Deep Learning-Based Fault Prediction and Adaptive Temperature Control System and Method for Aquaculture Heat Pumps

Abstract

This invention relates to the field of aquaculture technology, specifically to a deep learning-based aquaculture heat pump fault prediction and adaptive temperature control system and method. The system includes data acquisition, feature extraction, intelligent analysis, decision control, and execution interaction modules. The data acquisition module collects temperature, pressure, flow rate, and power parameters of the heat pump system. The feature extraction module extracts time-domain, frequency-domain, time-frequency, and thermodynamic features from these parameters. The intelligent analysis module includes a fault prediction submodule, an energy efficiency optimization submodule, and a temperature control strategy submodule, used to predict faults and energy efficiency indicators, and generate a dynamic temperature difference setpoint. The decision control module generates a control strategy based on the fault prediction results and the temperature difference setpoint. The execution interaction module receives the control strategy and controls the heat pump system. This system improves water temperature control accuracy to ±0.5℃, far superior to the ±2.1℃ of traditional systems, meeting the aquaculture needs of high-value aquatic species.

Description

Technical Field

[0001] This invention relates to the field of aquaculture technology, specifically to a deep learning-based aquaculture heat pump fault prediction and adaptive temperature control system and method. Background Technology

[0002] In aquaculture, water temperature is a key factor affecting the growth rate and health of farmed organisms. Traditional aquaculture relies mainly on the natural environment, is easily affected by seasonal changes, has a limited breeding cycle, and makes it difficult to achieve stable production throughout the year. With the development of aquaculture, heat pump equipment is widely used in aquaculture environments to regulate and maintain suitable water temperatures.

[0003] However, existing aquaculture heat pump systems have the following problems: First, traditional heat pump control systems generally use simple on / off control or fixed-parameter PID control, resulting in low temperature control accuracy, typically only reaching ±2℃, which is insufficient to meet the precise temperature requirements of some high-value aquatic species; Second, energy utilization efficiency is low, and the lack of optimized control strategies leads to high energy consumption and high operating costs; Third, the heat pump system has insufficient fault prediction capability, and equipment failures often lead to temperature control interruptions, causing significant economic losses to aquaculture production.

[0004] Currently, fault diagnosis of heat pump systems mainly relies on manual experience and simple threshold judgments, lacking the ability to predict fault development trends and thus failing to achieve preventative maintenance. At the same time, existing temperature control systems generally lack adaptive capabilities, making it difficult to cope with the effects of environmental changes, equipment aging, and other factors, resulting in a decline in control performance over time.

[0005] Therefore, there is an urgent need for an intelligent heat pump system that can predict faults and control temperatures precisely to improve the stability and economic benefits of aquaculture. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based aquaculture heat pump fault prediction and adaptive temperature control system and method, which aims to solve the problems of low temperature control accuracy, low energy efficiency and lack of fault prediction capability in existing aquaculture heat pump systems.

[0007] This invention proposes a deep learning-based aquaculture heat pump fault prediction and adaptive temperature control system, comprising:

[0008] The data acquisition module is used to collect temperature, pressure, flow, and power parameters of the heat pump system.

[0009] The feature extraction module is communicatively connected to the data acquisition module and is used to receive parameters acquired by the data acquisition module and extract time-domain features, frequency-domain features, time-frequency features, and thermodynamic features.

[0010] An intelligent analysis module, communicatively connected to the feature extraction module, is used to receive features extracted by the feature extraction module. The intelligent analysis module includes:

[0011] The fault prediction submodule is used to establish a multi-scale convolutional-long short-term memory hybrid network model based on the features to predict the fault types and fault development trends of the heat pump system.

[0012] The energy efficiency optimization submodule is used to construct a dual-hidden-layer deep neural network model based on the features to predict the energy efficiency index of the heat pump system under different operating conditions.

[0013] The temperature control strategy submodule is used to generate a dynamic temperature difference setpoint based on the energy efficiency index.

[0014] A decision control module, communicatively connected to the intelligent analysis module, is used to receive the fault prediction results from the fault prediction submodule and the dynamic temperature difference setpoint from the temperature control strategy submodule, and generate a control strategy. The decision control module includes:

[0015] The power gradient calculation unit is used to calculate the rate of change of the total power of the heat pump system with respect to the temperature difference;

[0016] A temperature difference setting unit is used to determine the optimal temperature difference setting value based on the power gradient.

[0017] Fuzzy control unit is used to dynamically adjust PID control parameters based on temperature difference deviation, deviation change rate and system inertia estimate;

[0018] The execution interaction module is communicatively connected to the decision control module, and is used to receive the control strategy generated by the decision control module and control the actuators of the heat pump system.

[0019] Preferably, the multi-scale convolutional-long short-term memory hybrid network model includes:

[0020] The multi-scale convolutional layer consists of three parallel convolutional paths, each using a different kernel size for feature extraction.

[0021] A bidirectional long short-term memory layer is used to capture the forward and backward time dependencies of time-series data;

[0022] The attention layer is used to calculate weights for different time steps and feature channels;

[0023] The fully connected layer is used to output the fault type and fault probability.

[0024] Preferably, the fault prediction submodule is further used for:

[0025] Perform hierarchical fault type identification, including system-level fault identification, component-level fault identification, and specific fault mode identification;

[0026] Set up progressive forecast windows, including 1-hour forecast windows, 4-hour forecast windows, and 24-hour forecast windows;

[0027] A multi-level early warning mechanism will be established based on the forecast results, including alert level, warning level, and emergency level.

[0028] Preferably, the input features of the dual-hidden-layer deep neural network model include:

[0029] The four-dimensional core feature vector includes the inlet temperature on the heat source side, the outlet temperature on the heat source side, the inlet temperature on the user side, and the flow rate on the user side.

[0030] Environmental impact factors include ambient temperature, humidity, and water temperature;

[0031] System status characteristics, including compressor load rate, superheat, and subcooling;

[0032] Temporal correlation features include historical data points from the past 10 minutes, 30 minutes, and 1 hour.

[0033] Preferably, the temperature control strategy submodule is also used for:

[0034] The control domain is divided into an emergency start zone, a loading zone, a holding zone, an unloading zone, and an emergency stop zone.

[0035] The control zone is dynamically switched based on the relationship between the current temperature and the set temperature.

[0036] Different control cycles are set for different control zones, with the control cycles for the emergency opening and emergency stopping zones being shorter than those for the loading and unloading zones.

[0037] Preferably, the power gradient calculation unit is specifically used for:

[0038] Collect total system power data at a sampling frequency of 5 seconds per time.

