technology of environmental data is relatively mature and feasible, and better trade-off between cost and precision and other aspects can be achieved, the present application proposes a fish growth rate prediction method based on environment modeling, adopts a deep learning method to construct a high-precision water temperature prediction model, combines an empirical formula to predict the fish growth rate, and also comprehensively considers the historical water temperature in the current system, the ambient temperature and humidity around the water body, the wall temperature, the illuminance and other exogenous environmental factors, so as to obtain real-time and adaptive scene fish growth rate prediction values for guiding aquaculture practitioners to make better decisions.

[0057] Figure 1 is a flowchart of a fish growth rate prediction method based on environment modeling provided by the present application, as shown in Figure 1 , comprising:

[0058] Step S1, acquiring environmental data;

[0059] Step S2, constructing an adaptive time pattern network based on the environmental data to obtain a water temperature prediction inference model;

[0060] Step S3, inputting the to-be-predicted fish experience data into the water temperature prediction inference model to obtain a to-be-predicted fish growth rate prediction result.

[0061] Specifically, as shown in the overall technical block diagram of the present application, Figure 2 firstly, environmental data is acquired through a data acquisition and preprocessing module, an adaptive time pattern network is constructed by using a water temperature prediction inference module based on the acquired environmental data to obtain a water temperature prediction inference model, to-be-predicted fish experience data is input into the water temperature prediction inference model according to the type of fish, and a to-be-predicted fish growth rate prediction result is obtained by a growth rate conversion module.

[0062] It should be noted that the multi-step real-time fish growth rate of the present application is achieved by the water temperature reasoning model and the growth rate conversion module, without the need for preparing fish growth rate historical data in advance under specific production environment, through monitoring of surrounding environmental factors, greatly reducing the experimental cost during model construction, and since the surrounding environment is considered, the growth rate predicted by the model is more real-time and accurate.

[0063] The accurate prediction method for the thermal environment proposed in the present application, since the time patterns of different circulating water surrounding environmental factors are not the same, for example, the specific heat capacity of the thermal environment is larger, so its change is relatively slow, the time delay is larger, the model needs greater perception ability; and the change of dissolved oxygen is more sensitive, and the model needs to pay attention to the fine-grained change. Therefore, the adaptive interval mode mining method proposed in the present application can learn the respective time pattern features end to end according to different environmental factors, and increase the perception and adaptation ability of the model to the coarse and fine granularity of the environmental factors.

[0064] The model proposed in the present application comprehensively considers the deep and shallow features of the exogenous environmental factors: in order to improve the generalization ability of the model, the deep features mined by the adaptive time pattern module and the original shallow features of the environmental factors are comprehensively considered, which makes the model have more complex and deeper structure, improves the processing ability of the deep feature module, and avoids overfitting.

[0065] The present application realizes the prediction of the growth rate of different fish: the fish growth rate conversion module dynamically converts the circulating water temperature reasoning result into the fish growth rate according to the empirical value, so that the growth rate of different fish can be obtained by changing the empirical value according to different fish, without the need for retraining the model.

[0066] Based on the above embodiment, the method step S1 comprises:

[0067] The sensor adopts LoRa mode star networking, collects environmental monitoring data at a preset time interval, and the environmental monitoring data includes water temperature, water body surrounding environment temperature and humidity, wall temperature, illuminance, CO2 concentration and water body dissolved oxygen;

[0068] The sensor transmits the environmental monitoring data to the cloud through a preset communication network, the cloud integrates the moving average autoregressive model ARIMA algorithm, interpolates the missing values in the environmental monitoring data, and adopts wavelet transform to smooth and denoise the interpolated environmental monitoring data, to obtain the environmental data stored in the database in the message queue.

[0069] Specifically, as Figure 3As shown, LoRa (Long Range) is used for networking the sensors used in the circulating water data acquisition module, adopts a star-shaped topology with good fault tolerance, and the status of each monitoring module is equal, so that even if any node fails, it will not affect the normal work of other parts. The data monitored by the module include water temperature, ambient temperature and humidity of water body, wall temperature, illumination, CO2 concentration and water dissolved oxygen, and the data acquisition interval is generally set to 10 minutes.

