Fish shoal feeding behavior grading identification method and system based on vibration signals

Abstract

The invention relates to a vibration signal-based fish school ingestion behavior grading identification method, which comprises the steps of controlling a plurality of vibration sensors to synchronously collect vibration signals caused by fish school ingestion, carrying out synthesis preprocessing on the vibration signals to obtain synthesis signals, sequentially carrying out filtering, framing and wavelet transform processing on the synthesis signals to carry out feature extraction, and obtaining the fish school ingestion behavior grading identification result. The method comprises the following steps: obtaining a first feature matrix, carrying out dimension reduction processing on the first feature matrix to obtain a second feature matrix, inputting the second feature matrix into a pre-trained deep learning model to obtain a fish shoal feeding behavior vibration signal grading identification signal, realizing accurate feeding with low cost, and being capable of being widely popularized. The breeding efficiency is improved.

Description

Technical Field

[0001] The present application relates to the technical field of aquaculture and ecological monitoring, and specifically to the technical field of a method and system for grading and identifying the feeding behavior of fish schools based on vibration signals. Background Art

[0002] In traditional aquaculture, fish feeding mainly relies on the experience and judgment of farmers. This method makes it difficult to observe the status of fish and changes in water quality, resulting in ineffective adjustment of feeding strategies, and thus causing many problems such as overfeeding or underfeeding.

[0003] The concept of precision feeding proposed for this purpose is to monitor the behavior of fish through technical means, accurately reflect their activity status and hunger level, and then adjust the feeding amount to ensure that fish obtain the best nutritional supply while effectively reducing food waste and reducing the degree of water pollution.

[0004] As some existing technical means for monitoring and identifying the feeding behavior of fish schools, there are the following implementation schemes. For example, patent CN118303353A, a feeding system and method for bottom-feeding fish based on machine vision, discloses a scheme for calculating the quantitative results of the feeding desire of fish schools based on the screenshots of the fish school feeding video.

[0005] For example, patent CN113040081A, an intelligent feeding decision system for recirculating aquaculture fish based on fish school swimming energy consumption analysis, discloses a solution for real-time analysis and evaluation of the swimming energy consumption of recirculating aquaculture fish using machine vision and deep learning technology.

[0006] For example, patent CN110412230A discloses a water pollution monitoring and early warning device based on fish behavior and its early warning method, which uses fish fin sensors, CCD cameras, photoelectric sensors and collision sensors as collection units to collect the average swimming speed, average number of turns, average activity water depth and total collision frequency of fish to determine the degree of water pollution.

[0007] The core idea of ​​these existing technologies is to use vision to obtain images of fish activities. However, field visits to freshwater aquaculture bases have found that the water turbidity of fish such as snakehead fish, which require "rich" water for aquaculture, is high, and the visibility of aquaculture water such as tilapia is very low. Therefore, the recognition accuracy of methods based on visual image technology will be greatly reduced in turbid waters or when there are too many obstructions.

[0008] In this regard, patents such as CN113592896A, a fish feeding method, system, equipment and storage medium based on image processing, CN215173990U, a device for monitoring fish behavior and activities in water and obtaining fish characteristic information, CN118075424A, a water ecology monitoring device and method for fish AI online identification, CN118323339A, a portable online monitoring device based on fish AI identification, and CN117911790A, a fish detection and identification method based on an improved YOLOX-nano algorithm, have proposed improvement solutions by improving the camera shooting angle and algorithmic operations on blurred images.

[0009] However, even with the improvement, it is still difficult to make timely, accurate and comprehensive judgments on changes in the water quality of the monitored environment caused by extreme environmental changes, such as sudden increases in precipitation. In addition, underwater shooting involves waterproof and high-definition cameras, as well as image processing computers, and the equipment costs are high. Therefore, the limited application areas and high costs of methods based on visual images have hindered the application and promotion of technology. Summary of the invention

[0010] In response to the problems in practical applications of the above-mentioned prior art, such as the judgment of fish behavior based on images, the degradation of image quality due to turbid water, which leads to poor accuracy in identifying fish feeding behavior, and expensive equipment costs, this application proposes a method and system for grading and identifying fish feeding behavior based on vibration signals, aiming to improve the management efficiency and accuracy of fish farming while reducing implementation costs.