[0039] Calculate the rate of change of power with respect to temperature difference within a 30-second sliding window;

[0040] The search step size is dynamically adjusted based on the gradient magnitude. When the absolute value of the gradient is greater than 0.5 kW / °C, a step size of 1.0°C is used; when the absolute value of the gradient is between 0.1 and 0.5 kW / °C, a step size of 0.5°C is used; and when the absolute value of the gradient is less than 0.1 kW / °C, a step size of 0.2°C is used.

[0041] Preferably, the fuzzy control unit is specifically used for:

[0042] The temperature difference deviation is divided into 7 fuzzy sets;

[0043] The rate of change of deviation is divided into 7 fuzzy sets;

[0044] The estimated system inertia is divided into three fuzzy sets;

[0045] The PID control parameters are adjusted based on the fuzzy rule matrix. When the deviation is large, the proportional coefficient is increased and the integral coefficient is decreased; when the deviation is small, the proportional coefficient is decreased and the integral coefficient is increased. When the system inertia is large, the derivative coefficient is increased; when the system inertia is small, the derivative coefficient is decreased.

[0046] Preferably, the collaboration mechanism between the intelligent analysis module and the decision control module includes:

[0047] The fault prediction results are used as safety constraints for the temperature control system.

[0048] Adjust control strategy parameters based on fault risk prediction results;

[0049] Dynamically adjust load distribution based on component health status;

[0050] When a component is predicted to fail, the control strategy is automatically adjusted to avoid the risk area.

[0051] Preferably, the execution interaction module includes:

[0052] The control unit is used to regulate the frequency converter, electronic expansion valve, and four-way valve;

[0053] The human-computer interaction interface is used to display the system's operating status and receive user commands;

[0054] A remote monitoring system is used to enable remote monitoring, parameter adjustment, and system maintenance.

[0055] The alarm and notification unit is used to send abnormal status and fault warning information.

[0056] Deep learning-based methods for fault prediction and adaptive temperature control in aquaculture heat pumps include:

[0057] Collect temperature, pressure, flow, and power parameters of the heat pump system;

[0058] Based on the parameters, time-domain features, frequency-domain features, time-frequency features, and thermodynamic features are extracted;

[0059] A multi-scale convolutional-long short-term memory hybrid network model is constructed, and the fault types and fault development trends of the heat pump system are predicted based on the features.

[0060] A dual-hidden-layer deep neural network model is constructed, and the energy efficiency index of the heat pump system under different operating conditions is predicted based on the features.

[0061] A dynamic temperature difference setpoint is generated based on the energy efficiency index;

[0062] Calculate the rate of change of the total power of the heat pump system with respect to the temperature difference;

[0063] The optimal temperature difference setpoint is determined based on the power gradient.

[0064] Based on the temperature difference deviation, the rate of change of deviation, and the estimated system inertia, the PID control parameters are dynamically adjusted.

[0065] A control strategy is generated based on the adjusted PID control parameters;

[0066] Control the actuators of the heat pump system to achieve precise temperature control of the aquaculture environment.

[0067] This invention achieves accurate prediction of heat pump system faults through a multi-scale convolutional-long short-term memory hybrid network, evaluates system energy efficiency through a dual-hidden-layer deep neural network, and achieves precise temperature control through power gradient zero-point tracking and three-dimensional fuzzy self-tuning PID control, offering the following advantages:

[0068] 1. Improved temperature control accuracy: The system of this invention improves the water temperature control accuracy to ±0.5℃, which is far superior to the ±2.1℃ of the traditional system, meeting the breeding needs of high-value aquatic products.

[0069] 2. Improve energy efficiency: Through power gradient zero-point tracking technology, the system can automatically find the optimal operating point, improving the energy efficiency ratio by 62% and reducing operating costs by 55% to 70%.

[0070] 3. Fault prediction: The multi-scale convolutional-long short-term memory hybrid network can predict possible faults in the heat pump system 24 to 72 hours in advance, with a fault prediction accuracy of over 85%, providing a reliable basis for preventive maintenance.

[0071] 4. Extend equipment life: Dynamically adjust load distribution based on component health status to avoid potential risk areas and significantly extend equipment life.

[0072] 5. Increase aquaculture output: Through precise temperature control, the growth rate of aquaculture organisms can be increased by 15% to 25%, and year-round aquaculture can be achieved, with a yield increase of 40% to 60%.

[0073] 6. Reduced energy consumption and carbon emissions: The system can save 43,800 kWh of electricity per thousand tons of water per year and reduce CO2 emissions by 35.6 tons per thousand tons of water. Attached Figure Description

[0074] Figure 1 This is a diagram showing the overall architecture of the deep learning-based aquaculture heat pump fault prediction and adaptive temperature control system of the present invention.

[0075] Figure 2This is a structural diagram of the three-dimensional fuzzy self-tuning controller of the present invention;

[0076] Figure 3 This is a flowchart of the hierarchical fault type identification strategy of the present invention;

[0077] Figure 4 This is a schematic diagram of data flow between modules of the system of the present invention;

[0078] Figure 5 This is a flowchart of the method of the present invention. Detailed Implementation

[0079] Please refer to the attached document. Figure 1-5 The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0080] Reference Figure 1 This invention provides a deep learning-based aquatic heat pump fault prediction and adaptive temperature control system. The system includes a data acquisition module 1, a feature extraction module 2, an intelligent analysis module 3, a decision control module 4, and an execution interaction module 5. The modules are connected through a communication bus to form a complete information processing and control link.

[0081] Data acquisition module 1 is used to collect various operating parameters of the heat pump system, providing a data foundation for subsequent analysis and decision-making. Specifically, the parameters collected by this module include temperature, pressure, flow rate, and power.

[0082] In a preferred embodiment of the present invention, the temperature parameters include the inlet temperature of the heat pump system, the outlet temperature, the condenser temperature, the evaporator temperature, and the ambient temperature. Preferably, the temperature parameters are measured using a PT100 temperature sensor with an accuracy of ±0.1℃ and a sampling frequency of 5 seconds / time. This high-precision temperature acquisition is particularly important for the cultivation of high-value aquatic species such as abalone and grouper, as these species are extremely sensitive to changes in water temperature; temperature fluctuations exceeding 1℃ may lead to decreased growth rates or stress responses.

[0083] Pressure parameters include the high-pressure side pressure and low-pressure side pressure of the heat pump system. Preferably, the pressure parameters are measured using a piezoresistive pressure sensor with an accuracy of ±0.5% and a sampling frequency of 10 seconds per measurement. Accurate acquisition of pressure parameters can effectively monitor the refrigerant circulation status and is crucial for diagnosing faults such as compressor performance and system leaks.

[0084] Flow parameters include system water circulation flow rate and refrigerant flow rate. Preferably, flow parameters are measured using an electromagnetic flow meter with an accuracy of ±1% and a sampling frequency of 30 seconds per measurement. In aquaculture environments, stable water circulation is crucial for maintaining uniform water temperature and dissolved oxygen distribution, especially under high-density farming conditions.

