Visual system for fault diagnosis and health assessment of marine air conditioning system

Abstract

The application provides a kind of ship air conditioning system fault diagnosis and health evaluation visualization system, equipment edge side obtains the real-time data stream of original measurement variable corresponding to the measuring point of each subsystem of ship air conditioning system;Complete algorithm data set is displayed on the interface of equipment edge side;JSON message on the upper system side, get summary time series library;Through the interface rendering program on the upper system side, the summary time series library is displayed on the interface of the upper system side;The summary time series library is packaged into UDP broadcast data packet stream;By the UDP to HTTP data conversion processing module of the control overall side, get global structured operation and maintenance database;Through the interface rendering program on the control overall side, the global structured operation and maintenance database is displayed on the interface of the control overall side.The application integrates fault diagnosis and health evaluation functions into three-layer visualization architecture, and realizes accurate identification of faults and effective prediction of equipment health status through advanced diagnosis and evaluation methods.

Description

Technical Field

[0001] This invention belongs to the field of fault identification of marine air conditioning systems, and specifically relates to a visualization system for fault diagnosis and health assessment of marine air conditioning systems. Background Technology

[0002] With increasing demands for ship automation and passenger / cargo comfort, ship air conditioning systems are becoming increasingly larger, with more diverse equipment types and measurement points. Equipment such as centralized air conditioning units, duct loops, chillers, chilled water pumps, and cooling seawater pumps operate in tandem within the same system, often accompanied by hundreds of measurement variables (temperature, humidity, flow rate, pressure, electrical quantities, vibration, etc.). Long-term practical engineering and maintenance work has revealed inaccuracies in traditional ship air conditioning monitoring and maintenance, severely restricting system reliability and operational efficiency. Summary of the Invention

[0003] The purpose of this invention is to provide a visualization system for fault diagnosis and health assessment of marine air conditioning systems.

[0004] To address the above problems, this invention provides a visualization system for fault diagnosis and health assessment of marine air conditioning systems, comprising: On the edge of the equipment, it is used to obtain the real-time data stream of the original measurement variables corresponding to the measurement points of each subsystem of the ship's air conditioning system and the corresponding actuator control parameters issued by the ship's air conditioning automatic control system; At the device edge, a complete algorithm dataset is obtained based on the real-time data stream of the original measurement variables; the complete algorithm dataset is then displayed on the device edge interface through the interface rendering program at the device edge. At the device edge, based on the complete algorithm dataset, standardized JSON messages are generated and uploaded to the upper-level system. On the upper-level system side, it is used to upload JSON messages based on the device edge side to obtain the aggregated timing library; the aggregated timing library is displayed on the interface of the upper-level system side through the interface rendering program on the upper-level system side. On the upper-level system side, the data in the aggregated time-series database is divided into two categories and encapsulated differently: For the first category of data, namely real-time KPI values, it is directly encapsulated into the first UDP broadcast message to meet the high-frequency refresh requirements of the large screen on the centralized control side; For the second category of data, Chinese text such as fault diagnosis, health decision, and maintenance work orders, a preset text-to-number mapping table is matched to convert the Chinese text into the corresponding numerical encoding, and then encapsulated into the second UDP broadcast message to avoid the defect of Chinese character garbled characters in native UDP transmission; Based on the first and second UDP broadcast messages after classification and differential encoding, a standardized UDP single data packet is generated; The UDP broadcast push service on the upper-level system side pushes the UDP broadcast data packet stream containing the standardized UDP single data packet to the global broadcast address at a fixed period; The centralized control system receives UDP broadcast data packets. The UDP-to-HTTP data conversion and processing module on the centralized control system performs end-to-end fault tolerance processing on the UDP broadcast data packets to obtain a global structured operation and maintenance database. The global structured operation and maintenance database is then displayed on the interface of the centralized control system through the interface rendering program on the centralized control system.

[0005] Compared with the prior art, the advantages of the present invention are: (1) Three-layer progressive visualization architecture: Through the layered design of the first layer of visualization (device edge side), the second layer of visualization (upper system side) and the third layer of visualization (central control overall side), a seamless connection from local device detail monitoring to overall system macro control is realized, solving the problem of information dispersion and difficulty in collaborative management in traditional monitoring methods.

[0006] (2) Multi-dimensional monitoring interface system: For different levels and different subsystems / grandchild systems, a reasonable number of monitoring interfaces with comprehensive functions have been designed, covering all aspects from basic operating parameters to fault diagnosis and health assessment, meeting the needs of different users and management scenarios.

[0007] (3) Hybrid data transmission and storage strategy: Based on the characteristics and transmission requirements of data at different levels, a variety of data transmission methods such as HTTP-POST and UDP broadcast are adopted, as well as database storage methods such as PostgreSQL + Prisma, to ensure the real-time performance, reliability and efficiency of data transmission and storage.

[0008] (4) Integrated fault diagnosis and health assessment mechanism: The fault diagnosis and health assessment functions are integrated into the three-layer visualization architecture. Through advanced algorithms and formulas, the accurate identification of faults and the effective prediction of equipment health status are realized, realizing the transformation from passive maintenance to proactive maintenance. Attached Figure Description

[0009] Figure 1 shows a three-layer visualized overall architecture for fault diagnosis and health assessment of a ship air conditioning system according to an embodiment of the present invention; Figure 2 is a schematic diagram showing the division of the ship air conditioning system subsystem and its subsystems according to an embodiment of the present invention; Figure 3 is a flowchart of the UDP broadcast reception and UDP-HTTP data conversion processing of an embodiment of the present invention; Figure 4 is a schematic diagram of the device edge-side core computing program module according to an embodiment of the present invention; Figure 5 is a flowchart of the training and inference process of an improved BP neural network according to an embodiment of the present invention; Figure 6 This is a network structure diagram of a multilayer feedforward BP neural network according to an embodiment of the present invention; Figure 7 This is a flowchart of a Kalman filter state prediction process according to an embodiment of the present invention; Figure 8 This is a diagram illustrating the fuzzy hierarchical analysis health assessment decision-making process and weight calculation according to an embodiment of the present invention. Figure 9 This is a timing and processing link diagram of system communication and data flow according to an embodiment of the present invention; Figure 10 This is a schematic diagram of the software interface of the first industrial control computer (air conditioning unit and air system) on the device edge side (first layer of visualization) according to an embodiment of the present invention; Figure 11 This is a schematic diagram of the software interface of the first industrial control computer (air conditioning unit and air system) on the device edge side (first layer of visualization) according to an embodiment of the present invention; Figure 12 This is a schematic diagram of the upper-level system side (second-layer visualization) software interface according to an embodiment of the present invention; Figure 13 This is a schematic diagram of the centralized control overall side (third-layer visualization) software interface according to an embodiment of the present invention. Detailed Implementation

[0010] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0011] like Figure 1 As shown, the present invention provides a visualization system for fault diagnosis and health assessment of a ship air conditioning system, comprising: Step S1, Data Acquisition and Local Persistence of Visualization Layer at the Equipment Edge: The industrial control all-in-one machine at the equipment edge acquires the real-time data stream of the original measurement variables corresponding to the measurement points of each subsystem of the ship's air conditioning system and the corresponding actuator control parameters issued by the ship's air conditioning automatic control system.

[0012] Here, in one embodiment, referring to Figure 1 and Figure 2 As shown, the three-layer visualized overall architecture of the marine air conditioning system includes: (1) Equipment edge side (first layer of visualization), including two industrial control all-in-one computers. The first industrial control all-in-one computer uses the core algorithm to calculate and display the air conditioning unit and the air system; the second industrial control all-in-one computer uses the core algorithm to calculate and display the chiller unit and the water system.

[0013] Specifically, the equipment on the edge side of the first layer of visualization consists of two 11-inch edge-side industrial control all-in-one machines. The first machine corresponds to the first subsystem (centralized air conditioning unit) and the second subsystem (air system); the second machine corresponds to the third subsystem (chiller unit) and the fourth subsystem (water system).

[0014] The first layer of visualization includes the monitoring interface at the device edge, which includes: The first industrial control computer displays detailed information about the first subsystem (8 centralized air conditioning units) and the second subsystem (the air systems corresponding to the 8 centralized air conditioning units). Each air conditioning unit and air system has 9 monitoring interfaces: main interface, performance monitoring, steady-state operation judgment, sensor fault, equipment fault alarm, equipment fault early warning, health assessment decision, historical data query, and health assessment report. Therefore, the first industrial control computer has a total of 9 × (8 + 1) = 81 monitoring interfaces.

[0015] The second industrial control computer displays detailed information about the third subsystem (2 chiller units) and the fourth subsystem (water system, including 2 chilled water pumps, 2 cooling seawater pumps, and corresponding cabin water circulation systems, totaling 3 sub-systems). Each sub-system has 9 monitoring interfaces: main interface, operational performance monitoring, steady-state operation judgment, sensor fault, equipment fault alarm, equipment fault early warning, health assessment decision, historical data query, and health assessment report. Therefore, the second industrial control computer has a total of 9 × (2 + 1 + 1 + 1) = 45 monitoring interfaces.

[0016] The first layer of visualization involves data processing at the device edge: the database storage uses PostgreSQL, the data transmission uses HTTP-POST, and the processed data is uploaded to the second layer of visualization.

[0017] The first layer of visualization (equipment edge): As the lowest level of visualization, it is directly connected to the equipment and is responsible for real-time and detailed monitoring of specific equipment in each subsystem. It can collect various operating parameters of the equipment, perform steady-state operation assessment, fault symptom detection and diagnosis, and generate information such as maintenance decisions, historical data queries, and health report printing. Its key function is to monitor and quickly respond to localized equipment problems in real time, providing direct evidence for on-site operation and maintenance.

