CN117473445B

CN117473445B - Extreme learning machine-based equipment abnormality analysis method and device

Info

Publication number: CN117473445B
Application number: CN202311819988.XA
Authority: CN
Inventors: 许育锋; 刘冬亮; 徐坤扬; 安磊; 林永桐; 徐凌子
Original assignee: Shenzhen Mingxin Digital Intelligence Technology Co ltd
Current assignee: Shenzhen Mingxin Digital Intelligence Technology Co ltd
Priority date: 2023-12-27
Filing date: 2023-12-27
Publication date: 2024-04-16
Anticipated expiration: 2043-12-27
Also published as: CN117473445A

Abstract

The application relates to the technical field of artificial intelligence and discloses a device abnormality analysis method and device based on an extreme learning machine. The method comprises the following steps: acquiring a data set of the working state of the cultivation equipment in the cultivation water area to be detected; decomposing the working state data set of the cultivation equipment through a preset operation data decomposition model to obtain a plurality of working state decomposition results; inputting the multiple working state decomposition results into a preset extreme learning machine model to perform feature extraction to obtain multiple target dimension feature vectors; inputting the target dimension feature vectors into a preset abnormal detection model of the cultivation equipment to detect the abnormal operation of the cultivation equipment, and obtaining an equipment abnormal classification result; and creating a corresponding abnormal processing strategy of the cultivation equipment according to the abnormal classification result of the equipment, so that the accuracy and the efficiency of abnormal processing of the cultivation equipment are improved.

Description

Extreme learning machine-based equipment abnormality analysis method and device

Technical Field

The application relates to the technical field of artificial intelligence, in particular to a device abnormality analysis method and device based on an extreme learning machine.

Background

Aquaculture is an important agricultural and food production area, with great significance to global food supply. However, abnormal conditions of the farming equipment cause production interruption, resource waste, and environmental problems. Therefore, developing an efficient method for analyzing equipment anomalies of aquatic products based on extreme learning machines is important for improving production efficiency, reducing loss and maintaining environment.

With the development of sensor technology and data acquisition technology, aquaculture industry is beginning to utilize big data and intelligent monitoring to improve production efficiency. This data driven approach provides new opportunities for anomaly detection and device fault diagnosis. However, the farming equipment is affected by various factors including mechanical failure, environmental changes, water quality problems, etc., which lead to equipment anomalies and production problems, which in turn lead to low accuracy of existing solutions.

Disclosure of Invention

The application provides an equipment abnormality analysis method and device based on an extreme learning machine, which are used for improving the accuracy and efficiency of abnormal treatment of cultivation equipment.

The first aspect of the present application provides an equipment abnormality analysis method based on an extreme learning machine, including:

Acquiring a data set of the working state of the cultivation equipment in the cultivation water area to be detected;

decomposing the working state data set of the cultivation equipment through a preset operation data decomposition model to obtain a plurality of working state decomposition results;

inputting the multiple working state decomposition results into a preset extreme learning machine model to perform feature extraction to obtain multiple target dimension feature vectors;

inputting the target dimension feature vectors into a preset abnormal detection model of the cultivation equipment to detect the abnormal operation of the cultivation equipment, and obtaining an equipment abnormal classification result;

and creating a corresponding abnormal treatment strategy of the cultivation equipment according to the abnormal classification result of the equipment.

A second aspect of the present application provides an apparatus abnormality analysis apparatus based on an extreme learning machine, the apparatus abnormality analysis apparatus based on an extreme learning machine including:

the acquisition module is used for acquiring a data set of the working state of the cultivation equipment in the cultivation water area to be detected;

the decomposition module is used for decomposing the working state data set of the cultivation equipment through a preset operation data decomposition model to obtain a plurality of working state decomposition results;

the extraction module is used for respectively inputting the multiple working state decomposition results into a preset extreme learning machine model to perform feature extraction to obtain multiple target dimension feature vectors;

The detection module is used for inputting the target dimensional feature vectors into a preset abnormal detection model of the cultivation equipment to detect the abnormal operation of the cultivation equipment, so as to obtain an abnormal classification result of the equipment;

the creating module is used for creating a corresponding abnormal processing strategy of the cultivation equipment according to the abnormal classification result of the equipment.

In the technical scheme provided by the application, a data set of the working state of the cultivation equipment in a cultivation water area to be detected is obtained; decomposing the working state data set of the cultivation equipment through a preset operation data decomposition model to obtain a plurality of working state decomposition results; inputting the multiple working state decomposition results into a preset extreme learning machine model to perform feature extraction to obtain multiple target dimension feature vectors; inputting the target dimension feature vectors into a preset abnormal detection model of the cultivation equipment to detect the abnormal operation of the cultivation equipment, and obtaining an equipment abnormal classification result; according to the abnormal classification result of the equipment, a corresponding abnormal treatment strategy of the aquaculture equipment is created, the abnormal condition of the equipment is monitored and detected in real time, production interruption can be reduced by timely finding and treating the abnormal condition, normal operation of the equipment is ensured, the growth environment of aquatic animals and plants can be improved by timely adjusting the factors such as water quality and temperature, and the aquaculture yield and quality are improved. By early recognition of abnormal conditions, resource waste caused by equipment faults or bad environmental conditions can be reduced, and intelligent decision support is provided. According to the abnormal conditions, suggestions and strategies can be provided for the aquaculture, and the accuracy and efficiency of the abnormal treatment of the aquaculture equipment are improved.

Drawings

FIG. 1 is a schematic diagram of one embodiment of an equipment anomaly analysis method based on an extreme learning machine in an embodiment of the present application;

fig. 2 is a schematic diagram of an embodiment of an apparatus abnormality analysis device based on an extreme learning machine in an embodiment of the present application.

Detailed Description

The embodiment of the application provides an equipment abnormality analysis method and device based on an extreme learning machine, which are used for improving the accuracy and efficiency of abnormal treatment of cultivation equipment. The terms "first," "second," "third," "fourth" and the like in the description and in the claims of this application and in the above-described figures, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged where appropriate such that the embodiments described herein may be implemented in other sequences than those illustrated or otherwise described herein. Furthermore, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed or inherent to such process, method, article, or apparatus.

For ease of understanding, a specific flow of an embodiment of the present application will be described below, referring to fig. 1, and one embodiment of an apparatus anomaly analysis method based on an extreme learning machine in an embodiment of the present application includes:

s101, acquiring a data set of the working state of the cultivation equipment in a cultivation water area to be detected;

it is to be understood that the execution subject of the present application may be an apparatus abnormality analysis device based on an extreme learning machine, or may be a terminal or a server, which is not limited herein. The embodiment of the present application will be described by taking a server as an execution body.