[0085] Power parameters include compressor power, water pump power, and total system power. Preferably, power parameters are measured using a three-phase power analyzer with an accuracy of ±0.2% and a sampling frequency of 5 seconds per measurement. Real-time monitoring of power parameters is crucial for evaluating system energy efficiency and detecting abnormal energy consumption, enabling timely detection of problems such as compressor overload.

[0086] In addition, the data acquisition module 1 also includes a data preprocessing unit for filtering, outlier detection, and missing value processing of the acquired raw data. Preferably, a low-pass filter with a cutoff frequency of 0.5Hz is used for filtering to remove high-frequency noise. Outlier detection uses the 3σ criterion, marking data points exceeding the mean ± 3 times the standard deviation as outliers. Missing value processing uses linear interpolation, interpolating data segments with no more than 3 consecutive missing points. Data segments with more than 3 consecutive missing points are marked as invalid data segments and skipped in subsequent analysis.

[0087] The data acquisition module 1 also includes a historical data storage unit for continuously storing 30 days of system operation data, providing historical data support for fault prediction and energy efficiency optimization. Preferably, the historical data is stored in 10-second intervals according to a time series, with each data point containing the measurement values ​​of all sensors at that time and the corresponding timestamp.

[0088] The feature extraction module 2 is communicatively connected to the data acquisition module 1 and is used to receive the parameters acquired by the data acquisition module 1 and extract time-domain features, frequency-domain features, time-frequency features and thermodynamic features.

[0089] Time-domain features include statistical characteristics such as the mean, standard deviation, peak value, peak-to-peak value, skewness, and kurtosis of each parameter. Preferably, a 30-minute sliding window is used for calculating time-domain features, with a window movement step of 5 minutes. Taking the standard deviation calculation as an example, the formula is as follows:

[0090] ,

[0091] in: Standard deviation represents the degree of dispersion of the data. The number of data points in the window is 180 in a 30-minute window with a sampling interval of 10 seconds. For the first The value of a data point, such as the water temperature or pressure value at a certain moment; The average value of the data within the window is calculated using the following formula: In aquaculture heat pump systems, the standard deviation of temperature parameters is an important indicator for evaluating temperature control stability. The smaller the standard deviation, the more stable the temperature control, which is especially important for temperature-sensitive aquatic species such as grouper farming.

[0092] Frequency domain features are obtained by performing a Fast Fourier Transform (FFT) on the time-domain signal, including the amplitude and phase information of the characteristic frequency components. Preferably, frequency domain feature extraction is particularly effective for vibration signals, capable of identifying the characteristic frequencies of rotating components such as compressors and water pumps, providing important information for fault diagnosis. The core formula for frequency domain feature extraction is:

[0093] ,

[0094] in: For frequency domain features, representing the first Complex values ​​of each frequency component; This is a time-domain signal, such as raw vibration data collected by a compressor vibration sensor; The signal length determines the accuracy of the spectrum analysis; The frequency exponent ranges from 0 to N-1; The imaginary unit, ; The rotation factor is used to convert the time-domain signal to the frequency domain. For compressor vibration signals, the preferred acquisition frequency is 100Hz, and the analysis window length is 10 seconds, i.e., N=1000. In practical applications, compressors typically exhibit characteristic peaks at specific frequencies (such as rotor frequency) during normal operation, while abnormal conditions such as bearing failure and blade damage can generate abnormal peaks or sidebands in specific frequency ranges. These abnormal patterns can be accurately identified through spectrum analysis.

[0095] Time-frequency features are obtained through wavelet transform, which can simultaneously provide information in both time and frequency dimensions, making it particularly suitable for analyzing non-stationary signals. Preferably, wavelet decomposition is used to analyze the pressure pulsation signal, extracting the energy distribution characteristics of different frequency bands. The mathematical expression for wavelet decomposition is:

[0096] ,

[0097] in: These are wavelet transform coefficients, representing the signal energy distribution on the time-scale plane; This is a scale parameter, inversely proportional to frequency; large scales correspond to low frequencies, and small scales correspond to high frequencies. The translation parameter represents the time position; This refers to raw signals, such as pressure pulsation signals in a heat pump system; As the conjugate of wavelet basis functions, the present invention preferably uses the Daubechies wavelet (db4) as the basis function; The energy normalization factor is used; the integral represents the summation of the product of the signal and the translated and scaled wavelet basis functions over all time ranges. This invention employs 5-level wavelet decomposition, which can decompose the original signal into sub-signals of different frequency bands. Especially for pressure pulsation signals in heat pump systems, low-frequency components usually reflect macroscopic changes in the system, while high-frequency components reflect local transient characteristics, such as valve vibration and fluid pulsation. Combining the energy characteristics of different frequency bands allows for a more comprehensive assessment of the system state.

[0098] Thermodynamic characteristics include thermodynamic parameters such as superheat, subcooling, compression ratio, and cycle efficiency. Preferably, superheat is defined as the difference between the refrigerant temperature at the evaporator outlet and the evaporation temperature, and subcooling is defined as the difference between the condensation temperature and the refrigerant temperature at the condenser outlet. Compression ratio is defined as the ratio of high-pressure side pressure to low-pressure side pressure. Cycle efficiency is calculated using the enthalpy difference method, i.e., the ratio of cooling capacity to compressor power. These thermodynamic characteristics are crucial for evaluating the operating status and energy efficiency of a heat pump system. For example, excessive superheat may lead to compressor overheating, while excessively low superheat may lead to liquid slugging; changes in subcooling may reflect condenser scaling or abnormal refrigerant charge.

[0099] The intelligent analysis module 3 is communicatively connected to the feature extraction module 2, and is used to receive the features extracted by the feature extraction module 2 and perform in-depth analysis. This module includes a fault prediction submodule, an energy efficiency optimization submodule, and a temperature control strategy submodule.

[0100] The fault prediction submodule is used to build a multi-scale convolutional-long short-term memory hybrid network model based on the features extracted by the feature extraction module 2, and to predict the fault types and fault development trends of the heat pump system.

[0101] The multi-scale convolutional-long short-term memory hybrid network model includes multi-scale convolutional layers, bidirectional long short-term memory layers, attention layers, and fully connected layers.