[0070] The LoRa concentrator uploads sensor data to the cloud through a 4G network, the cloud uses an ARIMA (Autoregressive Integrated Moving Average model) algorithm to interpolate missing values in the data collection process, and performs wavelet transform on the data to smooth and denoise, and finally the data is real-time dropped into the database, so as to perform the next model reasoning and growth rate prediction.

[0071] The present application adopts LoRa networking to collect environmental data in real time, without the need to prepare fish growth rate historical data in advance under specific production environment, and uses the cloud to preprocess data to ensure the robustness of the data.

[0072] Based on any of the above embodiments, the method step S2 comprises:

[0073] Determine the exogenous environmental factor, target environmental factor and historical time stamp of the environmental data, obtain the preset deep feature of multiple time modes of different environmental factors through deep time sequence feature DFT;

[0074] Based on the multi-time mode convolution method, the preset deep feature and the preset shallow feature of the environmental data are spliced to obtain complete environmental features;

[0075] Based on the spatial attention mechanism, the preset deep feature and the preset shallow feature are learned end to end to obtain attention weights, and the complete environmental features are processed based on the attention weights to obtain the water temperature prediction reasoning model.

[0076] The determination of the exogenous environmental factor, the target environmental factor and the historical time stamp of the environmental data, and the obtaining of the preset deep feature of multiple time modes of different environmental factors through deep time sequence feature DFT comprise:

[0077] Splice the exogenous environmental factor and the target environmental factor to obtain a first channel of a DFT input three-dimensional data structure;

[0078] The historical time stamp is repeated several times according to each time step to obtain a second channel, a third channel and a fourth channel of the DFT input three-dimensional data structure.

[0079] Superimpose a multi-layer time pattern MTP convolution module on the DFT, adjust the intermediate layer network channel dimension through one-dimensional convolution, and add residual convolution to realize control of the number of network layers;

[0080] For each layer of the MTP convolution module, add a dilated convolution kernel MD convolution that fuses multiple different expansion rates, a rectified linear unit ReLU activation function, and a Dropout layer to obtain the preset deep-level feature.

[0081] For each layer of the MTP convolution module, add a dilated convolution kernel MD convolution that fuses multiple different expansion rates, a rectified linear unit ReLU activation function, and a Dropout layer to obtain the preset deep-level feature.

[0082] Obtain the output channel dimension and feature map time sequence length of any layer model MD convolution module;

[0083] Based on the output channel dimension and feature map time sequence length, obtain the feature map and hidden layer feature output by any layer model;

[0084] Connect the hidden layer feature of the current layer and the feature map of the previous layer through the one-dimensional convolution to obtain the feature map of the current layer;

[0085] Output and stretch the multi-layer feature maps to obtain the preset deep-level feature.

[0086] Based on the multi-time pattern convolution method, splice the preset deep-level feature and the preset shallow-level feature of the environmental data to obtain the complete environmental feature, including:

[0087] Add multiple different expansion factor convolution kernels to each layer in the time convolution network TCN to obtain a multi-time pattern convolution network;

[0088] Process the preset deep-level feature and the preset shallow-level feature through the multi-time pattern convolution network respectively to obtain a multi-time pattern preset deep-level feature and a multi-time pattern preset shallow-level feature;

[0089] Fuse and splice the multi-time pattern preset deep-level feature and the multi-time pattern preset shallow-level feature to obtain the complete environmental feature.

[0090] Based on the spatial attention mechanism, perform end-to-end learning on the preset deep-level feature and the preset shallow-level feature to obtain attention weights, process the complete environmental feature based on the attention weights, and obtain the water temperature prediction reasoning model, including:

[0091] Based on the preset deep features and the preset shallow features, obtain the context information of any environmental factor and several attention mechanism learning parameters;

[0092] Based on the context information of any one of the environmental factors and the learning parameters of the several attention mechanisms, the initial attention weights of each environmental factor are obtained.

[0093] The initial attention weights of each environmental factor are processed using the Softmax function to obtain the attention weights of each environmental factor.

[0094] Based on the number of steps to predict future water temperature, several inference model learning parameters are obtained. Based on the attention weights of each environmental factor and the several inference model learning parameters, and using a preset loss function, the complete environmental features are trained to obtain the water temperature prediction inference model.