[0011] To achieve the above objectives, this application adopts the following technical solutions:

[0012] In a first aspect, the present application provides a method for hierarchical identification of fish feeding behavior based on vibration signals, including controlling a plurality of vibration sensors to synchronously collect vibration signals caused by fish feeding;

[0013] Performing synthesis preprocessing on the vibration signal to obtain a synthetic signal;

[0014] The synthesized signal is filtered, framed, and subjected to wavelet transformation processing in sequence to extract features, thereby obtaining a first feature matrix;

[0015] Performing dimensionality reduction processing on the first characteristic matrix to obtain a second characteristic matrix;

[0016] The second feature matrix is ​​input into a pre-trained deep learning model to obtain a graded recognition signal of the vibration signal of fish feeding behavior.

[0017] Compared with the existing technology, this application is based on artificial intelligence and uses vibration sensors to detect vibrations caused by underwater fish schools, achieving precise feeding at a relatively low cost. It can be widely promoted and improve aquaculture efficiency.

[0018] The vibration signal of one-dimensional time series data is easily affected by various interference items and noise. Multiple vibration sensors are used to extract multi-dimensional signals, and the noise and interference are effectively reduced through an intelligent multi-dimensional signal processing chain, so that it can accurately judge the information of fish feeding behavior.

[0019] In a possible implementation, the synthesis pretreatment includes:

[0020] The collected multiple vibration signals are filtered separately, and the analog signals are converted into time series digital signals for storage;

[0021] Multiple time series digital signals are denoised and synthesized using a weighted method to obtain a synthetic signal.

[0022] Through synthesis and processing, the signal quality and feature detectability can be significantly improved, the computational efficiency can be increased, and the scope of application can be expanded. It also provides reliable basic data for further processing of subsequent data.

[0023] Furthermore, some embodiments further include the step of synthesizing the vibration signal to obtain a synthesized vibration signal Shi, wherein the synthesized vibration signal The obtained time series digital signal is stored in a buffer to be sent.

[0024] In a possible implementation manner, the filtering method for the composite signal is software filtering, and the frame processing method is adaptive frame processing.

[0025] In a preferred embodiment, the deep learning model is an LSTM deep learning model.

[0026] In a possible implementation, the method further includes the step of training the LSTM deep learning model before using the pre-trained LSTM deep learning model to classify the feeding behavior of the main characteristic signal. The training step of the LSTM deep learning model includes:

[0027] Build LSTM model;

[0028] Initialize the parameters of the LSTM deep learning model;

[0029] Load the sorted feature data into the model and perform forward propagation calculations and back propagation algorithms to calculate gradients;

[0030] Continuously adjust the parameters of the LSTM deep learning model to continuously reduce the loss function of the model, wherein the loss function is a cross entropy loss function;

[0031] Repeat the above steps until the loss function reaches the expected value and the training ends.

[0032] In a preferred embodiment, the deep learning model training adopts the TensorFlow deep learning framework.

[0033] In a second aspect, the present application also provides a fish feeding behavior classification identification system based on vibration signals, which uses the fish feeding behavior classification identification method based on vibration signals as described above, including a plurality of data acquisition modules for collecting vibration signals, a power supply module, a main control module for controlling the data acquisition module, and a human-computer interaction module for human-computer interaction with the main control module;

[0034] It also includes a data processing module for performing signal processing on the collected data and an edge computing device for performing deep learning and hierarchical identification on the processed data.

[0035] In some possible implementations of the second aspect, the data acquisition module includes a vibration sensor, a single-chip microcomputer, a wireless communication module, an ADC conversion module, and a filtering module to form an intelligent acquisition node, and the vibration sensor is an acceleration sensor.

[0036] In a preferred implementation of the second aspect, the power supply module is powered by a combination of lithium batteries and solar panels. The power supply module has a power monitoring unit and a control unit. The monitoring unit mainly performs power monitoring, and the control unit mainly realizes normal power supply under working conditions and is in node mode when idle. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is the application scenario layout diagram of the fish school feeding behavior classification identification system of the present application;

[0038] Figure 2 This is a schematic diagram of the fish feeding behavior classification identification system of the present application;

[0039] Figure 3 It is the overall workflow diagram of the hierarchical identification method of this application;

[0040] Figure 4 It is a signal processing flow chart. DETAILED DESCRIPTION

[0041] The features of the present application and other related features are further described in detail below through embodiments.