[0018] The database storage uses PostgreSQL and the Prisma Object Relational Mapping (ORM) management framework; the device edge (first-layer visualization) acts as the server and uses Node.js. Data transmission uses the Hypertext Transfer Protocol (HTTP) – POST requests are uploaded from the first layer to the second-layer visualization. HTTP is the foundational protocol of the Web, used for transferring data between clients (such as browsers) and servers; POST requests are used to send data to the server.

[0019] The equipment edge side (first-level visualization) also has a "core computing program" installed, which includes: data cleaning, expert rule-based detection of operational performance fault symptoms, fault diagnosis based on an improved backpropagation neural network, a Kalman filter-based equipment state prediction algorithm, and a health assessment decision based on fuzzy hierarchical analysis (FAHP). The input data for the "core computing program" is the system operational performance parameters collected on-site, and the output results are directly saved to a PostgreSQL database. The equipment edge side (first-level visualization) directly reads the results from this database and displays them visually.

[0020] (2) Upper system side (second layer visualization), including an industrial control all-in-one machine.

[0021] The second layer of the visualized upper-level system consists of a 15-inch industrial control all-in-one machine and the Galaxy Kylin operating system.

[0022] Specifically, such as Figure 2 As shown, the monitoring content of the second-level visualized upper-level system is as follows: it displays the main operating data of the entire air conditioning system, summarizing 4 subsystems and 14 subordinate subsystems, including 8 centralized air conditioning units, air system, 2 chiller units, 2 chilled water pumps, 2 cooling seawater pumps, and cabin water circulation system.

[0023] The second layer is the monitoring interface on the upper-level visualization system side: each subsystem has 7 monitoring interfaces, namely the main interface, device information, operating performance, fault detection, fault diagnosis, health assessment, and report printing. Therefore, there are a total of 14 × 7 = 98 monitoring interfaces.

[0024] The second layer of visualization involves data processing on the upper-level system side: database storage uses PostgreSQL and the Prisma Object Relational Mapping (ORM) management framework. Data transmission uses User Datagram Protocol (UDP) broadcast to upload the aggregated and processed data to the third layer of visualization.

[0025] Here, the second layer of visualization (upper-level system side) summarizes and processes information from each subsystem, displaying the main operational data of the entire large air conditioning system. It enables centralized monitoring and management of equipment information, operational performance, fault conditions, and health status of each subsystem, achieving information exchange and collaborative management between subsystems. Its key function is to integrate subsystem information, providing system administrators with a comprehensive view of the system's operational status, facilitating system-level fault diagnosis and the formulation of corresponding health assessment decisions.

[0026] (3) Centralized control overall side (third layer visualization), including a central control room display screen.

[0027] Specifically, the equipment on the third-layer visualized centralized control side consists of a large screen in the central control room.

[0028] The third layer of visualization involves data transmission and processing on the overall control side: key performance parameters (numerical variables) of each subsystem are transmitted and received via UDP broadcast; symptom detection, fault diagnosis, and health assessment decision results (Chinese strings converted to numerical encoding) of each subsystem are transmitted via UDP broadcast, and after reception, UDP-to-HTTP data processing (numerical encoding converted to Chinese strings) is performed.

[0029] The third layer is the centralized control system's monitoring interface: each subsystem has four monitoring interfaces, namely, operational performance, fault detection, fault diagnosis, and health assessment. Therefore, there are a total of 14 × 4 = 56 monitoring interfaces.

[0030] Here, the third layer of visualization (centralized control system): This involves using a large screen in the central control room to achieve macro-level monitoring of the entire air conditioning system, displaying key performance parameters, fault diagnosis results, and health assessment decisions. It provides control personnel with a global perspective, enabling them to quickly grasp the overall system operation and make timely decisions and commands. Its key function is overall control, ensuring the coordinated operation of the entire air conditioning system with other ship systems, and guaranteeing the safe navigation of the vessel.

[0031] The overall control side (third-layer visualization) is displayed on the large screen in the control room. The key performance parameters of each subsystem are transmitted and received using UDP broadcast. The fault diagnosis results and health assessment decisions of each subsystem are transmitted using UDP broadcast. After receiving the data uploaded by the second-layer visualization, UDP is converted to HTTP data for processing.

[0032] In one embodiment, referring to Figure 2, the division of the air conditioning system subsystem and its subsystems is as follows: (1) First subsystem: air conditioning unit, with a total of 8 subsystems: each centralized air conditioning unit corresponds to one subsystem, numbered AHU1 to AHU8. Each unit includes measurement points such as supply air temperature and humidity, fan three-phase current and voltage, operating frequency, filter pressure difference, and coil water temperature, totaling 168 measurement variables.

[0033] (2) Second subsystem: air system, with 1 subsystem: all the air ducts of the 8 centralized air conditioning units are defined as 1 subsystem. Total air volume, duct pressure difference, terminal static pressure, return air humidity, a total of 22 measurement variables.

[0034] (3) Third subsystem: Chiller unit. There are two subsystems: each chiller unit corresponds to one subsystem, numbered CHL1 and CHL2. Evaporation / condensation temperature, refrigerant flow rate, compressor current, vibration acceleration, etc., totaling 64 measured variables.

[0035] (4) Fourth subsystem: Water system, consisting of 3 subsystems: 2 chilled water pumps, 2 cooling seawater pumps, and a cabin water circulation system, each corresponding to 1 subsystem. A total of 40 measurement variables were measured, including pump inlet and outlet pressure, vibration, seawater temperature difference, and pipeline pressure at constant pressure points.

[0036] A total of 168 + 22 + 64 + 40 = 294 measurement points are used, corresponding to 294 measurement variables for this air conditioning system, including temperature, humidity, flow rate, pressure or differential pressure, current, voltage, frequency, vibration velocity, vibration acceleration, and equipment operating status. All measurement points are bound to their corresponding subsystem IDs, with an adjustable acquisition period of 1~10s, and a data acquisition configuration file is generated as the input for the data acquisition program.

[0037] Specifically, sensor measurement points are bound according to the preset rules of 4 subsystems and 14 grandchild systems: 168 measurement points in the first system (8 AHUs, 8 grandchild systems), 22 measurement points in the second system (1 grandchild system of the air system), 64 measurement points in the third system (2 grandchild systems of the chiller units), and 40 measurement points in the fourth subsystem (3 grandchild systems of the water pump and water circulation system), totaling 294 measurement variables. Each measurement point is bound to its respective grandchild system ID and sampling period (configurable from 1 to 10 seconds). The resulting data acquisition configuration file with 294 measurement variables bound to the grandchild system identifiers is directly used as the input to the program for data acquisition rules.

[0038] The industrial control all-in-one machine on the edge of the equipment periodically polls the PLC Modbus / TCP messages according to the configuration file, extracts the raw values ​​one by one, and adds UTC timestamps, subsystem IDs, and CRC check codes. The real-time data stream of the raw measurement variables is distributed in two ways: one way is written to the edge PostgreSQL temporary time series table for persistent cache, serving as the data source for historical queries in the visualization interface and for data entry after network outages; the other way is directly sent to the edge integrated intelligent computing module in step S2 as raw input to generate a standardized raw feature input stream.

[0039] Step S2: Based on the real-time data stream of the original measurement variables, the device edge obtains a complete algorithm calculation dataset, including: real-time data stream of the original measurement variables, fault labels, fault probabilities, trend predictions, NIS identifiers, health scores, maintenance work orders, and end-to-end intermediate features; like Figure 4 As shown, step S2 includes: Step S2-1: The data cleaning and normalization module at the device edge performs data cleaning and normalization on the real-time data stream of the original measurement variables obtained in step S1, obtains a real-time standardized feature vector, and outputs it to the expert rule fault symptom detection module at the device edge. Data cleaning: The original real-time data stream of the measurement variables is processed for missing values ​​(interpolation / previous value preservation), outlier filtering (based on sliding window statistics and steady-state discrimination), and noise reduction (low-pass filtering) to obtain the cleaned real-time data stream; Normalization processing: Zero mean variance normalization (or min-max normalization) is performed on each feature in the cleaned real-time data stream to obtain a real-time standardized feature vector, which is then output to the expert rule fault symptom detection module at the device edge. The same set of normalization parameters is used for subsequent training and inference to ensure consistency.

[0040] Here, the method for acquiring the real-time data stream of the original measurement variables is as follows: the original measurement variables (temperature, humidity, flow rate, differential pressure, current, voltage, frequency, vibration, etc.) are read from the Modbus / TCP protocol of the PLC / sensor interface at fixed time intervals.

[0041] The data cleaning and normalization module performs forward interpolation of missing values, 3σ sliding window anomaly removal, first-order low-pass denoising, and zero-mean or minimum-maximum normalization on the real-time data stream of the original measurement variables obtained in step S1 to obtain real-time standardized feature vectors, which are then output to the fault symptom detection module. Step S2-2, Expert Rule Fault Symptom Detection Module: Real-time standardized feature vector, obtain the corresponding real-time standardized feature vector and the corresponding symptom list, and output to the BP neural network module; Step S2-2-1: Retrieve the built-in expert threshold rule library (combined timing and threshold condition rules for refrigerant temperature difference exceeding tolerance, compressor current overload, vibration amplitude exceeding standard, sensor drift judgment, etc.). Step S2-2-2: Match the real-time standardized feature vector output from step S2-1 against the expert threshold rule base to identify faults such as abnormal temperature, backflow, electrical overload, and sensor drift in real time, and obtain the fault type; calculate the fault confidence value in the 0~1 interval corresponding to each fault type to generate a symptom list, which includes: fault type, fault confidence value, and fault trigger time, as prior information for subsequent neural network and alarm management; output the real-time standardized feature vector corresponding to the real-time data stream of the original measurement variable and the corresponding symptom list to the BP neural network module.