Specifically, the server first determines the source and the acquisition mode of the data of the cultivation equipment. The farming equipment operation data may be collected from a number of sources including sensor data, equipment logs, remote monitoring systems, and manual recordings. Sensor data is one of the most common sources because they are capable of providing real-time device status and environmental parameters. For example, in fish farming, a water quality sensor may be used to measure the pH, dissolved oxygen content, and temperature in water. The sensors may periodically collect data and send it to a data storage system. On the other hand, the device log records the running state, operation and event of the device. These logs, typically in the form of text files or database records, may provide information about device performance and potential anomalies. The remote monitoring system allows remote access to the device data, transmitting real-time data over a network, which is very useful for remote monitoring and real-time anomaly detection. In addition, manual recording by a human operator is also part of the data. The operator will periodically record key parameters, events and observations. These manual recordings may complement automatic data collection and provide manual insight regarding the operation of the device. Once the server determines the source of the data, it then establishes a data storage and management system. The choice of data storage location includes a local server, cloud storage, or a specialized database system, depending on the amount of data, access requirements, and security requirements. In addition, the format and structure of the data needs to be defined to ensure that the data is easy to process and analyze. This may involve creating a data table, using JSON files, CSV files, or other suitable data formats. Determining the data acquisition frequency is also one of the key factors. Some data requires high frequency acquisition in order to detect anomalies in time, while other data may be acquired more sparsely to reduce storage and processing costs. At the same time, it is important to implement data quality control measures to detect and process errors or outliers in the data to ensure accuracy and consistency of the data.

S102, decomposing a working state data set of the cultivation equipment through a preset operation data decomposition model to obtain a plurality of working state decomposition results;

specifically, the server classifies the working state data set of the cultivation equipment by using a preset parameter clustering model so as to obtain a plurality of initial equipment operation data. The initial data includes data of different parameters such as temperature data, humidity data, water quality data and equipment vibration data. For example, consider a farming equipment monitoring system that includes a temperature sensor, a humidity sensor, a water quality sensor, and a vibration sensor. The system records the values of these parameters every minute. By using a clustering model, the server classifies these data into different initial device operational data. For example, the model would classify temperature and humidity data as one set, water quality data as another set, and vibration data as another set. The server obtains a plurality of initial operational status data sets of the farming equipment. The server performs time sequence association processing on each initial device operation data to obtain a plurality of target device operation data. This step may include data interpolation, smoothing, or other time series processing techniques to ensure data continuity and consistency. For example, for temperature and humidity data, the server performs interpolation to fill in existing data missing or discontinuities. For water quality data, the server performs smoothing processing to remove noise and outliers. For vibration data, the server performs a spectral analysis to extract frequency domain features. These processing steps will allow the server to obtain multiple target device operational data, each with better timing continuity. The server performs parameter initialization on each target device operational data in preparation for complementary aggregate empirical mode decomposition. For example, for temperature data, parameter initialization may include determining an initial temperature range and a reference temperature value. For water quality data, the initialization may involve defining criteria and allowed ranges for water quality parameters. For vibration data, the initialization may include an estimation of vibration amplitude and frequency. These initialization parameters will be used in the subsequent data decomposition process. And the server uses a preset complementary set empirical mode decomposition model to decompose the operation data of each target device according to the initialization parameter set, so as to obtain a plurality of working state decomposition results of the operation data of each target device. The complementary set empirical mode decomposition is a signal processing technique that can decompose complex time series data into operating state decomposition results of different frequencies and amplitudes to better understand the characteristics and variations of the data. For example, assuming that the server performs a complementary set of empirical mode decomposition on the temperature data, the server obtains a plurality of operating state decomposition results, each function representing a temperature change over a different frequency range. For example, the server obtains a high frequency operating condition decomposition result indicating rapid temperature fluctuations, and a low frequency operating condition decomposition result indicating long-term trends in temperature. Similarly, for water quality data and vibration data, the server may also obtain a corresponding set of operating state decomposition results.

And setting the iteration times M and the convergence threshold epsilon by the server according to a preset complementary set empirical mode decomposition model. These parameters will be used in the subsequent decomposition process. For example, assume that the server sets the number of iterations M to 100 and the convergence threshold epsilon to 0.01. The server performs 100 iterations, each of which checks if the one-dimensional cumulative energy curve is less than 0.01, to determine if it converges. The server uses a preset mean curve calculation function to calculate a mean curve ci (t) for each target device operational data. The mean curve is obtained by performing multiple iterations of the target device operational data and a weighted average of the noise function. For example, for temperature data, the server performs an iterative process, each iteration calculating a mean curve ci (t) from the current noise function hm (t). These mean curves represent the average trend of the temperature data. Using the one-dimensional cumulative energy curve calculation function, the server calculates a one-dimensional cumulative energy curve Ei (t) for each target device operation data. This curve is used to evaluate the energy distribution of the data and for the subsequent decision process. For example, for humidity data, the server calculates a one-dimensional cumulative energy curve Ei (t) that represents the energy difference between the humidity data and the mean curve ci (t). The smaller Ei (t) value indicates how close the data is to the mean. The server determines whether the one-dimensional accumulated energy curve Ei (t) is smaller than a set convergence threshold epsilon. If Ei (t) is less than ε, the specification data has converged and the server continues to generate operational state decomposition results. Otherwise, the server will perform the next step to further process the data. For example, assuming that the server judges the humidity data, ei (t) =0.005 is found to be smaller than ε (0.01). Thus, the data has converged and the server continues to generate the working state decomposition result. If the one-dimensional accumulated energy curve Ei (t) is less than ε, indicating that the data has converged, the server generates a plurality of operating state decomposition results using a complementary set of empirical mode decomposition models, which capture the main features of the data. For example, for wetness data, the server generates a plurality of operating state decomposition results, each of the modal functions representing a different frequency and amplitude component of the wetness data. These modal functions help to understand the pattern of changes in the humidity data. If the one-dimensional accumulated energy curve Ei (t) is greater than or equal to ε, it is indicated that the data has not converged. In this case, the server will calculate a one-dimensional extremum curve Mi (t). For example, for vibration data, if Ei (t) is greater than ε, the server calculates a one-dimensional extremum curve Mi (t) that helps the server identify extremum points in the data. These extreme points are helpful in capturing important characteristics of the vibration. And the server repeats the steps according to the setting of the iteration times M until the designated iteration times are reached. The resulting operating state decomposition results may thus be further refined and improved to better reflect the characteristics of the target device operational data. For example, if the server sets the number of iterations to 100, 100 iterations will be performed. Each iteration generates a new working state decomposition result to continuously optimize the representation of the data.

S103, respectively inputting a plurality of working state decomposition results into a preset extreme learning machine model to perform feature extraction to obtain a plurality of target dimension feature vectors;