[0095] Specifically, since the accuracy of circulating water temperature inference determines the accuracy of the final fish growth rate prediction, this invention fully considers the temporal characteristics of circulating water temperature and proposes an Adaptive Time Pattern Network (ATPNet). Figure 4 As shown in the ATPNet model architecture diagram, the exogenous environmental factor u∈R in the recirculating aquaculture system T×(N-1) (Ambient temperature and humidity of the water body, wall temperature, illuminance, CO2 concentration, dissolved oxygen in the water body) and target environmental factor v∈R T×1 Historical data on circulating water temperature is used as input to the model to uncover its spatiotemporal relationships and predict future circulating water temperatures. Furthermore, since changes in the water temperature factor are strongly correlated with different times of day, a timestamp τ = [τ...] is introduced for each historical moment. m ,τ d ,τ h ]∈R T×3 (τ m ,τ d ,τ h These represent the time characteristics of month, day, and hour, respectively, further enhancing the effective information in the input data. The data undergoes processing by the Deep Temporal Feature (DTF) module to obtain deep feature representations of various time patterns of different environmental factors in the data, m = [m1, m2, ..., m...]. N ] T ∈R N×F (Where, F represents the feature dimension output by the DTF module). Then, the deep-level feature m is compared with the original shallow-level feature o = [u...]. 1...T ,v 1...T ] T ∈RN×T Concatenate to get the complete feature expression c = [m 1...N ,o 1...N ]∈R N×(F+T) .

[0096] In order to make the model better mine the feature correlation of environmental factors, the application introduces a spatial attention mechanism (SAM), so that the importance of environmental factors to target prediction can be adaptively mined through end-to-end learning. And the weighted features flow into the output module, and finally the prediction result is obtained.

[0097] Further, in order to adaptively mine various time patterns of different environmental factors, the application designs a deep time sequence feature DTF module as shown in Figure 5 Firstly, the input three-dimensional structure of the model is reconstructed by using exogenous environmental factors, target environmental factors and timestamp information. Specifically, the exogenous environmental factors and the target environmental factors are spliced as a channel x1 = [u, v] T ∈R N×T in the input three-dimensional structure, the timestamp information is repeated N times according to each time step, thereby serving as the remaining 3 channels x 2|3|4 = [τ m|d|h ,τ m|d|h ,…,τ m|d|h ] T ∈R N×T , and finally the data structure for DTF module input is X = [x1; x2; x3; x4]∈R 4×N×T . The DTF module combines multiple time pattern (MTP) convolution modules to more flexibly combine and mine time patterns. In order to reduce the overfitting of the model, the application adjusts the channel dimension of the intermediate layer network through 1*1 convolution, thereby adding a residual connection, so that the model can control the number of network layers in the end-to-end learning process. For each layer of MTP module, the feature is expressed in multiple time patterns by fusing multiple dilated convolution kernels with different dilation rates (multi-scale dilated convolution, MD convolution). In order to increase the generalization of the model, the application sets ReLU (rectified linear unit) activation function and Dropout layer in each layer of MTP. The calculation process of each layer of MTP module is as follows:

[0098] h l = dropout (relu (f md (z l-1 ))) (1)

[0099] z l = h l * f 1×1 + z l-1 (2)

[0100] wherein, is the feature map and hidden layer feature output by the lth layer model, C l , T l are the output channel dimension and feature map time sequence length of the MD convolution module f md (·) of the lth layer model respectively. The hidden output h 1×1 is connected in residual with the feature map z l of the previous layer through a 1x1 convolution kernel f l-1 , so as to obtain the feature map z l of the current layer. Finally, the DTF module outputs and stretches the top layer feature map after multi-layer multi-time mode convolution, and finally obtains the deep multi-time mode feature m of the circulating water environment data, m e R N×F , F = T L x C L .