[0042] Here, the detection rules are based on a rule base (thresholds, combination conditions, and temporal relationships) developed by experts and equipment specifications, and trigger symptoms in real time on the cleaned data (such as abnormal temperature difference, backflow of flow, overload current, etc.).

[0043] The test results include a list of symptoms (fault confidence values, fault trigger times) and serve as preliminary suggestions for improving BP networks and alarm management.

[0044] Fault type: Used to identify specific faults such as abnormal temperature difference, backflow of flow, electrical overload, and sensor drift. The source is the matching result of the expert threshold rule library built into this module. Fault confidence score: A value in the range of 0 to 1, representing the credibility of the fault symptom, derived from standardized feature vector calculations; the standardized feature vector can be matched against an expert threshold rule base one by one, and the fault confidence score of a single rule can be calculated using a piecewise linear interpolation method: using the fault normal threshold and alarm threshold as interpolation boundaries, the degree of deviation of the feature from the normal range is linearly normalized to obtain a single rule fault confidence value in the range of [0,1]; if a type of fault is jointly determined by multiple feature rules, a weighted fusion method is used to summarize the confidence scores of multiple rules to obtain the final comprehensive confidence score of the type of fault; a structured fault symptom list is generated based on the fault type, comprehensive confidence score, and data sampling time.

[0045] Fault trigger time: The timestamp when the fault matching rule is hit, which comes from the time sequence marker attached when collecting the original measurement variables in step S1.

[0046] like Figure 5 As shown, steps S2-3 include: a BP neural network module, used for: Step S2-3-1 yields the steady-state subset, transient load subset, and high / low temperature marine environment subset, along with their corresponding manually labeled fault tags: Step S2-3-1-1: Obtain the historical offline dataset. The historical offline dataset includes: 294 archived data of measurement variables of each subsystem of the ship's air conditioning system throughout the entire cycle, with manually labeled fault and operating condition labels. Here, the input for step S2-1 is the real-time dynamic measurement point stream, i.e., the real-time data stream of the original measurement variables. The input for this step is the archived historical stock samples, i.e., the historical stock offline dataset. The data sources are completely different, and only the mathematical formulas are cleaned and normalized for reuse.

[0047] The full cycle includes: the full cycle of the time sequence (time dimension) and the full system cycle of the device.

[0048] Manually labeled faults include: fault type, fault true confidence value, fault start and end timestamps, and the subsystem number to which the fault belongs.

[0049] Operating condition labels, including steady-state operation, transient variable load, and high and low temperature marine conditions, are used for stratified screening of training samples.

[0050] The time series period (time dimension) refers to continuously archived data covering the entire navigation cycle of a ship, including four types of complete runtime segments: 1) Stable operating period at full load in nearshore / deep-sea areas; 2) Variable load start-up and shutdown, and transient operating conditions during temperature rise and fall; 3) Operating conditions in high and low temperature sea areas and with alternating day and night temperature differences; 4) Long-term operating conditions with minor equipment malfunctions and slow performance degradation; Data storage duration meets ship operation and maintenance audit standards, with each archived dataset containing at least 30 days of continuous data collection, fully reproducing the air conditioning's changes from steady-state startup, variable operating conditions, performance degradation to shutdown, covering both short-term fluctuations and medium- to long-term performance degradation patterns. The time dimension covers all continuous operating conditions of the ship under full load, variable load, high and low temperatures, and performance degradation, with a storage duration of at least 30 days.

[0051] The entire system lifecycle (equipment dimension) refers to the complete coverage of all 294 measurement variables across the 4 subsystems and 14 grandchild systems defined in this invention, with no missing measurement points. It includes all measurement points for temperature, humidity, flow rate, pressure, electrical parameters, vibration, and operating status of 8 air conditioning units, air loops, 2 chillers, chilled / cooling water pumps, and cabin water circulation, without any incomplete datasets due to partial equipment or measurement point truncation. It fully covers four categories of equipment health conditions: normal, sub-healthy, warning, and faulty. The equipment dimension fully covers all measurement points across the 4 subsystems and 14 grandchild systems without any missing or truncated points. The dataset includes manually labeled fault and health condition tags, serving as dedicated samples for offline training of the improved BP neural network.

[0052] Step S2-3-1-2: Traverse all measurement point samples in the real-time historical offline dataset, calculate the global mean and global standard deviation for each feature dimension, encapsulate the global mean and global standard deviation of all features and bind them to the Z-score normalization operation logic to generate a reusable global normalization operator; based on the global normalization operator, perform the same data cleaning and normalization processing as in step S2-1 on the historical offline dataset to obtain the corresponding historical normalized feature vector; Step S2-3-1-3: Based on the equipment steady-state operating condition label or equipment transient operating condition label in the operating condition label, the historical standardized feature vector is split into three types of training samples: steady-state subset, transient variable load subset, and high and low temperature marine environment subset. Step S2-3-1-4: Bind and pair the steady-state subset, transient load subset, and high and low temperature marine environment subset with the corresponding manually labeled fault tags one by one to obtain the steady-state subset, transient load subset, and high and low temperature marine environment subset and the corresponding manually labeled fault tags. like Figure 6 As shown, in step S2-3-2, network structure and parameter initialization: the dimension values ​​of the historical standardized feature vectors are read as the number of input layer nodes to build a multi-layer feedforward neural network with 2 hidden layers; ReLU is selected as the activation function for the hidden layers, and Softmax is used for the output layer to achieve multi-fault classification; the Xavier method is used in combination with the input dimension to complete the network weight initialization, and the adaptive learning rate is preset. Preset adaptive momentum factor The model incorporates L2 regularization coefficients, an improved weight update formula based on adaptive momentum, and a cross-entropy loss function to obtain an initial neural network model with completed structure building and parameter initialization (including: an initial weight matrix matching the input dimension, activation functions for each layer, loss functions, and an improved weight update formula based on adaptive momentum).

[0053] Here, the adaptive momentum improvement weight update calculation formula is the Adam optimizer iterative update formula. The algorithm integrates the first-order momentum accumulation mechanism and the second-order gradient adaptive scaling mechanism: the first-order momentum coefficient adaptively smooths gradient fluctuations, and the second-order gradient accumulation dynamically matches the effective learning rate for each layer parameter; the network training hyperparameter configuration is completed by combining the preset basic learning rate, momentum factor, and L2 regularization coefficient, and combined with the multi-class cross-entropy loss function, a complete network training solution system is formed.

[0054] Step S2-3-3, Batch Iterative Training Process: Step S2-3-3-1: The steady-state subset, transient variable load subset, and high / low temperature marine environment subset obtained in step S2-3-1-4 are globally randomly shuffled, and then divided into a training subset and a validation subset. The training subset is divided into multiple mini-batch standardized offline feature vectors according to a preset batch size. Each mini-batch is synchronously bound to a corresponding manually labeled fault label for subsequent iterative forward and backward propagation and weight updates. The validation subset is stored separately and includes validation subset features and corresponding manually labeled fault labels, which are used to calculate the validation loss in the iteration stop determination step. Step S2-3-3-2, Forward Propagation: Based on the initial neural network model obtained in step S2-3-2, read the number of input layer nodes of the initial neural network model obtained in step S2-3-2, and verify that the dimension of the input vector is consistent with the dimension of the mini-batch standardized offline feature vector; if the verification is consistent, then send the mini-batch standardized feature vector into the input layer of the initial neural network model. Each component in the mini-batch standardized feature vector is the corresponding input neuron. The first layer of matrix weighting is performed, and the result is input into the ReLU activation function. The ReLU function outputs the first layer intermediate activation value. The first layer intermediate activation value is then fed into the second layer of matrix weighting, and passed through the ReLU activation function again to generate the second layer intermediate activation value. The second layer intermediate activation value is then fed into the output layer matrix operation, and after Softmax activation, the predicted probability of various faults in a single batch is output. The corresponding manually labeled fault labels are retrieved, and the cross-entropy loss value of this batch is calculated by combining the manually labeled fault labels with the predicted probability of various faults in a single batch. This yields the first layer intermediate activation value, the second layer intermediate activation value, the predicted probability of various faults in a single batch, and the batch cross-entropy loss value.

[0055] Here, the first layer weighting matrix and the second layer weighting matrix are the weight matrices of the first hidden layer and the second hidden layer inside the initial neural network model.