the server presets an extreme learning machine (Extreme Learning Machine, ELM) model. ELM is a machine learning model that generally includes an input layer, a plurality of hidden layers, and an output layer. Here, the server will use ELM for feature extraction to translate the operational state decomposition results into high-dimensional feature vectors. For example, assume that the server has an ELM model that includes an input layer, two hidden layers, and an output layer. Each operating state decomposition result will be input into the model for feature extraction. The server inputs a plurality of working state decomposition results of the operation data of each target device into a preset ELM model respectively so as to extract the characteristics. The hidden layer of the ELM model will automatically learn the feature information in the working state decomposition result and map it into the high-dimensional feature space. For example, for the operating state decomposition result of the temperature data, the server inputs it into the ELM model. The hidden layer of the model will learn how to extract features of the temperature data, such as trends, periodicity, and peaks and valleys of the temperature, from the operating state decomposition results. These features will be mapped into high-dimensional feature vectors. After each working state decomposition result passes through the ELM model, an initial high-dimensional characteristic vector is obtained. To obtain a combined high-dimensional feature vector for each target device operational data, the server combines these initial feature vectors. This may be a simple vector connection or other suitable means. For example, after the ELM model is passed through, the server connects the initial high-dimensional feature vectors of the working state decomposition result of the humidity data and the working state decomposition result of the water quality data to form a combined high-dimensional feature vector. This vector will contain characteristic information from the decomposition results of the different operating states. Weight optimization analysis, the server uses a preset goblet sea squirt optimization algorithm (Sea Cucumber Optimization Algorithm) to perform weight optimization analysis on the combined high-dimensional feature vector of each target equipment operation data. The goal of this optimization process is to find the best weight distribution so that the high-dimensional feature vectors better reflect the characteristics of the plant operational data. For example, assume that the server uses a goblet sea squirt optimization algorithm to optimize the weight distribution of the combined high-dimensional eigenvectors of temperature, humidity, and water quality data. The goal of the optimization is to maximize the distinguishability of the feature vectors to better capture the characteristics and anomalies of the device operational data. And obtaining a plurality of target dimension feature vectors by the server through weight optimization analysis. Each vector contains characteristics extracted from decomposition results of different working states, and the characteristics and abnormal conditions of the equipment operation data can be better reflected after weight optimization. For example, in this embodiment, the server obtains a plurality of target dimensional feature vectors, each representing a feature of different device operational data. These feature vectors may be used for subsequent plant anomaly detection to identify plant anomalies and take appropriate action.

Wherein a plurality of operating state decomposition results for each target device operational data are to be input into an input layer of the ELM model. These operating condition decomposition results represent different characteristics and frequency components of the device operating data. For example, for temperature data, the server has a plurality of operating state decomposition results, each of the modal functions representing a different frequency component of the temperature data, such as rapid fluctuations and long-term trends. These modal functions will serve as inputs to the ELM model. Multiple hidden layers of the ELM model will be used to extract hidden features for each operational state decomposition result. These hidden layers will automatically learn the characteristic information in the working state decomposition result, such as frequency, amplitude, time sequence pattern, etc. For example, in a first hidden layer of an ELM model, the model may learn to extract frequency features of temperature data, while in a second hidden layer, the model may learn to extract other features such as amplitude and timing patterns. And the output layer of the ELM model performs vector combination on the initial high-dimensional feature vectors obtained in the hidden layer according to preset initial vector weights, so as to generate combined high-dimensional feature vectors of the operation data of each target device. For example, assume that the server has a preset vector weight, e.g., [0.5,0.3,0.2], for combining feature vectors in the hidden layer. For the operating state decomposition results of the temperature data, the ELM model will use these weights to combine the extracted frequency, amplitude and timing characteristics to generate a combined high-dimensional feature vector of the temperature data. Through the vector combining process, the server successfully generates a combined high-dimensional feature vector for each target device operational data. The high-dimensional feature vectors contain feature information from decomposition results of different working states, and the feature information better reflects the characteristics and abnormal conditions of equipment operation data after weight combination. For example, the above steps are repeated for the humidity data, the water quality data, and the vibration data, respectively, to generate combined high-dimensional feature vectors of these data, respectively. Each vector will represent a characteristic of the different device operational data.

The server uses a preset sea-squirt optimizing algorithm to initialize the weight parameters of the combined high-dimensional feature vector of the operation data of each target device. This initialization process will generate an initial population of goblets, including a plurality of first candidate vector weights. For example, suppose the server has three combined high-dimensional feature vectors of target device operational data, V1, V2, and V3, respectively. The server generates an initial weight vector for each vector using a goblet sea squirt optimization algorithm, resulting in an initial goblet sea squirt population. This population includes a plurality of first candidate vector weights, such as W1, W2, and W3. The server performs fitness calculation on each first candidate vector weight in the initial goblet sea squirt population. Fitness functions are typically defined according to specific tasks or optimization objectives for evaluating the performance of each vector weight. For example, in this embodiment, the fitness function may be a performance metric for equipment anomaly detection, such as accuracy, recall, or F1 score. Each vector weight will be used to calculate a performance metric for the combined high-dimensional feature vector for the corresponding target device operational data. And according to the calculated fitness, the server performs position updating on the initial goblet-sea squirt population by using a goblet-sea squirt optimization algorithm to obtain a plurality of second candidate vector weights. This location update procedure aims to adjust the weights according to the fitness to improve performance. For example, assume that after fitness calculation, the server gets updated weight vectors, such as W1', W2', and W3'. These vector weights have been updated by location to better meet the performance metrics of device anomaly detection. An optimal solution, i.e., a target vector weight, is determined from the plurality of second candidate vector weights. This optimal solution is selected based on fitness, typically the vector weight with the best performance metric. For example, in this embodiment, based on fitness calculations, the server determines the best weight vectors, such as W1', W2', and W3', where one or more vectors have the best performance metrics. These optimal weight vectors will be used for subsequent weight optimization calculations. And carrying out weight optimization calculation on the combined high-dimensional feature vector of the operation data of each target device by using the determined weight of the target vector. The calculation process generates a plurality of target dimension characteristic vectors which better reflect the characteristics and abnormal conditions of the equipment operation data after weight optimization. For example, by applying optimal weight vectors, such as W1 ", W2", and W3 ", the server optimizes the combined high-dimensional feature vector for each target device operational data to obtain multiple target-dimensional feature vectors. These vectors will be used for the detection and analysis of the anomalies in the farming plant to improve the accuracy of the anomaly detection.

S104, inputting a plurality of target dimension feature vectors into a preset abnormal detection model of the cultivation equipment to detect the abnormal operation of the cultivation equipment, and obtaining an equipment abnormal classification result;

specifically, the server establishes a preset abnormal detection model of the cultivation equipment. This model is a deep learning model that includes multiple LSTM layers and two layers of fully connected networks. LSTM is used to process sequence data, while fully connected networks are used for final anomaly detection classification. For example, the farm anomaly detection model includes two LSTM layers and two fully connected layers. This model is trained to identify anomalies in the equipment operation data. The server inputs the target dimension feature vectors into a preset abnormal detection model of the cultivation equipment. These high-dimensional feature vectors are generated in the previous steps and contain feature information extracted from the equipment operation data. For example, assume that the server has three high-dimensional feature vectors of the target devices, F1, F2, and F3, respectively. These vectors will be input into the aquaculture device anomaly detection model for anomaly detection. In the aquaculture device anomaly detection model, a plurality of LSTM layers will be used to perform feature classification operations on the high-dimensional feature vectors. LSTM is a recurrent neural network adapted for sequential data that can effectively capture timing information in the data. For example, each of the high-dimensional feature vectors F1, F2, and F3 will undergo a feature classification operation through the LSTM layer to better understand its timing characteristics and potential anomalies. This will generate a plurality of feature classification operand values for subsequent anomaly detection. And inputting the multiple feature classification operation values into a two-layer fully-connected network for abnormality detection. The fully connected network comprehensively considers the characteristic classification operation values to obtain the abnormal classification result of the equipment. For example, through two layers of fully connected networks, the server integrates the feature classification operation values to obtain an abnormal classification result of the equipment. For example, if the integrated result indicates that the operational data of a certain device does not match the normal condition, it will be classified as an abnormal state.