[0102] The multi-scale convolutional layer consists of three parallel convolutional paths, each using different kernel sizes for feature extraction: 3×1, 5×1, and 7×1 kernels. Each path contains two convolutional operations: the first outputs 64 feature maps, and the second outputs 32 feature maps. The mathematical expression for the convolutional operation is:

[0103] ,

[0104] in: For the convolution result at position The value; The input feature matrix, in the aquaculture heat pump system, represents a feature matrix composed of time-domain, frequency-domain, time-frequency, and thermodynamic features; The convolution kernel represents the feature extractor; and N represents the height and width of the convolution kernel, respectively. For time-series data, convolution is performed along the time dimension, so N=1, and M are 3, 5, and 7, corresponding to short-term, medium-term, and long-term time dependencies, respectively. This multi-scale convolution design can simultaneously capture feature patterns at different time scales. For example, a 3×1 convolution kernel can capture short-term change patterns within 30 seconds, suitable for identifying rapid faults such as valve vibration; a 5×1 convolution can capture medium-term change patterns within 50 seconds, suitable for identifying medium-speed faults such as refrigerant flow fluctuations; and a 7×1 convolution kernel can capture long-term change patterns within 70 seconds, suitable for identifying slow faults such as slow temperature drift.

[0105] A bidirectional Long Short-Term Memory (LSTM) layer is used to capture the forward and backward temporal dependencies of time-series data. This layer contains 128 hidden units, and the output concatenates the forward and backward states. The core computations of the LSM unit include the input gate, forget gate, output gate, and cell state updates.

[0106] ,

[0107] ,

[0108] ,

[0109] ,

[0110] ,

[0111] in: The activation value of the input gate controls the degree to which the current input information flows into the unit state; The activation value of the forget gate controls the degree to which the cell state from the previous time step is retained; The activation value of the output gate represents the degree to which the control unit state affects the current output. This is a unit state, storing long-term memory; This is a hidden state, used as the output at the current moment; The input vector represents the multidimensional feature vector at the current moment in the aquaculture heat pump system. and These are the weights and bias parameters, obtained through training. The sigmoid activation function has an output range of (0,1). is the hyperbolic tangent activation function, with an output range of (-1, 1). This is element-wise multiplication. In a bidirectional LSTM, the above calculations are performed independently in both directions, and the final output is the concatenation of the hidden states in both directions: This two-way structure can take into account information from the past and the future simultaneously, which is particularly important for fault prediction in aquaculture heat pump systems. This is because some faults may exhibit certain characteristics some time before they occur. For example, compressor bearing failures often go through a stage of slight abnormal noise and increased vibration before they are completely destroyed.

[0112] The attention layer is used to calculate weights for different time steps and feature channels. This invention employs a dual attention mechanism, including temporal attention and channel attention. The formula for calculating temporal attention is:

[0113] ,

[0114] ,

[0115] ,

[0116] in: For time steps Attention score, indicating the importance of that time point; The normalized attention weights satisfy... ; This is a context vector that integrates information from all time steps; for The hidden state at any given moment; and These are learnable parameters; The total sequence length is denoted as . Channel attention generates channel descriptors through global average pooling and global max pooling, and then generates channel weights through a shared multilayer perceptron. In aquaculture heat pump systems, different time points and different sensor channels contribute differently to fault prediction. For example, in refrigerant leakage faults, changes in low-pressure side pressure and superheat are usually the earliest indicators, while in compressor mechanical faults, changes in vibration signals are more critical. The attention mechanism can automatically learn these differences in importance, improving prediction accuracy.

[0117] The fully connected layer is used to output the fault type and fault probability. The structure of the fully connected layer is a stepwise dimensionality reduction: 256 → 128 → 64 → 32 → number of fault types. The last layer uses the softmax activation function to output the probability distribution of each fault type.

[0118] ,

[0119] in: For the first The probability of a type of failure; This is the original output score for this class; The total number of fault types is 25 in this invention, including typical fault modes such as compressor suction valve failure, condenser scaling, evaporator frosting, and expansion valve blockage. In aquaculture applications, accurately predicting these faults is crucial for avoiding downtime and maintaining stable water temperatures, especially in the cultivation of temperature-sensitive high-value aquatic products, such as abalone farming, where water temperature fluctuations exceeding 1.5°C can lead to stress reactions or even death.

[0120] In a preferred embodiment of the present invention, the fault prediction submodule further performs hierarchical fault type identification, including system-level fault identification, component-level fault identification, and specific fault mode identification. System-level fault identification categorizes system states into three types: normal operation, minor anomalies, and severe faults. Component-level fault identification locates faulty components, including compressors, condensers, evaporators, expansion valves, and control systems. Specific fault mode identification identifies the specific fault type for each component. This hierarchical identification strategy meets the actual needs of heat pump system fault diagnosis, allowing the system to adopt different response strategies based on the severity of the fault.

[0121] In addition, the fault prediction submodule features progressive prediction windows, including 1-hour, 4-hour, and 24-hour prediction windows. Short-term prediction (1 hour) uses a fine-grained time-series model for early warning of urgent faults; medium-term prediction (4 hours) combines system dynamics for comprehensive prediction, used for maintenance planning within work shifts; and long-term prediction (24 hours) uses historical pattern matching for trend prediction, used for next day's maintenance planning. This multi-timescale prediction is particularly important for aquaculture farm management because it allows managers to rationally allocate manpower and spare parts based on prediction results, minimizing the impact of downtime on the aquaculture environment.

[0122] Based on the prediction results, the fault prediction submodule establishes a multi-level early warning mechanism, including alert-level, warning-level, and emergency-level early warnings. Alert-level early warnings indicate faults that may occur within 72 hours, with a prediction confidence level between 60% and 75%; warning-level early warnings indicate faults that may occur within 24 hours, with a prediction confidence level between 75% and 90%; and emergency-level early warnings indicate faults that may occur within 6 hours, with a prediction confidence level above 90%. Different levels of early warning correspond to different response strategies, from scheduling routine checks to preparing emergency spare parts and activating emergency plans, forming a complete preventative maintenance system.

[0123] The energy efficiency optimization submodule is used to construct a dual-hidden-layer deep neural network model based on the features extracted by the feature extraction module 2, and to predict the energy efficiency index of the heat pump system under different operating conditions.

[0124] The network topology of the dual-hidden-layer deep neural network model is as follows: Input layer (multi-dimensional features) → Hidden layer 1 (8 nodes) → Hidden layer 2 (4 nodes) → Output layer (energy efficiency index). Hidden layer 1 uses the ReLU activation function, hidden layer 2 uses the Tanh activation function, and the output layer uses the linear activation function.

[0125] The forward propagation calculation for this network is as follows:

[0126] ,

[0127] ,

[0128] ,

[0129] ,

[0130] ,

[0131] ,

[0132] in: The input feature vector contains multi-dimensional operating parameters of the aquaculture heat pump system; and The first The layer's weight matrix and bias vector are obtained through training; For the first The linear output of the layer, i.e., weighted and biased; For the first The layer's activation output is the result after processing by the activation function; ReLU is the modified linear unit activation function, which truncates negative values ​​to 0; tanh is the hyperbolic tangent activation function, which maps the input to the interval [-1, 1]. Final output This indicates the predicted energy efficiency indicators, including the coefficient of performance (COP) and total system power consumption. In aquaculture applications, the COP is a key indicator for measuring the efficiency of heat pump systems and directly affects operating costs. In large-scale applications such as aquaculture farms, a 1% improvement in energy efficiency can bring significant economic benefits.