[0056] Step S2-3-3-2, Backpropagation: Using the batch cross-entropy loss value as the total loss benchmark, the error of each neuron is derived layer by layer from the output layer to the input layer based on the deviation between the predicted probability of various faults in a single batch and the manually labeled faults, and the intermediate activation values ​​of the first and second layers. Then, by combining the weight matrices of each layer of the network, the gradients of all network parameters can be solved. ; Obtain the error of all neurons and gradients of all network parameters ; , among which, The error term of the j-th layer neuron, The symbol for partial differential represents the rate of change of a multivariate function with respect to only one variable. Loss, or batch cross-entropy loss, measures the difference between the model's fault prediction results and manually labeled results. Represents the connection weights in a neural network, representing the weight coefficients from the i-th input neuron to the j-th output neuron; This represents the input features of the i-th neuron; Step S2-3-3-2, Improve the weight update process: Based on the error of all neurons Network parameter gradient Preset iterative adaptive learning rate Let the adaptive momentum factor be set. Input characteristics of this round of forward propagation Previous iteration network weights The network weights in the previous two iterations , The network weights for this iteration are calculated using the following improved weight update formula. Iterative updates: ; The network weights in this round of iteration Write back to the initial neural network model obtained in step S2-4 to obtain the complete neural network after weight update; use it as the basic parameters for the next iteration. Here, the BP neural network module is based on the standardized feature vector and corresponding symptom list output in step S2-2; the network structure of the BP neural network module is: a 36-dimensional input layer (chiller unit example), two hidden layers with ReLU activation, and an output layer with Softmax. The weight update formula for the improved BP neural network for multi-fault classification is as follows: ; In the formula: For the first k Neuron at the next iteration i To neurons j The weights, η ( k ) is the first k Learning rate at the next iteration δ j For neurons j Error term, x i For neurons i Input, α ( k ) is the first k The momentum factor at the next iteration; offline periodic training and online real-time inference are performed. Low-confidence results trigger expert rule secondary review, outputting the fault probability distribution vector. The fault probability distribution vector and the corresponding standardized feature vector are then fed into the Kalman module.

[0057] Step S2-3-3-4, Iteration Stop Judgment: The validation subset features obtained in step S2-5 are fed into the complete neural network after weight update, and forward propagation is performed to calculate the validation loss; it is determined whether the validation loss has converged for multiple consecutive rounds, or whether the number of iterations has reached a preset maximum value; if either condition is met, all training loops are terminated, and a converged complete neural network is obtained; if not, the process returns to the forward propagation step in step S2-3-3-2 to start a new round of iteration until a converged complete neural network is obtained; lightweight compression and encapsulation are performed on the converged complete neural network to generate a lightweight optimal model weight file; the lightweight optimal model weight file is used as a static configuration parameter, internally encapsulating the network carrier matching the feature dimension, the weights completed in each layer iteration, the ReLU and Softmax activation function configurations, and the offline global normalization operator generated in S2-3-1-2, and is persistently stored after completion; Step S2-3-4: Read the feature dimension values ​​of the real-time standardized feature vector from step S2-2, and perform a consistency check between the read feature dimension values ​​and the historical sample standard dimensions built into the offline global normalization operator. If the dimensions do not match, output a dimension anomaly alarm and terminate the current inference process. If the dimensions match, call the offline global normalization operator from step S2-3-1-2 to perform scale alignment check and scale correction on the real-time standardized feature vector, eliminate the data scale drift caused by real-time single-sample normalization, and obtain the real-time standardized feature vector and corresponding feature dimension values ​​for inference. Step S2-3-5 includes: Step S2-3-5-1: Read the number of network input layer nodes built into the lightweight optimal model weight file, and combine it with the feature dimension value corresponding to the real-time standardized feature vector for inference passed in step S2-3-4 to verify that the feature dimension value of the real-time standardized feature vector for inference in step S2-3-4 is consistent with the number of network input layer nodes in the lightweight optimal model weight file; if the verification is consistent, send the real-time standardized feature vector for inference into the network input layer of the lightweight optimal model weight file. Step S2-3-5-2, Forward Inference of Real-Time Standardized Feature Vectors for Inference: Load the weight parameters of each layer stored in the lightweight optimal model weight file, and sequentially perform first-layer matrix weighting and ReLU activation to obtain the first-layer intermediate activation value; input the first-layer intermediate activation value into the second-layer matrix weighting and ReLU activation to obtain the second-layer intermediate activation value; send the second-layer intermediate activation value into the output layer matrix operation, and output the original predicted probability distribution vector of the real-time sample after Softmax activation operation; based on the probability dispersion and maximum predicted probability amplitude of the original predicted probability distribution vector of the real-time sample, calculate the baseline total confidence of the diagnosis, which is used to characterize the initial credibility of the overall fault diagnosis result and provide a baseline numerical basis for subsequent auxiliary verification and correction.

[0058] Step S2-3-5-3: Extract the fault types and fault confidence values ​​from the symptom list in S2-2, and perform auxiliary verification with the original predicted probability distribution vector of the real-time samples. First, sort all the original predicted fault probabilities in the real-time sample original predicted probability distribution vector, and select the two fault types with the highest probabilities. Then, match and compare these two fault types with the actual fault types recorded in the symptom list. If the actual fault type in the symptom list is included in the two fault types with the highest probability, then the prediction result is determined to match the actual symptoms. The corresponding fault confidence value is used to corroborate the validity of the abnormal working condition. The matching correction coefficient is set to 1.0. The original predicted probability distribution vector of the real-time sample is retained, and the baseline diagnostic total confidence obtained in step S2-3-5-2 is used and remains unchanged. Here, the original predicted probability distribution vector of the real-time sample is used to distinguish the relative probability of occurrence of various faults; the diagnostic total confidence is used to characterize the credibility of the overall fault diagnosis result. If the fault types in the matched symptom list and the two fault types with the highest probabilities belong to the same sub-fault type under the same major fault category, it is determined that the model's major category diagnosis is accurate, but the sub-fault diagnosis has bias. Based on the corresponding fault confidence value, an adaptive correction coefficient is selected within the range of 1.05 to 1.1. The higher the fault confidence, the closer the correction coefficient value is to the upper limit of the range. The predicted probabilities of all sub-faults under this major fault category are slightly weighted and increased. After weighting, global normalization is performed on all fault probabilities to ensure that the sum of all fault predicted probabilities is 1, resulting in a predicted probability distribution vector after auxiliary verification. At the same time, the baseline overall diagnostic confidence of the current sample is slightly increased for correction coefficients greater than 1, generating an updated overall diagnostic confidence. If the fault types in the symptom list do not have any major or sub-category matching relationship with the two most probable fault types, it is determined that the model prediction deviates significantly from the actual operating conditions. Based on the corresponding fault confidence value, an attenuation correction coefficient less than 1 is adaptively selected. The higher the fault confidence and the more significant the abnormal operating conditions, the greater the attenuation force. The baseline overall diagnostic confidence of this sample is lowered, an updated overall diagnostic confidence is generated, and this diagnostic result is marked as an abnormal and questionable state. The original predicted probability distribution vector of the real-time sample is retained.

[0059] Step S2-3-5-4: Based on the forward inference of the network's first-layer intermediate activation values, second-layer intermediate activation values, and the inference-specific real-time standardized feature vector output in step S2-3-4, combined with the network weight parameters in the lightweight optimal model weight file of step S2-3-3-4, the gradient attribution algorithm is used to calculate the contribution ratio of various fault probabilities in the original predicted probability distribution vector of the real-time sample output in step S2-3-5-2 dimensionally, and the core measurement point feature set with the highest contribution ratio is selected to finally obtain the key fault contribution features corresponding to this diagnosis; the predicted probability distribution vector after symptom list auxiliary verification, and the updated diagnostic total confidence and key fault contribution features are output; the predicted probability distribution vector after symptom list auxiliary verification includes: the original predicted probability distribution vector or the predicted probability distribution vector after auxiliary verification.

[0060] Step S2-3-6 includes: Step S2-3-6-1: The main fault is determined using the maximum probability criterion. The maximum probability criterion is as follows: extract the fault category with the largest value from the predicted probability distribution vector after symptom list-assisted verification, and take it as the current main fault of the equipment. The current main fault of the equipment serves as the core judgment basis for subsequent expert review, fault alarm, and fault tracing. The updated diagnostic total confidence is compared and verified with the system's preset confidence threshold. Step S2-3-6-2: If the updated overall confidence of the diagnosis is lower than the preset threshold, or if there are abnormal or questionable markers in this diagnosis, the expert rule secondary review mechanism is activated. Based on the identified current main fault of the device, the current main fault of the device is reviewed and corrected in conjunction with the symptom list in step S2-2. The current main fault of the device and the corresponding predicted probability distribution vector after review and correction are updated simultaneously. At the same time, the difficult time series sample carrying the updated label of the current main fault of the device is stored in the low confidence sample cache library for subsequent monthly offline incremental training to expand the fault dataset and optimize the model's generalization ability. Here, the expert rule secondary review mechanism is a set of logical judgment rules that are pre-fixed based on equipment operation and maintenance, fault mechanism and industry operation and maintenance experience. It is specifically designed for diagnostic samples with low confidence and marked as abnormal and questionable. It performs manual experience-based verification and correction on the main fault and various fault probabilities given by the model, making up for the limitations of pure data models that rely solely on measurement point data for judgment.

[0061] Expert rules, including three categories of rules: 1) Fault Coupling Matching Rules: Bind the actual alarm symptoms of the equipment output in step S2-2, and verify whether the main fault determined by the model conflicts with the currently synchronously reported fault symptoms. For example: The symptom list only records the "cooling pipe blockage" alarm, but the model determines that the main fault is motor overload. The rule identification category does not match at all. Reduce the probability of this fault and change the main fault.

[0062] 2) Measurement point threshold constraint rules: Combine the normal or abnormal interval threshold verification of each measurement point of the equipment: If the model determines a certain fault, but the corresponding key measurement point values ​​are all in the normal range, the rule reduces the weight of the fault; if the corresponding measurement points exceed the limit in a large area, the credibility of the corresponding common source fault is increased.