S105, creating a corresponding abnormal treatment strategy of the cultivation equipment according to the abnormal classification result of the equipment.

In particular, it is desirable to collect the results of device anomaly classification, which is typically accomplished through a device monitoring system, sensor, or anomaly detection model. These results are classified into different categories, typically including normal, warning, and abnormal conditions. Based on these classification results, different exception handling policies are formulated, each requiring specific countermeasures. These strategies should be defined in terms of best practices and related standards for the farming industry to ensure reliability, safety and efficiency of the equipment. For example, when the equipment is classified as normal, normal production or operation flows may continue, but periodic monitoring and maintenance is still required to ensure that the equipment remains in good condition. While preventive measures such as increasing the frequency of monitoring, performing more frequent maintenance, adjusting operating parameters, etc. are required when the equipment is classified as a warning situation. At the same time, the relevant personnel need to be notified to ensure that the problem is of sufficient concern. When the equipment is classified as abnormal, immediate emergency actions such as shutdown maintenance, emergency maintenance scheduling, use of spare equipment, etc. are required. At the same time, it is necessary to quickly notify a maintenance team or related responsible personnel in order to take emergency action. In some cases, an automated system may be used to implement these exception handling policies. For example, the device monitoring system may automatically detect anomalies and trigger corresponding operations, thereby reducing the need for human intervention and improving the speed of the reaction. For example, assume that an aquaculture farm is operated and an automatic water quality monitoring system is used. If the system detects that the water quality is abnormal, the following strategies can be automatically executed: if the water quality is slightly wrong, the system can automatically adjust the water quality control equipment to correct the problem; if the water quality problem is serious, the system can automatically stop the water pump and inform relevant staff to check and repair; if the water quality problem is severe, the system may automatically trigger the backup water treatment apparatus to ensure that the water quality is maintained at an acceptable level. Such automated decisions can greatly reduce response time, reducing potential losses.

In the embodiment of the application, a data set of the working state of the cultivation equipment in a cultivation water area to be detected is obtained; decomposing the working state data set of the cultivation equipment through a preset operation data decomposition model to obtain a plurality of working state decomposition results; inputting a plurality of working state decomposition results into a preset extreme learning machine model to perform feature extraction to obtain a plurality of target dimension feature vectors; inputting a plurality of target dimension feature vectors into a preset abnormal detection model of the cultivation equipment to detect the abnormal operation of the cultivation equipment, and obtaining an equipment abnormal classification result; according to the method, the corresponding abnormal treatment strategy of the aquaculture equipment is established according to the abnormal classification result of the equipment, the abnormal condition of the equipment is monitored and detected in real time, production interruption can be reduced by timely finding and treating the abnormal condition, normal operation of the equipment is ensured, the growth environment of aquatic animals and plants can be improved by monitoring factors such as water quality, temperature and the like and timely adjusting the factors, and the aquaculture yield and quality are improved. By early recognition of abnormal conditions, resource waste caused by equipment faults or bad environmental conditions can be reduced, and intelligent decision support is provided. According to the abnormal conditions, suggestions and strategies can be provided for the aquaculture, and the accuracy and efficiency of the abnormal treatment of the aquaculture equipment are improved.

In a specific embodiment, the process of executing step S102 may specifically include the following steps:

(1) Classifying operation parameters of the working state data set of the cultivation equipment through a preset parameter clustering model to obtain a plurality of initial equipment operation data, wherein the plurality of initial equipment operation data comprise: temperature data, humidity data, water quality data and equipment vibration data;

(2) Respectively carrying out time sequence association processing on a plurality of initial equipment operation data to obtain a plurality of target equipment operation data;

(3) Respectively carrying out parameter initialization on a plurality of target device operation data to obtain an initialization parameter set of each target device operation data;

(4) And decomposing the operation data of the plurality of target devices according to the initialization parameter set through a preset complementary set empirical mode decomposition model to obtain a plurality of working state decomposition results of the operation data of each target device.

Specifically, the server classifies the operation parameters of the working state data set of the cultivation equipment by establishing a parameter clustering model. The objective is to divide the device operational data into a plurality of different categories, each category representing a set of related operational parameters. This may be achieved by means of a clustering algorithm, such as K-means clustering or hierarchical clustering. For example, for aquaculture, parameters such as water temperature, salinity, dissolved oxygen, etc. can be categorized. And performing time sequence association processing on each initial equipment operation data set. The data of the different parameters are correlated in time to ensure that they correspond at the same point in time. This may be achieved by time stamping or time series alignment so that subsequent analyses can be compared and unified at the same point in time. And initializing parameters of the operation data of each target device. This step involves setting an initial value for each parameter according to the specifications of the apparatus, the production environment, and the like. The parameters are initialized to ensure that the characteristics and changes of the data are better understood during the subsequent decomposition process. And decomposing the operation data of the plurality of target devices according to the initialization parameter set by a preset complementary set empirical mode decomposition model to obtain a plurality of working state decomposition results of each device. This model is typically based on mathematical and signal processing techniques, which can decompose the raw data into a plurality of modal functions, each corresponding to a different frequency or time domain characteristic. For example, assuming an aquaculture farm, the server is aimed at monitoring parameters such as water temperature, salinity, dissolved oxygen, etc. The parameters are classified by using a parameter clustering model, and time sequence association processing is carried out on the data of each parameter. Then, appropriate initial values are initialized for each parameter to ensure accuracy and interpretability of the data. And applying the initialization parameter set to the equipment operation data of each parameter by the server through a complementary set empirical mode decomposition model, so as to obtain a plurality of working state decomposition results of each equipment. These modal functions will help the server to learn more about the characteristics and trends of the device operation.

In a specific embodiment, the executing step decomposes the operation data of the plurality of target devices according to the initialization parameter set through a preset complementary set empirical mode decomposition model, and the process of obtaining the decomposition results of the plurality of working states of the operation data of each target device may specifically include the following steps:

s1: setting iteration times M and convergence threshold epsilon according to an initialization parameter set by a preset complementary set empirical mode decomposition model;

s2: calculating a mean value curve ci (t) of the operation data of each target device through a preset mean value curve calculation function, wherein the mean value curve calculation function is as follows: ci (t) =1/M Σx (t) hm (t), ci (t) represents a mean curve, M represents the number of iterations, x (t) represents target device operation data, hm (t) represents a noise function;

s3: according to the mean value curve ci (t), calculating a one-dimensional accumulated energy curve Ei (t) of the operation data of each target device, wherein the calculation function of the one-dimensional accumulated energy curve Ei (t) is as follows: ei (t) = ≡ [ x (t) -ci (t) ]2dt, ei (t) represents a one-dimensional accumulated energy curve, x (t) represents target device operation data, ci (t) represents a mean curve ci (t);

s4: judging whether the one-dimensional accumulated energy curve Ei (t) is smaller than a convergence threshold epsilon;

S5: if the one-dimensional accumulated energy curve Ei (t) is less than the convergence threshold epsilon, generating a plurality of working state decomposition results of the operation data of each target device according to the one-dimensional accumulated energy curve Ei (t);

s6: if the one-dimensional accumulated energy curve Ei (t) is not less than the convergence threshold epsilon, calculating a corresponding one-dimensional extremum curve Mi (t) according to the one-dimensional accumulated energy curve Ei (t), and generating a plurality of working state decomposition results of the operation data of each target device according to the one-dimensional accumulated energy curve Ei (t) and the one-dimensional extremum curve Mi (t), wherein the calculation function of the one-dimensional extremum curve Mi (t) is as follows: mi (t) =max (Ei (t)) -Ei (t), mi (t) representing a one-dimensional extremum curve, max (Ei (t)) representing the extremum of a one-dimensional accumulated energy curve Ei (t), ei (t) representing a one-dimensional accumulated energy curve;

s7: and iteratively executing the steps S2-S6 until the execution of the iteration times M is finished, and outputting a plurality of working state decomposition results of the operation data of each target device.