[0042] It should be noted that the terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0043] First, the scenarios applied in the application and the environmental information involved are described in detail.

[0044] Fertilizing water refers to promoting the growth of plankton by increasing nutrients (such as nitrogen, phosphorus, etc.) in water bodies in aquaculture, thereby providing sufficient natural bait for fish. This method can improve the productivity of water bodies and is very useful for certain situations that require high-density farming. However, it can also lead to decreased water transparency and affect underwater lighting. Therefore, there is a need for this application.

[0045] The vibration signals generated by a school of fish mainly include the vibrations generated by the swinging of the fish body, the flapping of the fins and tail, the activities of the mouth and gills, etc. When multiple fish feed at the same time, their collective movements and activities will be superimposed to form a more obvious vibration signal. Therefore, the vibration signals of a school of fish have complex characteristics, with obvious interference and noise. Therefore, the processing of vibration signals has become a key link in judging the feeding behavior of a school of fish.

[0046] The vibration signal is characterized by the fact that the vibration signal generated by the school of fish when feeding usually contains multiple frequency components. Low-frequency components may come from the large-scale movement of the fish body, and high-frequency components may come from the rapid flapping of the fins and tail and the activity of the mouth and gills. The amplitude of the vibration signal depends on the size of the fish, the intensity of activity and the density of the group. The duration of the vibration signal with the same frequency and amplitude can also indicate the progress time of the school of fish feeding.

[0047] Vibration sensors are devices that can collect these vibration signals of fish schools. A specific application example is an accelerometer that can be installed in water to detect vibrations in the water. By analyzing the data from the accelerometer, the activity of the fish school can be understood. A hydrophone is a sensor used to detect underwater sound waves and vibrations. By placing a hydrophone, the vibration signals generated by the fish school when feeding can be recorded. There are also sonar systems that can emit sound waves and receive the reflected signals. By analyzing these signals, the location and activity of the fish school can be detected. The application can be selected according to the actual situation.

[0048] Please refer to Figure 1 This is a scene layout diagram of fish feeding used in this application. Intelligent collection points are set around the food feeding area. The intelligent collection points are equipped with vibration sensors. Acceleration sensors are preferred in this application. Acceleration sensors have the advantages of high sensitivity, multi-axis detection, low power consumption, miniaturization and integration, strong anti-interference ability, strong data processing and analysis capabilities, and high cost-effectiveness in sensing fish vibrations.

[0049] In this case, it is preferred to feed in an open area and use 4 vibration sensors around. In other environments, the number of vibration sensors can be increased or decreased according to demand. As mentioned in the background technology, the fish that are likely to cause increased underwater turbidity include black carp, tilapia, etc., which have the characteristics of living in groups, strong predation, and moving in the middle and lower layers of water. Therefore, it is also preferred to feed in an open area, that is, the feeding area is far away from the shore, which is one of the reasons why it is difficult to identify the feeding progress manually.

[0050] Please refer to Figure 2 , power supply is a problem that needs to be solved when feeding in open areas. Although the farm has insulated cables that can supply power, when it is arranged at multiple points, it also increases safety hazards. Battery power supply or solar power supply can be selected as the power supply of the data acquisition module. The data acquisition module consists of a vibration sensor, a single-chip microcomputer, a wireless communication module, an ADC (Analog-to-Digital Converter) conversion module, and a filter module to form an intelligent acquisition node. The vibration sensor and ADC conversion module are controlled by a single-chip microcomputer.

[0051] The single-chip microcomputer is preferably a low-power single-chip microcomputer, which can be a 32-bit single-chip microcomputer. It is a mature technology in the industry and will not be described in detail. Zigbee wireless communication is preferably used in the wireless communication module, and its short-distance and low-power wireless communication characteristics are used for wireless communication in open waters in the field of aquaculture.

[0052] The multiple data acquisition modules are controlled by the main control module, and each data acquisition module is connected to the active module through wireless communication. The main control module sends control instructions through the human-computer interaction module, which can be a dedicated control device, including a high-brightness digital display and a remote control command input device. Or it can also be achieved using general devices such as mobile phones and computers.

[0053] When solar energy is used for power supply in an application, a combination of solar energy and lithium batteries can be considered.