[0063] 3) Fault timing association rules: Combine historical timing samples to determine the fault occurrence logic: Some faults have a sequential triggering relationship. If the preceding fault does not occur, the subsequent fault is directly output. The rule is marked as unreasonable and the probability distribution is corrected.

[0064] In step S2-3-6-3, if the updated overall diagnostic confidence is higher than the preset threshold, the predicted probability distribution vector after symptom list-assisted verification and the corresponding current device main fault determination result are directly adopted in step S2-3-5-4, without the need for secondary verification and sample caching operations. Step S2-3-6-4: The predicted probability distribution vector after review and correction or direct acceptance is encapsulated to generate a deterministic fault probability distribution vector. The deterministic fault probability distribution vector carries the label of the current main fault of the equipment corresponding to this diagnosis.

[0065] Traditional BP neural networks suffer from slow convergence speed and susceptibility to local minima. This invention presents a fault diagnosis algorithm based on an improved BP neural network. This algorithm employs adaptive learning rate and momentum factor to enhance the BP neural network, dynamically adjusting these parameters to improve convergence speed and diagnostic accuracy. This algorithm can analyze a large number of operating parameters collected at the device's edge and identify different fault modes.

[0066] The BP neural network module functions as a fault discriminator with multiple inputs (feature vectors, symptom labels) and multiple outputs (fault category probability distributions). An improved weight update rule (including adaptive learning rate η(k) and momentum α(k)) is employed to enhance convergence and robustness.

[0067] Operational strategy: Primarily online inference, with support for periodic offline or incremental training at the edge; for uncertain or low-confidence diagnostic results, a manual confirmation process is triggered and samples are recorded for subsequent training.

[0068] Results output: Fault type, probability, key contribution features, and diagnostic confidence.

[0069] Specifically, in one embodiment of the present invention, as shown in Figure 5, the training and inference process of the BP neural network in the fault diagnosis of the ship air conditioning system is improved, emphasizing the key points of engineering implementation and real-time constraints.

[0070] (1) Data preparation and preprocessing: Data source: 294 measured variables collected from the edge of the equipment as raw input, including temperature, humidity, flow rate, differential pressure, current, voltage, frequency, vibration velocity / acceleration, and equipment operating condition signals. Supplemented by expert-annotated fault type labels and operating conditions (such as steady-state / transient, load level).

[0071] Standardization: Normalize each feature to zero mean and variance (or min-max normalization). The same set of normalization parameters is used for training and inference to ensure consistency.

[0072] (2) Network structure and parameter initialization: Network structure: A multi-layer feedforward backpropagation neural network is used (input layer: N standardized features; several hidden layers: usually 2-3 layers, the number of hidden units is adjusted according to experience and computing power; output layer: softmax or multi-label sigmoid for fault categories). The activation function can be ReLU or tanh, and the output layer uses softmax to implement multi-class diagnostic probabilities.

[0073] Improvements: An improved weight update formula (including adaptive learning rate η(k) and momentum factor α(k)) is adopted, and L2 regularization is added to suppress overfitting. Momentum and learning rate can adaptively decay with iterations or be self-adjusted based on the validation set.

[0074] (3) Training process (offline periodic training): Forward propagation: Calculate the output of each layer for a batch of samples, and finally obtain the predicted distribution and loss (commonly cross-entropy loss + L2 regularization term).

[0075] Backpropagation: Calculating the output error δj Then, the error term is calculated layer by layer to obtain the gradient. ∂L / ∂w ij .

[0076] Improved weight update: Uses a given weight update format. ,in, η (k) For adaptive learning rate (which can be based on validation error or use a scheduler such as Adam / learning rate decay strategy), α(k) As momentum, both can be gradually adjusted during training iterations to accelerate convergence and suppress oscillations.

[0077] Batch training and regularization: Update in mini-batch format, use Dropout or weight decay to avoid overfitting; evaluate validation set performance after each epoch.

[0078] Stopping criteria: Training is stopped when any of the following conditions are met: verification loss no longer decreases for several rounds (early stopping), the maximum number of iterations is reached, or both training error and verification error converge.

[0079] Model selection and calibration: Save the optimal model weights, and perform temperature scaling or threshold calibration on the output probabilities to obtain reliable confidence.

[0080] (4) Reasoning (online diagnosis) process: Feature construction and normalization: Construct time series and statistical features according to the training method and apply the same standardized parameters.

[0081] Forward inference: Standardized features are input into the network to calculate the output probability of each fault category. To ensure real-time performance, the forward inference only involves matrix multiplication and activation, and the computational complexity has been optimized during network pruning.

[0082] Decision logic: Determine the fault category based on the maximum probability or a multi-threshold strategy, and output the confidence level. When the confidence level is lower than the threshold or the probabilities of multiple categories are close, trigger further diagnosis (such as expert rules, spectrum analysis, or reporting for manual review).

[0083] Results fusion: The results are fused (weighted or voted) with the fault symptom detection and state prediction results based on expert rules to generate a final diagnosis and health rating, which is then written to PostgreSQL for visualization and historical tracking.

[0084] (5) Engineering deployment and online learning: Lightweighting and acceleration: Perform model pruning and quantization (e.g., INT8) and use an efficient linear algebra library; use TensorRT or ONNX Runtime for acceleration when necessary.

[0085] Incremental training and adaptation: Implement an incremental learning mechanism at the edge, periodically report new labeled samples or difficult cases to the upper system for batch retraining or fine-tuning, and distribute the updated model to the edge for replacement.

[0086] Interpretability and logging: Output the contribution of key features (e.g., based on sensitivity analysis or SHAP approximation) and record model inputs, outputs and confidence levels to facilitate post-analysis and accountability.

[0087] Preferably, in this embodiment, as shown in Figure 6, the network structure employs a multi-layer feedforward BP neural network. The activation function can be ReLU or tanh, and the output layer uses Softmax to implement multi-class diagnostic probabilities.

[0088] (1) Input layer: n A standardized feature.

[0089] (2) Several hidden layers: usually 2 to 3 layers, the number of hidden units is adjusted according to experience and computing power.

[0090] (3) Output layer: Softmax or multi-label Sigmoid for fault categories.

[0091] (4) Improvements: An improved weight update formula is adopted, which includes an adaptive learning rate. η(k) With momentum factor α(k) L2 regularization is added to suppress overfitting. Momentum and learning rate can be adaptively decayed with iterations or self-adjusted based on the validation set.

[0092] (5) Initialization: The weights are initialized using the Xavier or He method, the initial momentum is set to a small value, and the initial bias is 0 or a small positive value to prevent dead neurons.

[0093] like Figure 7 As shown, in steps S2-4, the Kalman time-series state prediction module is used to obtain short-term performance prediction values, NIS anomaly identifiers, and filtered and smoothed standardized feature vectors based on the real-time data stream of the original measurement variables. Step S2-4-1: Based on the components of the real-time standardized feature vector obtained in step S2-1, construct the device discrete-time state vector. Among them, each component of the real-time standardized feature vector corresponds to all monitoring points of the ship's air conditioning system; Step S2-4-2: Based on the real-time data stream of the original measured variables from step S1, construct the sensor observation vector. Based on the actuator control parameters obtained in step S1, the control input is constituted. (Control parameters such as fan frequency, actuator settings, and valve opening); Sampling time; Here, , Although data was collected synchronously at the same time, the objects being collected were different. For equipment status measurement. External control commands that drive changes in the state of the equipment; This is an index for discrete time-series sampling times. Represents the previous sampling period. This represents the current calculation sampling period.

[0094] It is the system's true state vector defined by the Kalman discrete-time model, using standardized eigenvectors as the modeling basis, and iteratively updated through time-series recursion, filtering, and correction; the posterior optimal state after iterative correction. After inverse mapping, the final output of the module is the filtered and smoothed standardized feature vector.

[0095] Step S2-4-3: Construct a discrete linear time-series state model for the equipment. The model includes state equations and observation equations. ; This is the state transition matrix; To control the input matrix; This is process noise; for Real-time raw sensor observation vectors are used to correct and predict the state, while retaining the true physical dimensions. The observation matrix; To observe noise; , The noise consists of uncorrelated, zero-mean Gaussian white noise. Step S2-4-4: The module first performs a time update (prior prediction): the expected value of the state equation is taken to eliminate zero-mean process noise and obtain the state prediction equation. By expanding the state prediction error covariance and removing irrelevant cross terms, the covariance prediction equation is obtained. The following system of time update equations is constructed simultaneously: ; Relying on historical status With control input Complete one-step prior prediction, and based on the prior state, continue multi-step extrapolation and iteration to generate short-term performance prediction values, thereby realizing the prediction of equipment parameter degradation trends; In the formula, For based on The state obtained at time step The prior estimate of the state at a given time is used, and based on this formula, multiple future time series are extrapolated iteratively to obtain short-term performance predictions of key parameters such as evaporation temperature, compressor load, and pump head. for Prior covariance matrix at time step; The process noise covariance matrix; , They are respectively The posterior state estimate at time t and the posterior error covariance matrix; Step S2-4-5: Perform measurement update (posterior correction), using the following set of equations to complete the entire process of residual calculation, Kalman gain solution, and state and covariance correction: ; In the formula, for The innovation vector is observed at all times, representing the residual between the original measured observations and the model predicted observations; To innovate the covariance matrix; To observe the noise covariance matrix; for The optimal Kalman gain at any given time is determined by adaptively assigning confidence weights between model predictions and sensor measurements. , After the update The posterior optimal state estimate and posterior error covariance matrix at time t; It is the identity matrix; Step S2-4-6: Based on the innovation vector obtained by solving the first two equations of the measurement update equation system. Innovative covariance matrix ,calculate : ; This is a scalar statistic, following a chi-square distribution that matches the observed dimension. A chi-square threshold is pre-set based on the confidence level. If the calculated... If the value exceeds the preset chi-square threshold range, the current measurement point is marked as abnormal and an NIS abnormality identifier is generated. After the abnormality occurs, robust processing measures such as reducing sensor weight, skipping the measurement update of this frame, and amplifying the state covariance are taken, and the sensor fault diagnosis subroutine is triggered. Step S2-4-7: Complete the state optimization update using the following three equations after measurement update: Correct the prediction bias through adaptive Kalman gain weighting to obtain a high-precision posterior state. Based on the global normalization operator generated in step S2-3-1-2, the posterior state of physical quantities is transformed into a standardized statistical scale, generating a filtered and smoothed standardized feature vector. This completely removes instantaneous noise and jumps in the original standardized features, resulting in higher temporal stability. Simultaneously, the posterior covariance is updated. This provides an uncertainty benchmark for the next iteration. ; Step S2-4-8: If the hardware communication message in step S1 is lost and the sensor cannot observe the vector... Then it cannot be solved. and It directly skips the measurement update equations and NIS discrimination calculations for this set of measurements, relying solely on the synchronously acquired control inputs. Execution time prediction; where the time prediction output is The prior error covariance matrix (or simply covariance matrix) is used to characterize the uncertainty of state estimation in pure time-series inference. When a message is lost, the diagonal elements of the prior error covariance matrix are amplified, indicating that there is no actual observation correction in this frame and the reliability of state prediction decreases. The amplified prior error covariance matrix will be used as the iterative input for the next sampling period, allowing subsequent filtering to incorporate more sensor measured data to compensate for prediction drift.