Specifically, to perform modal decomposition, the server must set two key parameters: iteration number M and convergenceA threshold epsilon. The number of iterations M represents the number of iterations in the decomposition process, and the convergence threshold epsilon is a threshold used to determine whether the decomposition is sufficiently accurate. The choice of these two parameters needs to be adjusted according to the specific problem and the nature of the data. Typically, the number of iterations M should be large enough to ensure adequate iterations, and the convergence threshold ε should be chosen based on the noise level and accuracy requirements of the data. And respectively calculating the mean value curve ci (t) of the operation data of each target device by using a preset mean value curve calculation function. The mean curve is obtained by averaging each data point over time, which helps to eliminate noise and random variations in the data. The specific form of the mean curve calculation function is as follows: ci (t) =1/M Σx (t) ×hm (t), where ci (t) represents a mean curve, M represents the number of iterations, x (t) represents target device operational data, hm (t) represents a noise function. Based on the mean curve ci (t), a one-dimensional cumulative energy curve Ei (t) for each target device operational data is calculated. The one-dimensional cumulative energy curve is used for quantifying the difference between the data and the mean curve, and is calculated as follows: ei (t) = ≡ [ x (t) -ci (t) ] ² dt, ei (t) represents a one-dimensional cumulative energy curve, x (t) represents target device operating data, and ci (t) represents a mean curve. It is determined whether the one-dimensional accumulated energy curve Ei (t) is smaller than the convergence threshold epsilon. If Ei (t) is less than ε, which means that the decomposition is sufficiently accurate, iteration may be stopped, yielding a plurality of operating state decomposition results. If Ei (t) is greater than or equal to ε, which means that the decomposition is not yet accurate enough, iteration needs to be continued to improve accuracy. There are two methods for generating the working state decomposition result according to the convergence condition: if Ei (t) < epsilon, directly generating a plurality of working state decomposition results, wherein the functions represent different modes or frequency domain components of the data; if Ei (t) is not less than epsilon, calculating a corresponding one-dimensional extremum curve Mi (t), and generating a working state decomposition result according to the one-dimensional accumulated energy curve Ei (t) and the one-dimensional extremum curve Mi (t). The one-dimensional extremum curves are used to further separate the principal modal components of the data. The steps S2 to S5 are repeatedly performed until the preset iteration number M is reached, or the one-dimensional cumulative energy curve Ei (t) is close enough to the mean curve ci (t) to satisfy the convergence condition. Through this complete process, the server obtains each target The multiple operating state decomposition results of the device operational data reflect different characteristics and frequency components of the data.

In a specific embodiment, the process of executing step S103 may specifically include the following steps:

(1) Inputting a plurality of working state decomposition results of the operation data of each target device into a preset extreme learning machine model respectively, wherein the extreme learning machine model comprises: an input layer, a plurality of hidden layers, and an output layer;

(2) Feature extraction is carried out on a plurality of working state decomposition results through an extreme learning machine model, an initial high-dimensional feature vector corresponding to each working state decomposition result is obtained, and the initial high-dimensional feature vectors are combined to obtain a combined high-dimensional feature vector of the operation data of each target device;

(3) And carrying out weight optimization analysis on the combined high-dimensional feature vector of the operation data of each target device by a preset goblet sea squirt optimization algorithm to obtain a plurality of target dimensional feature vectors.

Specifically, the server prepares a preset extreme learning machine (Extreme Learning Machine, ELM) model. ELM is a machine learning algorithm used for feature extraction and classification tasks. The ELM model includes an input layer, a plurality of hidden layers, and an output layer. Each hidden layer contains a plurality of neurons, and the number of hidden layers and the number of neurons can be set according to the complexity of the problem. And respectively inputting a plurality of working state decomposition results of the operation data of each target device into a preset ELM model. Each operating state decomposition result may be considered an input feature. The ELM model takes the decomposition result of each working state as input, and performs feature extraction in the hidden layer. This feature extraction process is achieved by random initialization of the weight parameters and the action of the activation function. And after the feature extraction, obtaining an initial high-dimensional feature vector corresponding to each working state decomposition result. And combining the initial high-dimensional feature vectors corresponding to the decomposition results of each working state to obtain combined high-dimensional feature vectors of the operation data of each target device. This may be achieved by simply concatenating the different feature vectors together or employing other combining strategies. And (3) carrying out weight optimization analysis on the combined high-dimensional feature vector of the operation data of each target equipment by using a preset sea squirt optimization algorithm. The optimization algorithm of the goblet sea squirt is an optimization algorithm based on the behaviors of the group of the goblet sea squirts in the nature and is used for searching an optimal solution. At this step, the algorithm optimizes the weights for each feature so that the combined high-dimensional feature vector better represents the features of the target device operational data. And obtaining a plurality of target dimension feature vectors through the operation of the goblet sea squirt optimization algorithm. Each vector contains important characteristics of the operation data of the corresponding target equipment, and is subjected to characteristic extraction and weight optimization. For example, assume that a server is monitoring the operation of an aquaculture device. The server collects sensor data for a plurality of devices, including information on temperature, humidity, water quality, and vibration of the devices. And decomposing the data of each device into a working state decomposition result by the server through a complementary set empirical mode decomposition model. And the server inputs the decomposition result of each working state by using a preset ELM model, and performs feature extraction to obtain an initial high-dimensional feature vector. The server combines the initial high-dimensional feature vectors for each device into a more comprehensive feature vector to better represent the operational state of the device. The server applies a goblet sea squirt optimization algorithm that optimizes the weights of each feature in the feature vector to ensure that the features are best suited for classification of the equipment operating conditions and anomaly detection. Finally, the server obtains a plurality of target dimensional feature vectors which contain important feature information of each device and have been subjected to feature extraction and weight optimization. These high-dimensional feature vectors can be used for subsequent device operational status classification and anomaly detection, helping the server to better manage and maintain the aquaculture device.

In a specific embodiment, the process of executing step S202 may specifically include the following steps:

(1) Receiving a plurality of working state decomposition results of the operation data of each target device through an input layer in the extreme learning machine model;

(2) Extracting hidden features of the working state decomposition results through a plurality of hidden layers in the extreme learning machine model to obtain initial high-dimensional feature vectors corresponding to each working state decomposition result;

(3) And carrying out vector combination on the initial high-dimensional feature vectors according to preset initial vector weights through an output layer in the extreme learning machine model to obtain combined high-dimensional feature vectors of the operation data of each target device.