[0101] Figure 6 The implementation method of multi-time mode mining is shown, a multi-dilation factor convolution is adopted, and the dilation convolution strategy of a temporal convolutional network (TCN) is improved, so that the temporal convolutional network has the perception ability of different time modes of the environmental factors of the circulating water system. For a general TCN structure, as shown in (a), a large receptive field is obtained by gradually increasing the dilation factor of each layer. However, the size of each layer of the convolution kernel and the dilation factor are fixed, which causes the model to mine more single time modes of the environmental factors. The present application sets multiple convolution kernels with different dilation factors for each layer of convolution, as shown in (b), these convolution kernels represent the perceptrons of different time modes of the environmental factors, and the feature maps output by the multiple convolution kernels in each layer are used as different channels of the input features of the next layer network for further mining. Specifically, the calculation method of the multi-scale dilation convolution (MD Conv.) is as follows: Figure 6 Figure 6

[0102]

[0103] h = concat(z*f1, z*f2, z*f4, z*f8) (4)

[0104] wherein, for the i-th variable of the input feature map ​​(To simplify the description, it is assumed here that the channel of the input feature map is 1), the convolution method with expansion factor d, as formula (3); k is the size of the convolution kernel, t is the feature point in the target feature map. In time series data, t can be regarded as the current time, (t-dxj) ensures that only historical data will be used for convolution in the convolution process, thereby avoiding information leakage. Then, the convolution results with different expansion factors (d=1, 2, 4, 8) are spliced in the channel, as formula (4), so as to obtain the hidden layer output of the current layer The number of channels of each convolution output with different convolution factors.

[0105] Then, the model uses the spatial attention mechanism to process the deep features and the shallow features of the original data extracted by the deep time series feature module. The importance of each environmental factor to the target prediction can be learned end-to-end, so that the model can avoid the noise of the environmental factors with low correlation while not ignoring the influence of any external environmental factor. For the target factor v∈R T , the model extracts the context information c=[m 1...N ,o 1...N ]∈R N ×(F+T) The attention weight calculation method for the prediction impact is as follows:

[0106]

[0107]

[0108] Where c i ∈R (F+T) is the context information of the i-th environmental factor, W attn ∈R m×(F+T) , U attn ∈R m×T and v attn ∈R m are learning parameters of the attention mechanism, and a i is the attention weight of each environmental factor after Softmax processing. The final prediction output of the model is:

[0109]

[0110] W out ∈R H×(F+T) and b out ∈R H are model learning parameters, is the final prediction of the future H steps of the circulating water temperature. Since the model itself is smooth and derivable, the training loss function is defined as:

[0111]

[0112] Here, Θ represents the trainable parameters in the model, B represents the number of samples in each training batch, and ||·||2 is the L 2 norm used to predict the prediction.

[0113] The model provided by the application comprehensively considers deep and shallow features of exogenous environmental factors: in order to improve the generalization ability of the model, the deep features mined by the self-adaptive time mode module and the original shallow features of the environmental factors are also comprehensively considered, so that the model can have a more complex and deeper structure, improve the processing capacity of the deep feature module, and avoid overfitting.

[0114] Based on any of the above embodiments, the method step S3 comprises:

[0115] Determine a plurality of temperature empirical data of the fish to be predicted, and obtain a predicted temperature value of the water temperature prediction reasoning model;

[0116] Compare the predicted temperature value with the maximum temperature value in the plurality of temperature empirical data, determine to use the smaller temperature value for calculation, and obtain the fish growth rate prediction result.

[0117] Specifically, the application has the ability of multi-step prediction of fish growth rate through the reasoning result of circulating water temperature, that is, through the historical surrounding environment and on-site conditions of circulating water, the fish growth rate and its change curve in the next few hours or a day without human intervention are predicted, which has very meaningful reference value for the decision of producers. Specifically, the application converts the water temperature reasoning result into fish growth rate prediction value by using an empirical formula:

[0118]

[0119] In formula (9), the unit of fish growth rate FR is cm / month; T max , T base , and TU base are empirical values, which are different for different fish; if the water temperature is greater than T max , the growth rate is calculated according to the empirical value of T max , but too high water temperature will affect the growth of fish and the feed conversion rate, so it should be controlled in production. Table 1 shows the empirical values of four different fish, which are in ℃:

[0120] Table 1

[0121]

[0122] The present application predicts the growth rate of different fish, and the fish growth rate conversion module dynamically converts the circulating water temperature inference result into the fish growth rate according to the empirical value, so that the growth rate of different fish can be obtained by changing the empirical value according to different fish, without retraining the model.

[0123] The fish growth rate prediction system based on environment modeling provided by the present application is described below, and the fish growth rate prediction system based on environment modeling described below can be correspondingly referred to the fish growth rate prediction method based on environment modeling described above.