[0096] Here, the S2-4 module performs all four calculations and outputs, which are then fed into the S2-5 FAHP fuzzy hierarchical comprehensive health assessment module: 1) Short-term performance predictions obtained by time-updated extrapolation; 2) NIS anomaly markers obtained by NIS chi-square test; 3) The filtered and smoothed standardized feature vector obtained from the posterior optimal state mapping; 4) Deterministic failure probability of upstream link pass-through.

[0097] The Kalman filter state prediction module in step S2-4 is based on the system dynamics model for state estimation and short-term prediction (state prediction equation and covariance prediction equation), and is used for smooth estimation and fault precursor prediction of key quantities (such as evaporation temperature, compressor load, pump head, etc.).

[0098] The Kalman filter algorithm can provide optimal estimation of a system's state even in the presence of noise. This invention applies it to the state prediction of ship air conditioning equipment. By establishing the equipment's state equation and observation equation, it predicts the changing trends of the equipment's performance parameters in real time, and detects potential faults and performance degradation in advance.

[0099] Preferably, in one embodiment of the present invention, the algorithm implementation, parameter settings, and anomaly handling are explained step by step, taking into account the characteristics of the ship's air conditioning system (multiple sensors, high noise, unstable communication, and real-time requirements) in conjunction with the Kalman filter state prediction shown in Figure 7.

[0100] (1) Modeling and Quantization: Determine the state vector x ( k Select the state variables of the equipment / subsystem to be predicted, such as chiller evaporation temperature, condensation temperature, compressor speed, chilled water outlet temperature, pump flow rate, or vibration amplitude. Example: .

[0101] Determine the observation vector : These are measurable sensor quantities on-site, typically linear or approximately linear mappings of state vectors, such as cabin return air temperature, chilled water temperature difference, current value, pressure difference, etc.

[0102] Establish a discrete-time linear system model (or a linearized model): ； ,in and These are process noise and measurement noise, respectively, both assumed to be zero-mean Gaussian white noise, with covariances of respectively. Q and R .

[0103] (2) Initialization: Initial state estimation The first frame can be used to measure the pass rate. Estimate, or use empirical steady-state values.

[0104] Initial error covariance The settings are based on the understanding of the uncertainty of the initial value (a larger diagonal value can be set for unknown values).

[0105] choose Q ,R The settings are based on experience with sensor accuracy, operating condition fluctuations, and modeling errors. It is recommended to estimate using engineering experience first, and then adjust using historical data.

[0106] (3) Prediction (time update): State prediction: ; Covariance prediction: ； Note: This step is controlled by input. u ( k (e.g., actuator settings, fan frequency) and prior estimates from the previous step are used to advance to the current moment, which are suitable for real-time calculations on the edge side.

[0107] (4) Update (Measurement Update): Computational Innovation: ; Innovation covariance: ； Kalman gain: ; Status Update: ; Covariance update: Or in the Joseph form; (5) Abnormal data handling: Normalized Innovation Square (NIS): .like If the threshold is exceeded (as set by the chi-square distribution), an anomaly in measurement or a model mismatch is suspected.

[0108] In case of an anomaly, take the following measures: reduce the weight of the sensor (increase the corresponding weight). R Alternatively, the update of the measurement can be temporarily skipped (only a prediction is made), and the sensor fault diagnosis subroutine can be triggered (in combination with expert rules to determine whether manual repair is required).

[0109] (6) Adaptation and robustness: Adaptive adjustment Q / R : Based on historical innovation statistics, gradually adjusted Q or R In order to cope with changes in operating conditions and drift in noise characteristics.

[0110] Covariance inflation / contraction: This refers to the effect of covariance inflation / contraction during communication interruptions or when samples are sparse. P Appropriate inflation is implemented to reflect the increase in uncertainty.

[0111] Multi-sensor fusion: When multiple sensors measure the same physical quantity, the data is processed uniformly. HMatrix fusion utilizes Kalman gain for automatic weighting.

[0112] (7) Multi-rate and packet loss handling: For sensors with different sampling frequencies, a multi-rate Kalman update is employed: prediction is run at the fastest sampling step size, and an update is performed when a measurement arrives; if a measurement is lost, the update step is skipped, and only prediction and expansion are performed. P .

[0113] (8) Nonlinear extension and engineering implementation: For significantly nonlinear systems, the state transition and observation equations are linearized or sigma points are used by employing the Extended Kalman Filter (EKF) or the Unscented Kalman Filter (UKF).

[0114] To achieve deployment at the device edge (first-level visualization): consider floating-point precision, computation cycle, and memory limitations. Generally, the prediction and update cycle is consistent with the sensor sampling cycle (e.g., 1s~10s), while complex models can use longer step sizes or batch updates.

[0115] Results output: The state estimate, covariance, NIS, etc. are stored in the local PostgreSQL database for first-level visualization and reported to the second and third-level visualizations for health assessment decision-making.

[0116] like Figure 9 As shown, steps S2-5, the FAHP fuzzy hierarchical comprehensive health assessment module, are used for: S2-5-1, Multi-source index access and credibility correction processing: Based on the standardized feature vector after filtering and smoothing in step S2-4-6, steady-state operating parameters such as equipment temperature, differential pressure, flow rate, and vibration of the ship's air conditioning system are extracted, and steady-state operating indices are constructed and quantified to avoid noise and jump interference from the original sensors; Based on the deterministic fault probability distribution vector in step S2-3-6-4, the probability of occurrence of each fault mode is analyzed, and after normalization, a quantitative fault risk index is constructed to characterize the current fault risk level of the equipment; Based on the deviation residual calculation between the short-term performance prediction value and the rated standard operating condition in step S2-4-3, the degree of future operational degradation of the equipment is quantified, and a performance degradation index is constructed to characterize the short-term performance degradation trend of the equipment; Based on the NIS anomaly identification in step S2-4-5, a sensor reliability index is constructed. This index is used to correct the credibility of the three types of indices: steady-state operating index, fault risk index, and performance degradation index. If sensor drift or distortion is determined, the weights of the steady-state operating index, fault risk index, and performance degradation index of the corresponding measurement point will be reduced simultaneously to suppress abnormal data interference.

[0117] Here, the sensor reliability index is quantitatively calculated using the NIS chi-square test results, as shown in the following formula: ; In the formula: This is a sensor reliability value (ranging from 0 to 1). The threshold for the chi-square test; This represents the maximum exceeding NIS threshold. The lower the sensor reliability value, the more severe the distortion of the measurement point data, and the stronger the correction and suppression effect on the subsequent three types of operational evaluation indicators.