[0133] In a preferred embodiment of the present invention, the input features of the dual-hidden-layer deep neural network model include a four-dimensional core feature vector, environmental influence factors, system state features, and temporal correlation features.

[0134] The four-dimensional core feature vector includes the inlet temperature on the heat source side, the outlet temperature on the heat source side, the inlet temperature on the user side, and the flow rate on the user side. These four parameters are the core parameters that determine the operating status of the heat pump system and directly affect the system's heat exchange efficiency and energy consumption. In aquaculture applications, the heat source side is usually connected to the aquaculture water body, while the user side is connected to the heat exchange equipment. The temperature difference and flow rate between the two determine the system's load and efficiency.

[0135] Environmental factors include ambient temperature, humidity, and water temperature. Ambient temperature affects condenser heat dissipation efficiency, humidity affects evaporator frosting, and water temperature is a direct indicator of system load. In open-air aquaculture ponds, these environmental factors change significantly and have a substantial impact on system performance. For example, in high-temperature and high-humidity environments, condenser heat dissipation efficiency decreases, and the system COP may drop by more than 20%.

[0136] System status characteristics include compressor load rate, superheat, and subcooling. Compressor load rate reflects the system load, while superheat and subcooling reflect whether the system is operating at its optimal state. In practical applications, superheat is typically maintained within the range of 5-8°C, and subcooling within the range of 3-5°C to achieve optimal energy efficiency.

[0137] Temporal correlation features include historical data points from the past 10 minutes, 30 minutes, and 1 hour. This historical data reflects the dynamic characteristics and trends of the system, helping to improve prediction accuracy. Especially in aquaculture environments where water temperature changes slowly, systems often exhibit significant temporal correlations, and using this historical data can more accurately predict future conditions.

[0138] When training a two-hidden-layer deep neural network model, the mean squared error (MSE) is used as the loss function:

[0139] ,

[0140] in: This refers to actual energy efficiency indicators, such as the measured COP value; These are the model's predicted values; The sample size is specified. Preferably, the Adam optimizer is used for model training, with an initial learning rate of 0.001 and a learning rate decay strategy, reducing the learning rate to 0.9 times its original value every 50 epochs. This training strategy improves the model's generalization ability while ensuring convergence speed, enabling it to adapt to the energy efficiency prediction needs of heat pump systems in different aquaculture environments.

[0141] The temperature control strategy submodule is used to generate dynamic temperature difference setpoints based on the energy efficiency indicators predicted by the energy efficiency optimization submodule.

[0142] The temperature control strategy submodule divides the control domain into an emergency opening zone, a loading zone, a holding zone, an unloading zone, and an emergency stop zone. The system dynamically switches between these control zones based on the relationship between the current temperature and the set temperature.

[0143] In a preferred embodiment of the present invention, the boundary values ​​of the control region are set as follows:

[0144] Emergency opening zone: Set temperature ±P03, where P03 is the emergency opening hysteresis temperature, with a default value of 2.0℃.

[0145] Loading zone: Set temperature ±P02 (cooling mode) or ±P01 (heating mode), where P02 is the cooling hysteresis temperature, with a default value of 1.0℃, and P01 is the heating hysteresis temperature, with a default value of 1.5℃.

[0146] Holding range: set temperature ±0.2℃

[0147] Unloading area: Set temperature ±P02 or ±P01

[0148] Emergency Stop Zone: Set temperature ±P04, where P04 is the emergency stop hysteresis temperature, with a default value of 3.0℃.

[0149] These parameter values ​​are set based on the actual needs of aquaculture, especially taking into account the temperature sensitivity of different farmed species. For example, for temperature-sensitive species such as abalone, PO2 can be set to 0.8℃, and the holding zone can be set to ±0.1℃ to provide more precise temperature control; while for species with strong temperature adaptability such as grass carp, PO2 can be set to 1.5℃ to reduce the frequency of equipment start-up and shutdown and extend its service life.

[0150] The temperature control strategy submodule sets different control cycles for different control zones. The control cycle for the emergency start and emergency stop zones is P05 seconds (default 5 seconds), while the control cycle for the loading and unloading zones is P06 minutes (default 3 minutes). This differentiated control cycle design ensures system response speed while avoiding damage to the equipment from frequent start-stop cycles. For example, a rapid response is needed in the emergency start zone to restore water temperature, while a longer control cycle can be used in the loading zone to gradually adjust the system output and reduce energy consumption fluctuations.

[0151] The temperature difference setpoint is calculated every 30 seconds. Within each calculation cycle, the optimal supply / return water temperature difference setpoint is determined based on the output of the energy efficiency prediction model. The determination of the temperature difference setpoint considers both energy efficiency maximization and control stability. A quadratic curve is used to fit the relationship between energy efficiency and temperature difference, identifying the temperature difference value corresponding to the highest point of the energy efficiency curve. In practical applications, the optimal temperature difference may vary significantly under different aquaculture environments. For example, in high-density fish farming, due to greater biological heat production, the optimal temperature difference is typically larger (approximately 5-6℃), while in low-density shellfish farming, the optimal temperature difference is smaller (approximately 3-4℃).

[0152] The decision control module 4 is communicatively connected to the intelligent analysis module 3. It receives fault prediction results from the fault prediction submodule and dynamic temperature difference setpoints from the temperature control strategy submodule, and generates control strategies. This module includes a power gradient calculation unit, a temperature difference setting unit, and a fuzzy control unit.

[0153] The power gradient calculation unit is used to calculate the rate of change of the total power of the heat pump system with respect to the temperature difference.

[0154] The power gradient calculation unit collects total system power data at a sampling frequency of 5 seconds per sampling, and calculates the rate of change of power with respect to temperature difference within a 30-second sliding window. The formula for calculating the power gradient is:

[0155] ,

[0156] in: The power gradient, in units of kW / °C, represents the change in system power when the temperature difference changes by 1°C. and The total system power at two time points is expressed in kW. and The supply / return water temperature difference at the corresponding moment is expressed in °C. The power gradient reflects the degree of influence of temperature difference changes on system energy consumption and is a key indicator for finding the optimal operating point. In aquaculture heat pump systems, due to the slow change in water temperature and relatively stable system load, the power gradient typically exhibits good stability, which provides the possibility for zero-point tracking control.