Specifically, a plurality of operation state decomposition results of the operation data of each target device are respectively transmitted as inputs to an input layer of the ELM model. Each operating state decomposition result may be considered an input feature of the model. The number of hidden layers and the number of neurons in the ELM model need to be preset. Each hidden layer is responsible for carrying out feature extraction and data mapping on the input working state decomposition result. Specifically, the hidden layer maps the input working state decomposition result into the feature space of the hidden layer through the functions of the weight and the activation function, and an initial high-dimensional feature vector corresponding to each working state decomposition result is obtained. And at the output layer of the ELM model, vector combination is carried out on the initial high-dimensional feature vectors corresponding to each working state decomposition result according to the preset initial vector weight. This process combines the individual feature vectors into a higher dimensional feature vector that contains information about the outcome of each operational state decomposition. For example, assume that a server is monitoring the operation of aquaculture equipment, each having a plurality of sensors, each of which records different operational state decomposition results, such as temperature, humidity, water quality, and equipment vibration. The server has built an ELM model that includes an input layer, multiple hidden layers, and an output layer. The goal of the server is to translate these operating state decomposition results into useful high-dimensional feature vectors through the ELM model to better understand the operating state of the device. The results of the decomposition of the operating state of each device are passed to the input layer of the ELM model. For example, for device a, the server has a plurality of operating state decomposition results such as temperature, humidity, water quality, and device vibration. The hidden layers of the ELM will perform feature extraction on these modal functions separately, one for each hidden layer. In the hidden layer, the initial high-dimensional feature vector corresponding to each modal function is extracted through the functions of the weight and the activation function. These high-dimensional feature vectors contain an abstract representation of each of the modal functions. At the output layer of the ELM model, these high-dimensional feature vectors are combined into a higher-dimensional feature vector according to a preset initial vector weight, which represents the overall feature of the device a. The combined feature vector contains information of the decomposition results of each working state of the equipment A, and can be used for subsequent data analysis, classification or anomaly detection.

In a specific embodiment, the performing step performs weight optimization analysis on the combined high-dimensional feature vector of the operation data of each target device through a preset goblet sea squirt optimization algorithm, and the process of obtaining a plurality of target high-dimensional feature vectors may specifically include the following steps:

(1) Carrying out weight parameter initialization on the combined high-dimensional feature vector of the operation data of each target device through a preset goblet sea squirt optimization algorithm to obtain an initialized goblet sea squirt group, wherein the initialized goblet sea squirt group comprises a plurality of first candidate vector weights;

(2) Performing fitness calculation on the plurality of first candidate vector weights to obtain target fitness of each first candidate vector weight;

(3) According to the target fitness, performing goblet-sea squirt position updating on the initialized goblet-sea squirt group to obtain a plurality of second candidate vector weights;

(4) And determining an optimal solution from the plurality of second candidate vector weights to obtain target vector weights, and carrying out weight optimization calculation on the combined high-dimensional feature vector of the operation data of each target device according to the target vector weights to obtain a plurality of target dimensional feature vectors.

Specifically, the server initializes the weight parameters of the combined high-dimensional feature vector of the operation data of each target device through a preset sea squirt optimizing algorithm. In this process, the weight parameter of each vector is initialized to the first candidate vector weight in the group of goblet sea squirts. The goblet-sea squirt algorithm uses the idea of swarm intelligence, where each goblet-sea squirt individual has a set of weight parameters for weight optimization. For each first candidate vector weight, its fitness is calculated. Fitness is an indicator of the quality of a weighting parameter, typically measured using an objective function. In this context, the fitness function should be designed to take into account the optimization objectives of the feature vectors, such as minimizing errors or maximizing classification performance, etc. And (3) according to the calculated fitness, carrying out position updating on the initialized goblet sea squirt group. In the goblet sea squirt algorithm, the individual's location update is achieved by fine tuning of weight parameters. In particular, individuals with higher fitness will be more selected and their weighting parameters updated to better meet the optimization objectives. After a number of iterations, an optimal solution is determined from the plurality of second candidate vector weights. The optimal solution is typically the weight parameter of the individual with the highest fitness. These parameters are considered as the weight parameters that best meet the optimization objective. And using the determined weight of the target vector to perform weight optimization calculation on the combined high-dimensional feature vector of the operation data of each target device. By adjusting the weight parameters, the high-dimensional feature vector of each target device can be better expressed for subsequent data analysis, classification or anomaly detection tasks. For example, assuming that the server is monitoring the operational status of different aquaculture devices, the server has used the ELM model to convert the operational status decomposition results into high-dimensional feature vectors and combine these vectors into a combined high-dimensional feature vector. The server then further optimizes these feature vectors by the ascidian optimization algorithm to better satisfy the monitoring and classification tasks of the server. The server first initializes a group of goblet sea squirts, each individual goblet sea squirt having a set of weight parameters for the weights of the feature vectors. The server calculates the fitness of each ecteinascidity individual, and the fitness function can be designed according to the task requirement of the server. Individuals with high fitness will be selected more in the next iteration and their weight parameters will be slightly adjusted to better meet the optimization objectives of the server. After a number of iterations, the server determines the individual with the highest fitness, the weight parameter of which is considered as the parameter that best meets the optimization objectives of the server. And the server uses the optimal weight parameters to perform weight optimization calculation on the combined high-dimensional feature vector of the operation data of each target device. The optimized feature vectors can be better used for monitoring, classifying or abnormality detecting tasks of equipment states, and help a server to manage aquaculture equipment more effectively.

In a specific embodiment, the process of executing step S104 may specifically include the following steps:

(1) Inputting a plurality of target dimension feature vectors into a preset culture equipment abnormality detection model, wherein the culture equipment abnormality detection model comprises a plurality of long-time and short-time memory networks and two layers of fully-connected networks;

(2) Performing feature classification operation on the target dimension feature vectors through the long-short memory networks to obtain feature classification operation values;

(3) And inputting the multiple feature classification operation values into a two-layer fully-connected network for abnormality detection to obtain an equipment abnormality classification result.