[0054] The power supply module has a power monitoring unit and a control unit. The monitoring unit mainly completes power monitoring, and the control unit mainly realizes normal power supply under working conditions and is in node mode when idle. In specific use, when the microcontroller receives the acquisition instruction from the main control module, it controls the power supply unit to power all circuits of the node. When idle, the control node is in energy-saving mode, and the solar panel charges the lithium battery.

[0055] Furthermore, after the vibration data is collected, it is processed by the data processing module, and then the edge computing device completes the deep learning machine to perform vibration signal classification identification. Then it is fed back to the main control module for further monitoring.

[0056] Based on the above-mentioned fish feeding behavior classification identification system, the following introduces a method for using it to classify and identify fish feeding behavior. Using a vibration sensor to detect the vibration caused by underwater fish can achieve accurate feeding at a relatively low cost, so that it can be widely promoted and improve aquaculture efficiency.

[0057] Please refer to Figure 3 The main working process is to first control multiple vibration sensors to collect vibration signals caused by fish feeding. When working, the vibration sensors are sent by the main control module to collect data synchronously. Then, the vibration signals are synthesized and preprocessed to obtain a synthetic signal.

[0058] For the signal processing flow, please refer to Figure 4 The method for synthesizing and preprocessing the vibration signal includes first filtering the collected multiple vibration signals separately, and then converting the analog signal into a time series digital signal for storage. Preferably, it also includes a signal synthesis function. When the vibration sensor selects an acceleration sensor, the three signals X, Y, and Z of the sensor are synthesized to obtain a synthetic vibration signal of the position. The obtained time series digital signal is stored in a buffer to be sent.

[0059] After that, multiple time series digital signals are denoised and synthesized using a weighted method to obtain a synthetic signal. Specifically, the main control module collects the time series digital signals in the intelligent collection node through the zigbee wireless communication module. The time series digital signals of each intelligent collection node are synthesized using a weighted method to obtain a synthetic signal S=a1Sh1+a2Sh2+a3Sh3+a4Sh4. At this time, a1 and others represent their weighted proportions, which can be a comprehensive evaluation based on factors such as the position of the vibration sensor in the entire system or the positional relationship with the feeding direction.

[0060] In this way, through synthesis and processing, the quality of the signal and the detectability of the features can be significantly improved, the computing efficiency can be improved, the application scope can be expanded, and reliable basic data can be provided for the further processing of subsequent data.

[0061] Then, the synthesized signal is filtered, framed, and processed by wavelet transform in sequence to extract features, thereby obtaining a first feature matrix.

[0062] Specifically, the synthesized signal is sent to the data processing module for signal processing. The signal processing is preferably processed by DSP (Digital Signal Processing). Of course, other processors that can achieve equivalent functions can also be used. First, the synthesized signal is filtered to ensure that the quality and characteristics of the signal meet the requirements of subsequent processing, prevent aliasing, remove noise, and select the required frequency band. For example, mechanical vibrations caused by the surrounding environment and feeding equipment can be removed by software filtering. In actual application, it is not limited to this and can be selected according to working conditions.

[0063] The filtered signal is then processed by framing. Specifically, adaptive framing can be used. This is a technology that dynamically adjusts the frame length and frame shift according to the characteristics of the signal. It can better capture the transient characteristics of the signal and improve the accuracy and robustness of signal analysis.

[0064] The framing process is to perform framing processing on the synthesized signal to obtain the data of each frame, which is recorded as Si = {s(k)|k = ti_start, ti_start+1, ..., ti_end}, where i represents the i-th frame, ti_start and ti_end are the start and end time of the i-th frame respectively, k is the index of the time point, and s(k) represents the signal value at the time point k. In this way, the original signal is divided into several shorter time periods (frames) to facilitate subsequent localized analysis.

[0065] After that, each frame signal Si is subjected to wavelet transform one by one to obtain wavelet transform coefficients. The wavelet coefficients of the i-th frame signal at the j-th scale and the k-th position are recorded as ci,j,k. The wavelet coefficients are used to extract features of the signal to obtain the first feature matrix M.

[0066] i represents the i-th frame signal. A frame refers to dividing a long signal into several shorter segments. Each frame contains a certain number of sampling points. j represents the j-th scale. The scale parameter determines the width of the wavelet basis function. The k-th position. The position parameter determines the translation of the wavelet basis function on the time axis.