[0118] S2-5-2, Construction of the Hierarchical Structure for Health Assessment: Based on steady-state operation indicators, failure risk indicators, performance degradation indicators, and sensor reliability indicators, a three-layer FAHP health assessment hierarchy is constructed, resulting in a complete hierarchical health assessment architecture comprising a target layer, a criterion layer, and an indicator layer. The target layer is used for comprehensive health status assessment of ship air conditioning equipment; The criteria layer establishes four major evaluation criteria: operational stability, failure risk, performance trend, and measurement reliability. The indicator layer is used to embed the steady-state operation indicators, failure risk indicators, performance degradation indicators and sensor reliability indicators into the four evaluation criteria of the criterion layer, forming a structured, weighted, hierarchical evaluation model. Here, the four evaluation criteria include: Operational stability criterion: Match the steady-state operation indicators of the indicator layer; Failure risk criteria: Matching failure risk indicators at the indicator layer; Performance trend criteria: Matching performance degradation metrics at the metric layer; Measurement reliability criteria: Sensor reliability indices matching the index layer; S2-5-3, Construction of the triangular fuzzy judgment matrix: Based on the experience of ship air conditioning equipment operation and maintenance experts, pairwise fuzzy comparison of importance and matrix construction were completed in two stages: Criterion layer comparison: The importance of the four criteria—operational stability, failure risk, performance trend, and measurement reliability—is compared pairwise. The relative importance of the criteria is quantified using triangular fuzzy semantics, and a fuzzy judgment matrix for the criterion layer is constructed. Indicator layer comparison: For the four types of bottom-level indicators of the indicator layer, namely steady-state operation indicators, fault risk indicators, performance degradation indicators and sensor reliability indicators, the importance of each pair is compared. The relative importance between each indicator is quantified by triangular fuzzy semantics, and a fuzzy judgment matrix of the indicator layer is constructed. Here, a two-level matrix is ​​used to quantify the differences in the contribution of criteria and underlying indicators to the health status of equipment. Triangular fuzzy numbers are used to accommodate the subjective judgment range of experts, which solves the problems of fuzziness, distorted results, and difficulty in quantification in traditional single precise scoring subjective evaluation. Drawing on the experience of ship air conditioning equipment operation and maintenance experts, pairwise importance fuzzy comparison and matrix construction were completed in two stages. The triangular fuzzy numbers were expressed in interval quantization form as follows: ; In the formula: This is the lower limit critical value. For the most likely value, For the upper limit critical value, the matrix elements satisfy the reciprocal property: ; Criterion-level comparison: The importance of each of the four criteria—operational stability, failure risk, performance trend, and measurement reliability—is compared pairwise. The relative importance of the criteria is quantified using triangular fuzzy semantics, and a fuzzy judgment matrix at the criterion level is constructed. ; Indicator-level comparison: For the four categories of bottom-level indicators—steady-state operation indicators, fault risk indicators, performance degradation indicators, and sensor reliability indicators—pairwise importance comparisons are performed. Triangular fuzzy semantics is used to quantify the relative importance of each indicator, and a fuzzy judgment matrix for the indicator layer is constructed. ; S2-5-4, Fuzzy Weight Solution and Consistency Verification: For the fuzzy judgment matrix of the criterion layer and the fuzzy judgment matrix of the index layer, matrix synthesis, triangular fuzzy number defuzzification and geometric mean calculation are performed respectively to obtain the initial weights corresponding to the four criterion indicators and the four underlying indicators. The initial weights of the four criteria indicators and the four underlying indicators are normalized to obtain the standard weight coefficients of the four underlying indicators of steady-state operation, fault risk, performance degradation and sensor reliability and the standard weight coefficients of the four criteria. Consistency checks are performed independently on the fuzzy judgment matrix at the criterion layer and the fuzzy judgment matrix at the index layer: the maximum eigenvalue of each matrix is ​​calculated separately. Consistency index CI and consistency ratio CR; the consistency threshold is CR≤0.1. If any fuzzy judgment matrix CR>0.1 in the fuzzy judgment matrix of the criterion layer and the fuzzy judgment matrix of the index layer, it means that there is a subjective logical contradiction in the pairwise importance scores of the experts in the fuzzy judgment matrix. The corresponding fuzzy judgment matrix needs to readjust the expert scores and correct the matrix elements until the consistency requirements are met. After consistency verification and correction, a vector set of the four underlying indicators and the four criteria with reasonable allocation and no subjective contradictions is obtained, wherein the optimal weight is the corresponding standard weight coefficient after consistency verification.

[0119] Here, the four principles form a framework for four major evaluation categories and perspectives; the four underlying indicators are the specific scoring criteria for each category, with real-time data, and are the actual numerical values ​​used for calculation.

[0120] For the fuzzy judgment matrix of the criterion layer and the fuzzy judgment matrix of the index layer, matrix synthesis, triangular fuzzy number defuzzification, and geometric mean calculation are performed respectively to obtain the initial weights corresponding to the four criterion indices and the four underlying indices. The formula for calculating the geometric mean of a single row and column is as follows: ; The initial weights of the four criteria indicators and the initial weights of the four types of underlying indicators are normalized to obtain the standard weight coefficients of the four types of underlying indicators (steady-state operation, fault risk, performance degradation, and sensor reliability) and the standard weight coefficients of the four criteria. The normalization formula is as follows: ; Consistency checks are performed independently on the fuzzy judgment matrix at the criterion layer and the fuzzy judgment matrix at the index layer, respectively. The core formula for the consistency check is: ; In the formula: The largest eigenvalue of the matrix. Let be the order of the matrix. This is a random consistency indicator. The consistency threshold is... If any matrix This indicates that there is a subjective logical contradiction in the pairwise importance scores of the experts in the matrix. The corresponding fuzzy judgment matrix needs to be readjusted to adjust the expert scores and correct the matrix elements until the consistency requirements are met.

[0121] After consistency verification and correction, a vector set of the four types of underlying indicators and the optimal weights of each level of criteria are obtained, which are reasonably allocated and free from subjective contradictions.

[0122] S2-5-5, Calculation of Multi-Source Indicator Weighted Fusion and Comprehensive Health Index: Based on the optimal weights of the four underlying indicators, the basic scores for each item are calculated by weighting: The steady-state operation indicator is weighted by combining the optimal weights of the underlying indicators corresponding to the steady-state operation indicator to obtain the basic score for the real-time operation stability of the equipment; The fault risk indicator is weighted by combining the optimal weights of the underlying indicators corresponding to the fault risk indicator to obtain the basic score for the current fault risk of the equipment; The performance degradation indicator is weighted by combining the optimal weights of the underlying indicators corresponding to the performance degradation indicator to obtain the basic score for the future performance degradation trend of the equipment. The specific calculations are as follows: ; The sensor reliability index is introduced to simultaneously make a secondary correction to the above three basic scores: the sensor reliability value in the sensor reliability index is used. As a conversion factor, if the sensor reliability is too low, the corresponding sub-item score conversion factor is reduced to shield the evaluation bias caused by abnormal measurement points, and the operation stability score, fault risk score and performance degradation trend score are obtained after sensor reliability conversion correction. The specific calculations are as follows: ; Based on the optimal weights of the four criteria, a weighted fusion of secondary criteria is carried out: the corrected operational stability score, fault risk score, and performance degradation trend score, and the measurement reliability sub-score obtained by direct mapping from the sensor reliability index are multiplied by the optimal weights of the four criteria of operational stability, fault risk, performance trend, and measurement reliability, respectively, and summed to finally synthesize a comprehensive equipment health index that is continuously quantified in the range of 0 to 1.

[0123] Here, the measurement reliability sub-score is directly mapped from the sensor reliability index: ; Finally, a comprehensive equipment health index with continuous quantization in the range of 0 to 1 is obtained. ,as follows: .

[0124] S2-5-6, Health Level Grading and Intelligent Operation and Maintenance Strategy Output: Based on the preset health grading threshold and the comprehensive health index of the equipment, combined with the four-dimensional information linkage judgment of steady-state operation level, fault risk level, performance degradation trend, and sensor reliability status, the equipment status is divided into four health levels: healthy, sub-healthy, warning, and faulty. According to different health levels, corresponding intelligent operation and maintenance strategies are matched, including: routine monitoring, periodic inspection, spare parts warning, and shutdown maintenance, to achieve data-driven and precise operation and maintenance decision-making.

[0125] Here, the grading determination formula is as follows: .

[0126] The FAHP fuzzy hierarchical comprehensive health assessment module integrates fault diagnosis results, predictive information, and operational performance indicators. It uses a weight matrix and consistency test to calculate the weights of each factor and outputs health level, maintenance recommendations, and priorities based on fuzzy membership mapping.

[0127] Results output: Health score of the cascade system, recommended actions (such as maintenance, continued observation, spare parts preparation), and estimated remaining lifespan.

[0128] The calculation results are written to a PostgreSQL database: PostgreSQL + Prisma ORM is used to store raw data, cleaned results, model inference, logs and historical reports; indexed queries, archiving and periodic compression are supported.

[0129] Fuzzy Hierarchical Analysis (FAHP) combines qualitative and quantitative indicators to comprehensively evaluate complex systems. This invention utilizes this algorithm to assess the health status of a ship's air conditioning system. It establishes an evaluation index system from multiple dimensions, including equipment performance, operational reliability, and maintenance, determines the weight of each index, and ultimately obtains the system's health assessment results.

[0130] Preferably, in an embodiment of the present invention, as shown in Figure 8, the fuzzy hierarchical analysis (FAHP) health assessment decision-making process and weight calculation cover key steps such as modeling, fuzzy judgment, weight solving, consistency testing, indicator scoring, and health decision threshold setting.

[0131] (1) Hierarchical structure construction; The health assessment task is structured in a multi-level manner according to the objectives, criteria, and sub-criteria. The top level is the objective of "equipment / subsystem health assessment", the middle level consists of several performance dimensions (such as thermal performance, electrical performance, vibration / mechanical condition, sensor reliability, energy efficiency / operational efficiency, etc.), and the bottom level consists of specific measurement indicators corresponding to each performance dimension (number of temperature deviations, cooling capacity deviation, pump current imbalance, vibration acceleration amplitude, sensor drift rate, etc.).

[0132] Examples of criteria: operational performance degradation (C11), failure frequency / history (C12), maintenance cost / availability (C13), safety impact / criticality (C14), remaining life / prediction (C15). For each criterion, expert scoring is used to construct a fuzzy judgment matrix. The weight vector is obtained through fuzzy hierarchical analysis (FAHP). Finally, the on-site quantitative scores are aggregated according to their weights to obtain the overall health score. H This is mapped to operation and maintenance suggestions.