[0157] In a preferred embodiment of the present invention, the power gradient calculation unit dynamically adjusts the search step size according to the gradient magnitude. When the absolute value of the gradient is greater than 0.5 kW / ℃, a step size of 1.0℃ is used for a large-range search; when the absolute value of the gradient is between 0.1 and 0.5 kW / ℃, a step size of 0.5℃ is used for a medium-range search; and when the absolute value of the gradient is less than 0.1 kW / ℃, a step size of 0.2℃ is used for a fine search. This dynamic step size strategy can improve the accuracy of optimal point location while ensuring search speed. For example, in the early stage of system startup, since the operating point is far from the optimal state, the power gradient is usually large. At this time, using a large step size can quickly approach the optimal region; while when the system approaches the optimal operating point, the power gradient becomes smaller. At this time, using a small step size can accurately locate the optimal point and avoid oscillations caused by excessively large step sizes.

[0158] The temperature difference setting unit is used to determine the optimal temperature difference setting value based on the power gradient calculated by the power gradient calculation unit.

[0159] The temperature difference setting unit employs an iterative search method to find an operating point where the power gradient is close to zero. When the absolute value of the gradient is less than a preset threshold (default 0.05kW / ℃), the current operating point is considered close to the optimum, and the current temperature difference is output as the setpoint. This threshold is set based on actual system testing; 0.05kW / ℃ means that a 1℃ change in temperature results in a power change of no more than 0.05kW. For a typical 10kW aquaculture heat pump system, this equates to a power fluctuation of 0.5%, which is sufficiently accurate.

[0160] During the search process, the temperature difference setting unit follows these principles:

[0161] 1. If the current gradient is positive (power increases with temperature difference), then decrease the temperature difference setpoint.

[0162] 2. If the current gradient is negative (power decreases as temperature difference increases), then increase the temperature difference setpoint.

[0163] 3. If the current gradient is close to zero, maintain the current temperature difference setpoint.

[0164] In addition, the temperature difference setting unit also sets upper and lower limits for the temperature difference to ensure safe system operation. Generally, the lower limit is set at 3°C ​​to ensure sufficient heat exchange efficiency; the upper limit is set at 8°C to avoid excessive temperature differences that could lead to decreased system efficiency or instability. These limits are set considering the actual operating characteristics of aquaculture heat pump systems. For example, a small temperature difference can increase pump energy consumption, while a large temperature difference may cause evaporator frosting or insufficient heat exchange.

[0165] The fuzzy control unit is used to dynamically adjust the PID control parameters based on the temperature difference deviation, the rate of change of the deviation, and the estimated value of the system inertia.

[0166] Reference Figure 2The fuzzy control unit divides the temperature difference deviation (e) into seven fuzzy sets {NB, NM, NS, ZO, PS, PM, PB}, representing negative large, negative medium, negative small, zero, positive small, positive medium, and positive large, respectively. In the specific implementation, the membership functions of these fuzzy sets are defined as follows: NB corresponds to the range of [-3, -2]℃, NM corresponds to the range of [-2, -1]℃, NS corresponds to the range of [-1, -0.2]℃, ZO corresponds to the range of [-0.2, 0.2]℃, PS corresponds to the range of [0.2, 1]℃, PM corresponds to the range of [1, 2]℃, and PB corresponds to the range of [2, 3]℃. The rate of change of deviation (ec) is also divided into 7 fuzzy sets, whose physical meaning is the rate of change of temperature difference deviation, in °C / min. Specifically, NB corresponds to [-0.6, -0.4] °C / min, NM corresponds to [-0.4, -0.2] °C / min, NS corresponds to [-0.2, -0.05] °C / min, ZO corresponds to [-0.05, 0.05] °C / min, PS corresponds to [0.05, 0.2] °C / min, PM corresponds to [0.2, 0.4] °C / min, and PB corresponds to [0.4, 0.6] °C / min. The estimated value of system inertia (J) is divided into 3 fuzzy sets {S, M, L}, representing small, medium, and large, respectively. The corresponding physical meaning is the system's response speed to the control signal. Specifically, S corresponds to [0, 5] minute response time, M corresponds to [5, 15] minute response time, and L corresponds to [15, 30] minute response time.

[0167] The fuzzy control unit adjusts the PID control parameters based on a fuzzy rule matrix. Preferably, the PID parameter adjustment follows these rules:

[0168] The proportional gain Kp increases when the deviation is large and decreases when the deviation is small. This is because under large deviation conditions, strong correction is required to quickly approach the target value; while under small deviation conditions, reducing the proportional gain can avoid overshoot.

[0169] Integral coefficient Ki: Decrease when the deviation is large, increase when the deviation is small. This is because, under large deviation conditions, reducing the integral action can avoid integral saturation; while under small deviation conditions, increasing the integral coefficient can eliminate static error.

[0170] The differential coefficient Kd increases when the system inertia is large and decreases when the system inertia is small. This is because systems with large inertia have a delayed response and require stronger differential action for predictive control; while systems with small inertia respond quickly, and excessively strong differential action may cause system oscillations.

[0171] Fuzzy inference employs the Mamdani method, fuzzification uses triangular membership functions, and defuzzification uses the central averaging method. Taking the adjustment of the scaling factor Kp as an example, its fuzzy rule can be expressed as:

[0172] IFeisNBANDecisNBTHENΔKpisPB

[0173] IFeisNBANDecisNMTHENΔKpisPB ...

[0175] IFeisPBANDecisPBTHENΔKpisPB

[0176] By using similar rules, the adjustment amounts of the integral coefficient Ki and the derivative coefficient Kd under different conditions can be obtained. In practical applications, this fuzzy logic-based PID self-adjustment strategy can effectively cope with various disturbances in the aquaculture environment, such as diurnal temperature variations and sudden load changes caused by rainfall, maintaining the stable control performance of the system.

[0177] The fuzzy control unit also employs a first-order filter to smooth the control output, preventing system instability caused by sudden parameter changes. The filter formula is:

[0178] ,

[0179] in: This is the filtered output, which is the final control parameter applied to the system; This is the output from the previous time step; This is the current raw output, which is the direct result of fuzzy inference; This is the filter coefficient, with a value range of [0,1] and a default value of 0.8. The selection of the filter coefficient is based on the dynamic characteristics of the system. For aquaculture water bodies with high thermal inertia, a larger filter coefficient can prevent the control parameters from changing too quickly and improve system stability.

[0180] The execution interaction module 5 is communicatively connected to the decision control module 4, and is used to receive the control strategy generated by the decision control module 4 and control the actuators of the heat pump system.

[0181] In a preferred embodiment of the present invention, the execution interaction module 5 includes an execution control unit, a human-machine interface, a remote monitoring system, and an alarm and notification unit.