[0067] The first feature matrix M is a two-dimensional matrix, in which each row or column represents a feature vector, which contains the features extracted from the wavelet coefficients. Specifically, each row represents the feature vector of a frame signal. Each column represents the features of all frame signals at the same scale.

[0068] As mentioned above, in addition to environmental information, fish vibration information also contains signals of different frequencies in fish movement itself. Therefore, in order to remove noise and redundant information in the feature matrix, improve the signal-to-noise ratio of the signal, and make the features more prominent, the first feature matrix M is subjected to dimensionality reduction to obtain the second feature matrix N. The second feature matrix N after dimensionality reduction is easier to be processed by the machine learning model, preventing overfitting and improving the generalization ability and classification accuracy of the model.

[0069] Specifically, the first feature matrix M is standardized to obtain a new first feature matrix M', the covariance matrix C of M' is calculated, and the eigenvalue decomposition is performed on C to obtain the eigenvalues ​​and eigenvectors. The eigenvectors corresponding to the first K eigenvalues ​​are selected from large to small as the main features to obtain the principal component matrix E, where the K value is the number of features whose contribution rate (i.e., the ratio of the sum of the first k eigenvalues ​​to the total sum of eigenvalues) reaches the threshold value ɑ.

[0070] The first feature matrix M is projected onto the new first feature matrix M', spatially, to obtain the second feature matrix N after dimensionality reduction, N = EM'. The second feature matrix N is input into a pre-trained deep learning model to obtain a graded recognition signal of the vibration signal of the fish feeding behavior.

[0071] In this way, the problem that the vibration signal of one-dimensional time series data is easily affected by various interference items and noise is solved by using multiple vibration sensors to extract multi-dimensional signals, and the noise and interference are effectively reduced through a multi-dimensional signal processing chain, so that it can accurately determine the information of fish feeding behavior.

[0072] The deep learning model is described in detail as follows. Preferably, the deep learning model used in the hierarchical recognition method is an LSTM (Long Short-Term Memory) deep learning model. Using the LSTM model to process the vibration signals generated when the fish are feeding can effectively capture and memorize the temporal characteristics of the vibration signals, and improve the prediction and classification accuracy of the model.

[0073] First, build an edge computer system, which mainly includes key devices such as processors, network interface cards, and memory. The processor is responsible for completing all computing tasks. The memory is a hardware device used to store data and programs.

[0074] The second feature matrix N after dimension reduction is normalized and data cleaned, and the main feature data is divided into a training set, a validation set, and a test set, and the training samples are labeled. The training samples are labeled with the corresponding feeding behavior type (such as normal feeding, fast feeding, slow feeding, etc.) for each sample in the training set. The step of training the LSTM deep learning model is performed before using the pre-trained LSTM deep learning model to classify the feeding behavior of the main feature signal.

[0075] Training the model with the training set includes the following steps:

[0076] Build the LSTM model structure, including the input layer, LSTM layer, fully connected layer, and output layer, using the deep learning framework TensorFlow. The TensorFlow framework can be used to easily build, train, and deploy LSTM models, providing efficient computing power and rich tool support to ensure high performance and scalability of the model.

[0077] Set the network parameters, use Adam (Adaptive Moment Estimation) as the optimizer, use cross entropy loss as the loss function, input the second feature matrix N into the model for training, optimize the parameters, and repeat the process of forward propagation, loss calculation, and back propagation until the predetermined loss threshold is reached.

[0078] The specific steps of the training process are as follows:

[0079] Build LSTM model;

[0080] Initialize the parameters of the LSTM deep learning model;

[0081] Load the sorted feature data into the model and perform forward propagation calculations and back propagation algorithms to calculate gradients;

[0082] Continuously adjust the parameters of the LSTM deep learning model to continuously reduce the loss function of the model, wherein the loss function is a cross entropy loss function;

[0083] Repeat the above steps until the loss function reaches the expected value and the training ends.

[0084] After that, the validation set is used to evaluate the model performance, the model is deployed, the operation mode is started, the collected vibration signal is acquired and the current fish feeding behavior intensity level is predicted, and the output is displayed in the human-computer interaction module.