Specifically, the server establishes a preset abnormal detection model of the cultivation equipment. This model consists of a plurality of LSTM networks and two layers of fully connected networks. LSTM networks are used to process sequence data, while fully connected networks are used for feature classification and anomaly detection. And transmitting the high-dimensional feature vectors of the plurality of target devices as input to the culture device abnormality detection model. Each target device corresponds to a high-dimensional feature vector that contains the feature data collected from the device. Each target high-dimensional feature vector is subjected to feature classification operation through a corresponding LSTM network. LSTM networks are designed to be able to process time series data, so that time dependencies and sequence patterns in the data can be captured. The output of each LSTM network will be the feature classification operand for that device. The output of the LSTM network (feature classification operand) is passed into a two-layer fully connected network. The first layer of fully-connected network is used for further feature extraction and mapping, and the second layer of fully-connected network is used for performing an anomaly detection task. This network architecture is trained to map feature classification operand values to device anomaly classification results. And inputting the plurality of target dimension feature vectors into the model through the culture equipment abnormality detection model, and obtaining an equipment abnormality classification result through processing of an LSTM network and a two-layer fully connected network. This result may indicate whether the status of the device is normal or an abnormal situation exists. For example, assume that a server monitors the operational status of a plurality of aquaculture devices, each device having a plurality of sensors, and the collected data is characterized to form a high-dimensional feature vector. The server monitors the equipment state in real time and detects the abnormality through an abnormal detection model of the cultivation equipment. The server establishes a culture equipment abnormality detection model which consists of a plurality of LSTM networks and two layers of fully-connected networks. Each LSTM network is responsible for processing the high-dimensional feature vector of one device to capture timing information in the data. The two-layer fully connected network is used to map the output of the LSTM network to the device anomaly classification result. The server inputs the high-dimensional feature vector of each device into the aquaculture device anomaly detection model. And each LSTM network performs feature classification operation on the feature vectors to obtain feature classification operation values of each device. These values reflect the current state of each device. And mapping the feature classification operation value to an abnormal classification result of the equipment through a two-layer fully connected network. For example, if the abnormality detection result is "normal" or "abnormal", the state of each device may be known. If the abnormality detection result is "abnormality", the cause of abnormality of the apparatus may be further analyzed. Through the abnormal detection model of the cultivation equipment, the server monitors a plurality of equipment in real time and detects the abnormality, helps the server to discover equipment problems in time and takes necessary measures so as to ensure the smooth proceeding of the cultivation process. This helps to improve the efficiency and reliability of aquaculture.

The method for analyzing the device abnormality based on the extreme learning machine in the embodiment of the present application is described above, and the device for analyzing the device abnormality based on the extreme learning machine in the embodiment of the present application is described below, referring to fig. 2, one embodiment of the device for analyzing the device abnormality based on the extreme learning machine in the embodiment of the present application includes:

an acquisition module 201, configured to acquire a data set of an operating state of a cultivation device in a cultivation water area to be detected;

the decomposition module 202 is configured to decompose the working state dataset of the cultivation device through a preset operation data decomposition model to obtain a plurality of working state decomposition results;

the extracting module 203 is configured to input the multiple working state decomposition results into a preset extreme learning machine model to perform feature extraction, so as to obtain multiple target dimension feature vectors;

the detection module 204 is configured to input the multiple target dimension feature vectors into a preset abnormal detection model of the cultivation equipment to perform abnormal detection on operation of the cultivation equipment, so as to obtain an abnormal classification result of the equipment;

the creating module 205 is configured to create a corresponding abnormal processing policy of the farm equipment according to the abnormal classification result of the equipment.

Acquiring a data set of the working state of the cultivation equipment in the cultivation water area to be detected through the cooperation of the components; decomposing the working state data set of the cultivation equipment through a preset operation data decomposition model to obtain a plurality of working state decomposition results; inputting the multiple working state decomposition results into a preset extreme learning machine model to perform feature extraction to obtain multiple target dimension feature vectors; inputting the target dimension feature vectors into a preset abnormal detection model of the cultivation equipment to detect the abnormal operation of the cultivation equipment, and obtaining an equipment abnormal classification result; according to the abnormal classification result of the equipment, a corresponding abnormal treatment strategy of the aquaculture equipment is created, the abnormal condition of the equipment is monitored and detected in real time, production interruption can be reduced by timely finding and treating the abnormal condition, normal operation of the equipment is ensured, the growth environment of aquatic animals and plants can be improved by timely adjusting the factors such as water quality and temperature, and the aquaculture yield and quality are improved. By early recognition of abnormal conditions, resource waste caused by equipment faults or bad environmental conditions can be reduced, and intelligent decision support is provided. According to the abnormal conditions, suggestions and strategies can be provided for the aquaculture, and the accuracy and efficiency of the abnormal treatment of the aquaculture equipment are improved.

It will be clear to those skilled in the art that, for convenience and brevity of description, specific working procedures of the above-described systems, apparatuses and units may refer to corresponding procedures in the foregoing method embodiments, which are not repeated herein.

The integrated units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present application may be embodied in essence or a part contributing to the prior art or all or part of the technical solution in the form of a software product stored in a storage medium, including several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the embodiments of the present application. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a read-only memory (ROM), a random access memory (random access memory, RAM), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

The above embodiments are merely for illustrating the technical solution of the present application, and not for limiting the same; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the corresponding technical solutions.

Claims

1. The equipment abnormality analysis method based on the extreme learning machine is characterized by comprising the following steps of:

decomposing the working state data set of the cultivation equipment through a preset operation data decomposition model to obtain a plurality of working state decomposition results; the method specifically comprises the following steps: classifying the operation parameters of the working state data set of the cultivation equipment through a preset parameter clustering model to obtain multiple operation parametersA plurality of initial device operational data, wherein the plurality of initial device operational data comprises: temperature data, humidity data, water quality data and equipment vibration data; respectively carrying out time sequence association processing on the plurality of initial equipment operation data to obtain a plurality of target equipment operation data; respectively carrying out parameter initialization on the plurality of target equipment operation data to obtain an initialization parameter set of each target equipment operation data; decomposing the operation data of the plurality of target devices according to the initialization parameter set through a preset complementary set empirical mode decomposition model to obtain a plurality of working state decomposition results of the operation data of each target device; the method for obtaining the multiple working state decomposition results of the operation data of each target device comprises the following steps: s1: setting iteration times M and convergence threshold epsilon according to the initialization parameter set by a preset complementary set empirical mode decomposition model; s2: calculating a mean curve ci (t) of the operation data of each target device through a preset mean curve calculation function, wherein the mean curve calculation function is as follows: ci (t) = (1/M) Σx (t) ×hm (t), ci (t) represents a mean curve, M represents the number of iterations, x (t) represents target device operation data, hm (t) represents a noise function; s3: according to the mean value curve ci (t), calculating a one-dimensional accumulated energy curve Ei (t) of the operation data of each target device, wherein a calculation function of the one-dimensional accumulated energy curve Ei (t) is as follows: ei (t) = ≡ [ x (t) -ci (t) ] ² dt, ei (t) represents a one-dimensional accumulated energy curve, x (t) represents target device operation data; s4: judging whether the one-dimensional accumulated energy curve Ei (t) is smaller than the convergence threshold epsilon; s5: if the one-dimensional accumulated energy curve Ei (t) is less than the convergence threshold epsilon, generating a plurality of working state decomposition results of the operation data of each target device according to the one-dimensional accumulated energy curve Ei (t); s6: if the one-dimensional accumulated energy curve Ei (t) is not less than the convergence threshold epsilon, calculating a corresponding one-dimensional extremum curve Mi (t) according to the one-dimensional accumulated energy curve Ei (t), and generating a plurality of working state decomposition results of the operation data of each target device according to the one-dimensional accumulated energy curve Ei (t) and the one-dimensional extremum curve Mi (t), wherein the calculation function of the one-dimensional extremum curve Mi (t) is that: mi (t) =max (Ei (t)) -Ei (t), mi (t) representing a one-dimensional extremum curve, max (Ei (t)) representing the extremum of a one-dimensional accumulated energy curve Ei (t), ei (t) representing a one-dimensional accumulated energy curve; s7: iteratively executing the steps S2-S6 until the execution of the iteration times M is finished, and outputting a plurality of working state decomposition results of the operation data of each target device;

inputting the multiple working state decomposition results into a preset extreme learning machine model to perform feature extraction to obtain multiple target dimension feature vectors; the method specifically comprises the following steps: inputting a plurality of working state decomposition results of the operation data of each target device into a preset extreme learning machine model respectively, wherein the extreme learning machine model comprises: an input layer, a plurality of hidden layers, and an output layer; extracting features of the working state decomposition results through the extreme learning machine model to obtain initial high-dimensional feature vectors corresponding to each working state decomposition result, and combining the initial high-dimensional feature vectors to obtain combined high-dimensional feature vectors of the operation data of each target device; carrying out weight optimization analysis on the combined high-dimensional feature vector of the operation data of each target device by a preset goblet sea squirt optimization algorithm to obtain a plurality of target dimensional feature vectors;