[0124] Figure 7 The fish growth rate prediction system based on environment modeling provided by the present application is described below, and the fish growth rate prediction system based on environment modeling described below can be correspondingly referred to the fish growth rate prediction method based on environment modeling described above. Figure 7 As shown in the structural schematic diagram of the fish growth rate prediction system based on environment modeling provided by the present application, it comprises an acquisition module 71, a construction module 72 and a prediction module 73, wherein:

[0125] The acquisition module 71 is used for acquiring environment data; the construction module 72 is used for constructing an adaptive time pattern network based on the environment data to obtain a water temperature prediction inference model; and the prediction module 73 is used for inputting empirical data of a fish to be predicted into the water temperature prediction inference model to obtain a growth rate prediction result of the fish to be predicted.

[0126] The present application constructs a water temperature prediction inference model by collecting comprehensive environment data, combines different fish growth empirical values, and obtains the prediction result of the growth rate of different fish, which has the advantages of wide adaptation range, comprehensive environmental factors, and no need to repeatedly train the prediction model.

[0127] Figure 8 An example of an entity structure schematic diagram of an electronic device is shown in Figure 8 As shown in the structural schematic diagram of the fish growth rate prediction system based on environment modeling provided by the present application, it comprises an acquisition module 71, a construction module 72 and a prediction module 73, wherein:

[0128] In addition, the logic instructions in the memory 830 described above can be implemented in the form of software functional units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0129] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the fish growth rate prediction method based on environment modeling provided by the above-mentioned methods, the method comprising: obtaining environment data; based on the environment data, constructing an adaptive temporal pattern network to obtain a water temperature prediction inference model; inputting experience data of a fish to be predicted into the water temperature prediction inference model to obtain a fish growth rate prediction result to be predicted.

[0130] In another aspect, the present application also provides a non-transitory computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the fish growth rate prediction method based on environment modeling provided by the above-mentioned methods, the method comprising: obtaining environment data; based on the environment data, constructing an adaptive temporal pattern network to obtain a water temperature prediction inference model; inputting experience data of a fish to be predicted into the water temperature prediction inference model to obtain a fish growth rate prediction result to be predicted.

[0131] The device embodiments described above are only schematic, wherein the units illustrated as separate components can or can not be physically separate, and the components illustrated as units can or can not be physical units, i.e., they can be located in one place, or distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment. Those skilled in the art can understand and implement without creative labor.

[0132] Those skilled in the art can clearly understand the technical solutions of the various embodiments from the above description of the embodiments, and the various embodiments can be implemented by means of software with the necessary general hardware platforms, and of course, can also be implemented by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part of the prior art that makes a contribution, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, and the like, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0133] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features therein; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting fish growth rate based on environmental modeling, characterized in that, include: Environmental data is acquired using sensors in a LoRa star network, which collect environmental monitoring data at preset time intervals. This environmental monitoring data includes water temperature, ambient temperature and humidity around the water body, wall temperature, and illuminance. CO 2 Concentration and dissolved oxygen in water; Based on the environmental data, an adaptive time-pattern network is constructed to obtain a water temperature prediction inference model, including: The exogenous environmental factors, target environmental factors, and historical time timestamps of the environmental data are determined, and preset deep-level features of various time patterns of different environmental factors are obtained through deep time series features (DFT). Based on the multi-temporal mode convolution method, the preset deep features are concatenated with the preset shallow features of the environmental data to obtain complete environmental features; The preset deep features and the preset shallow features are learned end-to-end based on the spatial attention mechanism to obtain attention weights. The complete environmental features are then processed based on the attention weights to obtain the water temperature prediction inference model. The empirical data of the fish species to be predicted are input into the water temperature prediction inference model to obtain the predicted growth rate of the fish species, specifically: Determine multiple empirical temperature data of the fish species to be predicted, and obtain the predicted temperature value of the water temperature prediction inference model; The predicted temperature value is compared with the maximum temperature value among the multiple empirical temperature data, and the smaller temperature value is used for calculation to obtain the predicted growth rate of the fish to be predicted, as shown in the formula: ,in, FR This represents the fish growth rate, expressed in cm / month. These are empirical values, in °C; if the water temperature is greater than the stated value... Then according to the above Growth rate can be calculated using empirical values, but excessively high water temperatures can affect fish growth and feed conversion rates, so this should be controlled during production.