[0085] As described above, the present application protects a method and system for grading and identifying the feeding behavior of fish schools based on vibration signals. All technical solutions that are identical or similar to the present case should be deemed to fall within the protection scope of the present case.

Claims

1. A method for classifying and identifying the feeding behavior of fish schools based on vibration signals, characterized in that: It includes controlling multiple vibration sensors to synchronously collect vibration signals caused by fish feeding; Performing synthesis preprocessing on the vibration signal to obtain a synthetic signal; The synthesized signal is filtered, framed, and subjected to wavelet transformation processing in sequence to extract features, thereby obtaining a first feature matrix; Performing feature selection and dimensionality reduction processing on the first feature matrix to obtain a second feature matrix; The second feature matrix is ​​input into a pre-trained deep learning model to obtain a graded recognition signal of the vibration signal of fish feeding behavior.

2. The method for classifying and identifying the feeding behavior of a school of fish based on vibration signals according to claim 1, characterized in that: The synthesis pretreatment comprises: The collected multiple vibration signals are filtered separately, and the analog signals are converted into time series digital signals for storage; The multi-channel signals of a single sensor are synthesized by vector synthesis method to obtain the vibration synthesis vibration signal Shi at the location of the single sensor; The synthetic signals of multiple sensors arranged in the feeding area are synthesized by weighted method to obtain a synthetic signal S.

3. The method for classifying and identifying the feeding behavior of a school of fish based on vibration signals according to claim 2, characterized in that: The invention also includes the step of synthesizing the single multi-path vibration signal to obtain a synthesized vibration signal Shi, wherein the synthesized vibration signal Then, the synthetic signals of multiple sensors are synthesized, wherein the synthetic vibration signal S=a1Sh1+a2Sh2+a3Sh3+a4Sh4, and the obtained time series digital signal is stored in a buffer to be sent.

4. The method for classifying and identifying the feeding behavior of a school of fish based on vibration signals according to claim 1, characterized in that: The filtering method for the composite signal is software filtering, and the frame processing method is adaptive frame processing.

5. The method for classifying and identifying the feeding behavior of a school of fish based on vibration signals according to claim 1, characterized in that: The deep learning model is an LSTM deep learning model.

6. The method for classifying and identifying the feeding behavior of a school of fish based on vibration signals according to claim 5, characterized in that: The method further includes the step of training the LSTM deep learning model before using the pre-trained LSTM deep learning model to classify the feeding behavior of the main characteristic signal. The training step of the LSTM deep learning model includes: Label and preprocess feature data; Divide the data set into training set, validation set and test set; Build LSTM model; Initialize the parameters of the LSTM deep learning model; Load the sorted feature data into the model and perform forward propagation calculations and back propagation algorithms to calculate gradients; Continuously adjust the parameters of the LSTM deep learning model to continuously reduce the loss function of the model, wherein the loss function is a cross entropy loss function; Repeat the above steps until the loss function reaches the expected value and the training ends.

7. The method for classifying and identifying the feeding behavior of a school of fish based on vibration signals according to claim 5, characterized in that: The deep learning model training adopts the TensorFlow deep learning framework.

8. A fish feeding behavior classification identification system based on vibration signals, characterized in that: Applying the method for grading and identifying the feeding behavior of a school of fish based on vibration signals as described in any one of claims 1 to 7; It also includes a plurality of data acquisition modules for collecting vibration signals, a power supply module, a main control module for controlling all data acquisition modules, and a human-computer interaction module for human-computer interaction with the main control module; It also includes a data processing module for performing signal processing on the collected data and an edge computing device for performing deep learning and hierarchical identification on the processed data.

9. The fish feeding behavior classification identification system based on vibration signals according to claim 8, characterized in that: The data acquisition module comprises a vibration sensor, a single chip microcomputer, a wireless communication module, an ADC conversion module, and a filtering module, forming an intelligent acquisition node. The vibration sensor is an acceleration sensor.

10. The fish feeding behavior classification identification system based on vibration signals according to claim 8, characterized in that: The power supply module is powered by a combination of lithium batteries and solar panels. The power supply module has a power monitoring unit and a control unit. The monitoring unit mainly completes power monitoring, and the control unit mainly realizes normal power supply under working conditions and is in energy-saving mode when idle.

Images