[0182] The control unit regulates the frequency converter, electronic expansion valve, and four-way valve. The frequency converter controls the compressor speed, achieving stepless speed regulation; the electronic expansion valve controls the refrigerant flow, adjusting the system superheat; and the four-way valve controls the refrigerant flow direction, enabling switching between cooling and heating modes. Preferably, the frequency converter control uses a PWM (Pulse Width Modulation) signal with a frequency range of 20–50Hz, corresponding to a compressor load rate of 30%–100%; the electronic expansion valve control uses a stepper motor drive with an opening range of 0–480 steps, corresponding to an opening degree of 0%–100%; and the four-way valve control uses a 24V DC signal. This precise control enables continuous adjustment of the heat pump system, meeting the requirements for precise temperature control in aquaculture environments.

[0183] The human-machine interface (HMI) is used to display the system's operating status and receive user commands. The HMI uses a 10.1-inch color touchscreen with a resolution of 1280×800, displaying real-time temperature curves, system operating status, energy efficiency indicators, and fault warning information. Users can set parameters such as target temperature, operating mode, and timed start / stop through the interface. The interface design conforms to the usage habits of aquaculture personnel, is simple and intuitive, and key information is clearly displayed, facilitating daily monitoring and operation.

[0184] The remote monitoring system is used for remote monitoring, parameter adjustment, and system maintenance. It is available in both mobile app and web application formats, allowing users to view system status, adjust control parameters, and receive alarm information anytime, anywhere. Preferably, the remote monitoring system uses the MQTT protocol for data transmission to ensure real-time performance and reliability. This functionality is particularly important for large-scale aquaculture bases, as these bases are typically geographically dispersed; remote monitoring reduces inspection costs and improves management efficiency.

[0185] The alarm and notification unit is used to send abnormal status and fault warning information. Alarm information is divided into three levels: prompt information (blue), warning information (yellow), and emergency alarm (red), corresponding to different levels of urgency. Alarm methods include interface display, audible and visual alarms, SMS notifications, and APP push notifications. Preferably, the alarm threshold can be customized by the user, while the system default threshold is determined based on historical data statistical analysis. In aquaculture applications, a timely alarm mechanism is crucial, especially at night or during unattended periods, effectively preventing temperature control failures and losses of farmed organisms due to equipment malfunctions.

[0186] The present invention establishes a collaborative mechanism between the intelligent analysis module 3 and the decision control module 4, enabling the fault prediction and temperature control system to work together and improve the overall system performance.

[0187] Reference Figure 4The fault prediction results serve as safety constraints for the temperature control system, influencing the generation of the control strategy. Preferably, the safety constraint evaluation cycle is 5 minutes, and the system adopts different control strategy adjustment measures based on the risk level of the fault prediction.

[0188] Low risk (prediction confidence level <60%): Maintain the original control strategy.

[0189] Medium risk (prediction confidence level 60%–85%): Adjust control parameters to avoid the risk zone.

[0190] High risk (prediction confidence > 85%): Activate backup control strategies to reduce the load on affected components.

[0191] These thresholds are set based on real-world application testing, taking into account both the timeliness of warnings and the false alarm rate. For example, in a practical application at an aquaculture base, when the medium-risk threshold is set to 75%, the system can issue warnings an average of 36 hours before a fault occurs, while keeping the false alarm rate below 10%, achieving a good balance.

[0192] The system also dynamically adjusts load distribution based on component health status, extending the lifespan of weaker components. For example, when abnormal vibration of the compressor bearing is detected, the system appropriately reduces the compressor load and increases its operating time to balance the overall cooling capacity while reducing the burden on components. This intelligent load management is particularly important in aquaculture applications, which are often located in remote areas with inconvenient maintenance; extending equipment lifespan and reducing the need for emergency repairs has significant economic value.

[0193] Furthermore, when a component failure is predicted, the system automatically adjusts its control strategy to avoid the risk area. For example, if the expansion valve is predicted to become clogged, the system will increase the minimum opening setting to reduce the risk of blockage; if scaling is predicted in the condenser, the system will decrease the condensing pressure setting to reduce the load on the condenser. This preventative control strategy maximizes equipment uptime while ensuring temperature control performance, providing maintenance personnel with sufficient response time and avoiding adverse effects on aquaculture organisms caused by sudden temperature fluctuations due to unforeseen failures.

[0194] Reference Figure 5 The present invention also provides a deep learning-based method for predicting and adaptively controlling the faults of aquatic heat pumps, comprising the following steps:

[0195] Step 1: Collect the temperature, pressure, flow rate, and power parameters of the heat pump system.

[0196] Step 2: Extract time-domain features, frequency-domain features, time-frequency features, and thermodynamic features based on the parameters.

[0197] Step 3: Construct a multi-scale convolutional-long short-term memory hybrid network model, and predict the fault type and fault development trend of the heat pump system based on the features.

[0198] Step 4: Construct a dual-hidden-layer deep neural network model and predict the energy efficiency index of the heat pump system under different operating conditions based on the features.

[0199] Step 5: Generate a dynamic temperature difference setpoint based on the energy efficiency index.

[0200] Step 6: Calculate the rate of change of the total power of the heat pump system with respect to the temperature difference.

[0201] Step 7: Determine the optimal temperature difference setting value based on the power gradient.

[0202] Step 8: Dynamically adjust the PID control parameters based on the temperature difference deviation, the rate of change of deviation, and the estimated system inertia.

[0203] Step 9: Generate a control strategy based on the adjusted PID control parameters.

[0204] Step 10: Control the actuator of the heat pump system to achieve precise temperature control of the aquaculture environment.

[0205] The system of this invention demonstrates superior performance in practical applications. Its temperature control effect is significantly better than traditional control systems for various aquaculture species. For high-value species sensitive to temperature, such as abalone, the system improves temperature control accuracy to ±0.5℃, shortening the aquaculture cycle by 25%, increasing survival rate by 15%, and improving economic benefits by over 40%.

[0206] In terms of energy efficiency, the system of this invention uses power gradient zero-point tracking technology to ensure that the system always operates at the optimal energy efficiency point, improving the energy efficiency ratio by 62% and reducing operating costs by 55% to 70%. Taking a 1,000-ton aquaculture farm as an example, the annual electricity savings reach 43,800 kWh, which translates to an annual electricity cost saving of 35,040 yuan at a price of 0.8 yuan / kWh.

[0207] Regarding fault prediction, the system of this invention can predict potential faults in the heat pump system 24 to 72 hours in advance, with a fault prediction accuracy rate of over 85%. In one real-world case, the system successfully predicted early signs of fault in the compressor's intake valve, issuing a warning 48 hours in advance, thus preventing downtime due to a sudden fault and saving approximately 200,000 yuan in economic losses.

[0208] In addition, the remote monitoring function of the system of the present invention greatly reduces the need for manual inspection, reduces labor costs by about 40%, and improves response speed, shortening the average fault response time from 4 hours to less than 30 minutes.

[0209] The system of this invention has been successfully applied in multiple aquaculture bases, with significant comprehensive benefits, and has gained high recognition from users.