2. The method for predicting fish growth rate based on environmental modeling according to claim 1, characterized in that, The sensor transmits the environmental monitoring data to the cloud via a preset communication network. The cloud uses the ARIMA algorithm to interpolate missing values ​​in the environmental monitoring data and employs wavelet transform to smooth and reduce noise in the interpolated environmental monitoring data, thereby obtaining the environmental data stored in the database in a message queue.

3. The method for predicting fish growth rate based on environmental modeling according to claim 1, characterized in that, The process involves determining the exogenous environmental factors, target environmental factors, and historical time stamps of the environmental data, and obtaining preset deep-level features of various time patterns for different environmental factors through deep time series feature analysis (DFT), including: The exogenous environmental factors and the target environmental factors are spliced ​​together to obtain the first channel of the DFT input three-dimensional data structure; The historical time stamps are repeated several times according to each time step to obtain the second, third and fourth channels of the DFT input three-dimensional data structure. Multi-layer temporal pattern MTP convolutional modules are superimposed on the DFT. The channel dimension of the intermediate network is adjusted by one-dimensional convolution, and residual convolution is added to control the number of network layers. For each MTP convolutional module, multiple dilated convolutional kernels (MD convolutions) with different dilation rates are added, along with a modified ReLU activation function and a Dropout layer, to obtain the preset deep features.

4. The method for predicting fish growth rate based on environmental modeling according to claim 3, characterized in that, For each MTP convolutional module, multiple dilated convolutional kernels (MD convolutions) with different dilation rates are added, along with a modified ReLU activation function and a Dropout layer, to obtain the preset deep-level features, including: Obtain the output channel dimension and feature map temporal length of the MD convolutional module in any layer of the model; Based on the output channel dimension and feature map temporal length, the feature map and hidden layer features of any layer model output are obtained; The hidden features of this layer are residually connected with the feature map of the previous layer through the one-dimensional convolution to obtain the feature map of this layer. The multi-layer feature map is output and stretched to obtain the preset deep-level features.

5. The method for predicting fish growth rate based on environmental modeling according to claim 1, characterized in that, The multi-temporal mode convolution method concatenates the preset deep features with the preset shallow features of the environmental data to obtain complete environmental features, including: By adding multiple convolutional kernels with different dilation factors to each layer of the temporal convolutional network (TCN), a multi-temporal mode convolutional network is obtained. The preset deep features and the preset shallow features are processed by the multi-temporal mode convolutional network to obtain multi-temporal mode preset deep features and multi-temporal mode preset shallow features. The multi-time pattern preset deep features and the multi-time pattern preset shallow features are fused and spliced ​​together to obtain the complete environmental features.

6. The method for predicting fish growth rate based on environmental modeling according to claim 1, characterized in that, The method involves end-to-end learning of the preset deep-level features and the preset shallow-level features based on a spatial attention mechanism to obtain attention weights. Based on these attention weights, the complete environmental features are processed to obtain the water temperature prediction inference model, which includes: Based on the preset deep features and the preset shallow features, obtain the context information of any environmental factor and several attention mechanism learning parameters; Based on the context information of any one of the environmental factors and the learning parameters of the several attention mechanisms, the initial attention weights of each environmental factor are obtained. The initial attention weights of each environmental factor are processed using the Softmax function to obtain the attention weights of each environmental factor. Based on the number of steps to predict future water temperature, several inference model learning parameters are obtained. Based on the attention weights of each environmental factor and the several inference model learning parameters, and using a preset loss function to train the complete environmental features, the water temperature prediction inference model is obtained.

7. The method for predicting fish growth rate based on environmental modeling according to claim 1, characterized in that, The step of inputting empirical data of the fish species to be predicted into the water temperature prediction inference model to obtain the predicted growth rate of the fish species includes: Determine several empirical temperature data of the fish species to be predicted, and obtain the predicted temperature value of the water temperature prediction inference model. The predicted temperature value is compared with the maximum temperature value among the several empirical temperature data, and the smaller temperature value is determined for calculation to obtain the predicted growth rate of the fish to be predicted.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the fish growth rate prediction method based on environmental modeling as described in any one of claims 1 to 7.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the fish growth rate prediction method based on environmental modeling as described in any one of claims 1 to 7.

Images