2. The method for analyzing equipment abnormality based on an extreme learning machine according to claim 1, wherein the feature extraction is performed on the plurality of working state decomposition results by the extreme learning machine model to obtain an initial high-dimensional feature vector corresponding to each working state decomposition result, and the initial high-dimensional feature vectors are combined to obtain a combined high-dimensional feature vector of each target equipment operation data, and the method comprises the steps of:

receiving a plurality of working state decomposition results of the operation data of each target device through an input layer in the extreme learning machine model;

extracting hidden features of the working state decomposition results through a plurality of hidden layers in the extreme learning machine model to obtain initial high-dimensional feature vectors corresponding to each working state decomposition result;

and carrying out vector combination on the initial high-dimensional feature vector according to preset initial vector weights through an output layer in the extreme learning machine model to obtain a combined high-dimensional feature vector of the operation data of each target device.

3. The method for analyzing equipment abnormality based on an extreme learning machine according to claim 2, wherein the weight optimization analysis is performed on the combined high-dimensional feature vector of the operation data of each target equipment by a preset goblet sea squirt optimization algorithm to obtain a plurality of target dimensional feature vectors, and the method comprises the following steps:

carrying out weight parameter initialization on the combined high-dimensional feature vector of the operation data of each target device through a preset goblet sea squirt optimization algorithm to obtain an initialized goblet sea squirt group, wherein the initialized goblet sea squirt group comprises a plurality of first candidate vector weights;

performing fitness calculation on the plurality of first candidate vector weights to obtain target fitness of each first candidate vector weight;

according to the target fitness, performing goblet-sea squirt position updating on the initialized goblet-sea squirt population to obtain a plurality of second candidate vector weights;

and determining an optimal solution from the plurality of second candidate vector weights to obtain a target vector weight, and carrying out weight optimization calculation on the combined high-dimensional feature vector of the operation data of each target device according to the target vector weight to obtain a plurality of target dimensional feature vectors.

4. The method for analyzing equipment abnormality based on an extreme learning machine according to claim 1, wherein the step of inputting the plurality of target dimensional feature vectors into a preset equipment abnormality detection model for detecting the operation abnormality of the equipment to obtain an equipment abnormality classification result comprises the steps of:

Inputting the target dimension feature vectors into a preset culture equipment abnormality detection model, wherein the culture equipment abnormality detection model comprises a plurality of long-time and short-time memory networks and two layers of fully-connected networks;

performing feature classification operation on the target dimension feature vectors through the long-short time memory networks to obtain feature classification operation values;

and inputting the characteristic classification operation values into the two-layer fully-connected network to perform abnormality detection, so as to obtain an equipment abnormality classification result.

5. An apparatus abnormality analysis device based on an extreme learning machine, characterized in that the apparatus abnormality analysis device based on an extreme learning machine includes:

the decomposition module is used for decomposing the working state data set of the cultivation equipment through a preset operation data decomposition model to obtain a plurality of working state decomposition results; the method specifically comprises the following steps: classifying the operation parameters of the working state data set of the cultivation equipment through a preset parameter clustering model to obtain a plurality of initial equipment operation data, wherein the plurality of initial equipment operation data comprise: temperature data, humidity data, water quality data and equipment vibration data; respectively carrying out time sequence association processing on the plurality of initial equipment operation data to obtain a plurality of target equipment operation data; respectively carrying out parameter initialization on the plurality of target equipment operation data to obtain an initialization parameter set of each target equipment operation data; decomposing the operation data of the plurality of target devices according to the initialization parameter set through a preset complementary set empirical mode decomposition model to obtain a plurality of working state decomposition results of the operation data of each target device; the method for obtaining the multiple working state decomposition results of the operation data of each target device comprises the following steps: s1: through a preset complementary set empirical mode decomposition model Setting iteration times M and convergence threshold epsilon according to the initialization parameter set; s2: calculating a mean curve ci (t) of the operation data of each target device through a preset mean curve calculation function, wherein the mean curve calculation function is as follows: ci (t) = (1/M) Σx (t) ×hm (t), ci (t) represents a mean curve, M represents the number of iterations, x (t) represents target device operation data, hm (t) represents a noise function; s3: according to the mean value curve ci (t), calculating a one-dimensional accumulated energy curve Ei (t) of the operation data of each target device, wherein a calculation function of the one-dimensional accumulated energy curve Ei (t) is as follows: ei (t) = ≡ [ x (t) -ci (t)] ² dt, ei (t) represents a one-dimensional accumulated energy curve, x (t) represents target device operation data; s4: judging whether the one-dimensional accumulated energy curve Ei (t) is smaller than the convergence threshold epsilon; s5: if the one-dimensional accumulated energy curve Ei (t) is less than the convergence threshold epsilon, generating a plurality of working state decomposition results of the operation data of each target device according to the one-dimensional accumulated energy curve Ei (t); s6: if the one-dimensional accumulated energy curve Ei (t) is not less than the convergence threshold epsilon, calculating a corresponding one-dimensional extremum curve Mi (t) according to the one-dimensional accumulated energy curve Ei (t), and generating a plurality of working state decomposition results of the operation data of each target device according to the one-dimensional accumulated energy curve Ei (t) and the one-dimensional extremum curve Mi (t), wherein a calculation function of the one-dimensional extremum curve Mi (t) is as follows: mi (t) =max (Ei (t)) -Ei (t), mi (t) representing a one-dimensional extremum curve, max (Ei (t)) representing the extremum of a one-dimensional accumulated energy curve Ei (t), ei (t) representing a one-dimensional accumulated energy curve; s7: iteratively executing the steps S2-S6 until the execution of the iteration times M is finished, and outputting a plurality of working state decomposition results of the operation data of each target device;

The extraction module is used for respectively inputting the multiple working state decomposition results into a preset extreme learning machine model to perform feature extraction to obtain multiple target dimension feature vectors; the method specifically comprises the following steps: inputting a plurality of working state decomposition results of the operation data of each target device into a preset extreme learning machine model respectively, wherein the extreme learning machine model comprises: an input layer, a plurality of hidden layers, and an output layer; extracting features of the working state decomposition results through the extreme learning machine model to obtain initial high-dimensional feature vectors corresponding to each working state decomposition result, and combining the initial high-dimensional feature vectors to obtain combined high-dimensional feature vectors of the operation data of each target device; carrying out weight optimization analysis on the combined high-dimensional feature vector of the operation data of each target device by a preset goblet sea squirt optimization algorithm to obtain a plurality of target dimensional feature vectors;