[0210] This invention provides a deep learning-based aquaculture heat pump fault prediction and adaptive temperature control system and method. It achieves heat pump system fault prediction through a multi-scale convolutional-long short-term memory hybrid network, and achieves precise temperature control and energy efficiency optimization through a dual-hidden-layer deep neural network and power gradient zero-point tracking technology, significantly improving the stability and economic benefits of aquaculture. This invention solves key problems in existing aquaculture heat pump systems, such as low temperature control accuracy, low energy efficiency, and lack of fault prediction capabilities, providing technical support for the intelligent and green development of aquaculture.

[0211] The above description is merely a preferred embodiment of the present invention and does not limit the scope of patent protection of the present invention. Any equivalent structural transformations made based on the inventive concept of the present invention and the contents of the specification and drawings of the present invention, or direct / indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A deep learning-based aquaculture heat pump failure prediction and adaptive temperature control system, characterized in that, The method comprises the steps of: a data acquisition module for acquiring temperature parameters, pressure parameters, flow parameters and power parameters of the heat pump system; a feature extraction module in communication with the data acquisition module, configured to receive the parameters collected by the data acquisition module and extract time domain features, frequency domain features, time-frequency features and thermodynamic features; an intelligent analysis module in communication with the feature extraction module, configured to receive the features extracted by the feature extraction module, the intelligent analysis module comprising: a fault prediction submodule for establishing a multi-scale convolution-long short-term memory hybrid network model based on the features to predict the fault type and fault development trend of the heat pump system; an energy efficiency optimization submodule for constructing a double-hidden-layer deep neural network model based on the features to predict the energy efficiency indicators of the heat pump system under different operating conditions; a temperature control strategy submodule for generating a temperature difference dynamic set value based on the energy efficiency indicators; a decision control module in communication with the intelligent analysis module, configured to receive the fault prediction results of the fault prediction submodule and the temperature difference dynamic set value of the temperature control strategy submodule, and generate a control strategy, the decision control module comprising: a power gradient calculation unit for calculating the rate of change of the total power of the heat pump system with respect to the temperature difference; a temperature difference setting unit for determining an optimal temperature difference set value based on the power gradient; a fuzzy control unit for dynamically adjusting the PID control parameters based on the temperature difference deviation, the deviation rate of change and the system inertia estimate value; an execution interaction module in communication with the decision control module, configured to receive the control strategy generated by the decision control module and control the actuators of the heat pump system; The input features of the double-hidden-layer deep neural network model include: a four-dimensional core feature vector including the heat source side inlet temperature, the heat source side outlet temperature, the user side inlet temperature and the user side flow rate; environmental factors including ambient temperature, humidity and water temperature; system state features including compressor load rate, superheat and subcooling; time series correlation features including historical data points for the past 10 minutes, 30 minutes and 1 hour; The power gradient calculation unit is specifically configured to: acquire system total power data with a sampling frequency of 5 seconds per sample; calculate the rate of change of power with respect to temperature difference within a 30-second sliding window; dynamically adjust the search step size according to the gradient size, using a 1.0°C step size when the gradient absolute value is greater than 0.5kW / °C, a 0.5°C step size when the gradient absolute value is between 0.1-0.5kW / °C, and a 0.2°C step size when the gradient absolute value is less than 0.1kW / °C.

2. The system of claim 1, wherein, The multi-scale convolution-long short-term memory hybrid network model comprises: a multi-scale convolution layer including three parallel convolution paths, each path using a different convolution kernel size for feature extraction; a bidirectional long short-term memory layer for capturing forward and backward time dependence of time series data; an attention layer for calculating weights for different time steps and feature channels; a fully connected layer for outputting fault types and fault probabilities.

3. The system of claim 1, wherein, The fault prediction submodule is further configured to: perform hierarchical fault type identification, including system-level fault identification, component-level fault identification and specific fault mode identification; A progressive prediction window is set, including a 1-hour prediction window, a 4-hour prediction window, and a 24-hour prediction window; A multi-level early warning mechanism is established based on the prediction results, including a prompt level warning, a warning level warning, and an emergency level warning.

4. The system of claim 1, wherein, The temperature control strategy sub-module is further configured to: divide the control domain into an emergency start zone, a loading zone, a holding zone, an unloading zone, and an emergency stop zone; dynamically switch the control region according to the relationship between the current temperature and the set temperature; set different control periods for different control regions, wherein the control periods of the emergency start zone and the emergency stop zone are shorter than those of the loading zone and the unloading zone.

5. The system of claim 1, wherein, The fuzzy control unit is specifically configured to: divide the temperature difference deviation into 7 fuzzy sets; divide the deviation change rate into 7 fuzzy sets; divide the system inertia estimate value into 3 fuzzy sets; adjust the PID control parameters based on the fuzzy rule matrix, wherein the proportional coefficient is increased and the integral coefficient is decreased when the deviation is large, the proportional coefficient is decreased and the integral coefficient is increased when the deviation is small, the differential coefficient is increased when the system inertia is large, and the differential coefficient is decreased when the system inertia is small.

6. The system of claim 1, wherein, The coordination mechanism between the intelligent analysis module and the decision control module includes: using the fault prediction results as safety constraints for the temperature control system; adjusting the control strategy parameters based on the fault risk prediction results; dynamically adjusting the load distribution according to the component health status; automatically adjusting the control strategy to avoid the risk area when predicting that the component is about to fail.

7. The system of claim 1, wherein, The execution interaction module includes: an execution control unit for adjusting the frequency converter, electronic expansion valve, and four-way valve; a human-computer interaction interface for displaying system operation status and receiving user instructions; a remote monitoring system for realizing remote monitoring, parameter adjustment, and system maintenance; an alarm and notification unit for sending abnormal state and fault warning information.

8. A deep learning method for fault prediction and adaptive temperature control of aquaculture heat pumps, using the system of any one of claims 1-7, characterized in that, It includes: collecting temperature parameters, pressure parameters, flow parameters, and power parameters of the heat pump system; extracting time domain features, frequency domain features, time-frequency features, and thermodynamic features based on the parameters; constructing a multi-scale convolution-long short-term memory hybrid network model and predicting the fault type and fault development trend of the heat pump system based on the features; constructing a double-hidden-layer deep neural network model and predicting the energy efficiency indicators of the heat pump system under different working conditions based on the features; generating a temperature difference dynamic set value based on the energy efficiency indicators; calculating the change rate of the total power of the heat pump system to the temperature difference; determining the optimal temperature difference set value based on the power gradient; dynamically adjusting the PID control parameters based on the temperature difference deviation, the deviation change rate, and the system inertia estimate value; generating a control strategy according to the adjusted PID control parameters; controlling the actuators of the heat pump system to realize accurate temperature control of the aquaculture environment.

Images