[0133] (2) Establishment of fuzzy comparison and judgment matrix; Experts used a fuzzy semantic scale to assess any two factors i , j Relative importance is given, and triangular fuzzy numbers are used to represent it. r ij =( l ij , m ij , uij ),satisfy r ij ≥ 0 and r ji = (1 / u ij , 1 / m ij , 1 / l ij ).

[0134] Fuzzy linguistic variables (e.g., equally important, slightly important, significantly important, very important) are used and mapped to triangular fuzzy numbers. Where: equally important = (1, 1, 1), slightly important = (2 / 3, 1, 3 / 2), significantly important = (3 / 2, 2, 5 / 2), very important = (2, 3, 4).

[0135] (3) Fuzzy composition and weight solution; No. i The formula for calculating the row fuzzy comprehensive value is as follows: ; Among them, fuzzy addition / multiplication is performed using triangular fuzzy number arithmetic.

[0136] Each S i Perform centering / defuzzification using the centering method: v i = ( l i + m i + u i ) / 3.

[0137] Geometric mean of each row element: ; Normalization yields the weight vector: .

[0138] (4) Compatibility test and consistency correction; If the real-valued judgment matrix of the traditional AHP is used, the largest eigenvalue can be calculated. λ max ,via RI is the random consistency index ( n =1..10, commonly used RI tables: 0, 0, 0.58, 0.90, 1.12, 1.24, 1.32, 1.41, 1.45, 1.49).

[0139] If the dispersion of expert fuzzy judgments is inconsistent (CR>0.1 or large fluctuations in fuzzy membership), the following strategies can be adopted: review with experts, apply least squares consistency adjustment, or use the geometric mean method to integrate multiple expert opinions to reduce inconsistency.

[0140] (5) Criterion scoring and health index calculation; Mapping on-site quantitative indicators (such as pressure difference, temperature difference, vibration amplitude, energy consumption, number of failures, and Kalman prediction residuals) to the scores of each criterion. s i Normalization is applied to [0,1] (1 being the healthiest, 0 the worst). Normalization method: For unfavorable indicators (larger values ​​indicate worse performance), normalization is applied... s = max(0, 1 - ( x - x min ) / ( x max - x min )).

[0141] Calculate the overall health score: ; H The value range is [0, 1], which facilitates visualization and classification: healthy (0.8~1.0), sub-healthy (0.6~0.8), early warning (0.4~0.6), and alarm (<0.4).

[0142] (6) Mapping of decision-making rules and operation and maintenance suggestions; according to H Decisions triggered by key single-item thresholds: If H If the score is <0.4 and the individual C14 (safety impact) score is <0.5, immediately stop the machine for inspection; if 0.4 ≤ H <0.6, short-term inspection is recommended and spare parts should be scheduled in advance; if 0.6 ≤ H If the value is less than 0.8, it is recommended to include it in the planned maintenance. H ≥ 0.8 standard monitoring.

[0143] By combining the remaining life predicted by Kalman Analyzer with the fault types diagnosed by BP, specific work order suggestions (component replacement, lubrication, cleaning, software parameter adjustment, etc.) are generated and priorities and expected maintenance time are estimated.

[0144] (7) Engineering implementation and self-adaptation; The weights and judgment matrix can be adjusted online and saved to PostgreSQL through the "expert configuration interface" of the edge control system; the system supports multi-expert scoring and geometric mean to reduce subjective bias.

[0145] Regularly retrain the weights using historical work orders and fault recovery results: use real fault consequences as supervision signals, apply simple Bayesian or improved BP-based regression to correct the weights, and achieve online adaptive adjustment.

[0146] The visualization includes weighted bar charts, radar charts of contributions for each criterion, health time series and trigger logs, and supports the central control layer to summarize key health indicators and recommendations via UDP broadcast.

[0147] (8) Fault tolerance and interpretability; FAHP's fuzzy representation better accommodates expert subjective uncertainty, and the defuzzification step and consistency check ensure the interpretability of the results. The system simultaneously saves the expert's original fuzzy judgment and the defuzzification weights for auditing and review.

[0148] Here, the Fuzzy Hierarchical Analysis (FAHP) method can combine qualitative and quantitative indicators to comprehensively evaluate complex systems. This invention utilizes this algorithm to assess the health status of a ship's air conditioning system, establishing an evaluation index system from multiple dimensions such as equipment performance, operational reliability, and maintenance status, determining the weight of each index, and ultimately obtaining the system's health assessment results.

[0149] Taking the CHL chiller system as an example, 36 relevant measurement points are selected as module inputs; the criterion layer has 4 indicators: refrigeration performance, electrical safety, mechanical vibration, and water circulation adaptation. The expert judgment matrix is ​​R=[[1,1 / 4,3,5],[4,1,7,9],[1 / 3,1 / 7,1,3],[1 / 5,1 / 9,1 / 3,1]]; the calculated weight WB=[0.22,0.62,0.096,0.63], and after consistency verification and fine-tuning, CR<0.1; each indicator is mapped to a [0,1] membership degree, and the weighted comprehensive health score H is obtained. The grading standard is: H≥0.8 healthy, 0.6≤H<0.8 sub-healthy, 0.4≤H<0.6 warning, H<0.4 fault. Inspection and spare parts replacement work orders are output simultaneously.

[0150] In detail, in one embodiment of the present invention, as shown in Figure 8, taking the health assessment decision of the chiller unit subsystem (third subsystem) of the ship air conditioning system as an example, the weight calculation and decision derivation process of fuzzy hierarchical analysis (FAHP) is completely decomposed. All data are based on the actual operating parameters of the ship air conditioning system and expert experience, ensuring that the calculation logic can be transferred to other subsystems.

[0151] In detail, the process of basic preparation, core calculation, consistency verification, and health decision derivation of this invention is as follows: I. Basic Preparations: Determining the Indicator System and Data Sources: 1.1 Construct a three-level indicator system: For the two chiller units (Sun system), focusing on their core failure risk points and health influencing factors, the following indicator system was constructed: (1) Target layer (A): Health of chiller unit subsystem.

[0152] (2) Criteria layer (B, first-level indicators): 4 core influence dimensions.

[0153] B1: Cooling performance (cooling capacity output, COP value); B2: Electrical safety (current, voltage, insulation resistance); B3: Equipment status (vibration velocity, bearing temperature, noise); B4: System compatibility (chilled water / cooling water flow rate, inlet and outlet water temperature difference).

[0154] (3) Indicator layer (C, secondary indicators): specific monitoring items under each criterion layer.

[0155] B1 includes: C1 (cooling capacity deviation rate) and C2 (COP value fluctuation). B2 Sub-subjects: C3 (Frequency of operating current exceeding the standard), C4 (Insulation resistance value); Subcategories of B3 include: C5 (peak vibration velocity), C6 (bearing temperature), and C7 (operating noise). B4 Sub-subjects: C8 (chilled water flow deviation), C9 (cooling water inlet and outlet temperature difference).

[0156] 1.2 Data Sources and Preprocessing: Input data: 36 key parameters related to the chiller unit (such as chilled water inlet temperature 12℃, outlet temperature 7℃, operating current 180 A, etc.) were taken from 294 variables monitored at the edge of the equipment.

[0157] Preprocessing: The data is cleaned by the "core calculation program" to remove outliers (such as an abnormal current of 500 A caused by a momentary sensor failure) and convert the raw data into quantitative values ​​such as "deviation rate" and "frequency of exceeding the standard" required by the indicator layer (such as C1 cooling capacity deviation rate = (actual cooling capacity - rated cooling capacity) / rated cooling capacity × 100%).

[0158] II. Core Calculations: Construction of Fuzzy Judgment Matrix and Weight Calculation: 2.1 Invite experts to construct a fuzzy judgment matrix: Five experts with over 10 years of experience in the operation and maintenance of ship chiller units were selected. Based on the 1-9 fuzzy scaling method (Table 1), pairwise comparisons were made to assess the importance of the criterion layers (B1-B4), generating a fuzzy judgment matrix (elements). rij express B i Compared to B j (the degree of importance).

[0159] Fuzzy judgment matrix of the criterion layer after expert consensus (AB layer): Represented by matrix R = [rij]4×4, where rows / columns correspond to B1-B4 respectively: R = [1, 1 / 4, 3, 5; # Comparison of B1 and B1-B4; 4, 1, 7, 9; # Comparison of B2 with B1-B4; 1 / 3, 1 / 7, 1, 3; # Comparison of B3 with B1-B4; 1 / 5, 1 / 9, 1 / 3, 1 ] # Comparison of B4 and B1-B4.

[0160] 2.2 Calculate the weight vector of the criterion layer (B layer): Based on the formula for calculating the weight vector, the "weighted average method" commonly used in fuzzy hierarchical analysis (FAHP) is employed to solve the problem, which involves three steps: Step 1: Calculate the geometric mean of the elements in each row of the judgment matrix: ; in: n The number of criteria layer indicators, here n =4.

[0161] The geometric mean of B1: w1' = (1×1 / 4×3×5)^(1 / 4) = (15 / 4)^(1 / 4) ≈ (3.75)^0.25 ≈ 1.39; The geometric mean of B2: w2' = (4×1×7×9)^(1 / 4) = (252)^(1 / 4) ≈ (252)^0.25≈ 3.98; The geometric mean of B3 is: w3' = (1 / 3×1 / 7×1×3)^(1 / 4) = (1 / 7)^(1 / 4) ≈ 0.61; The geometric mean of B4 is: w4' = (1 / 5×1 / 9×1 / 3×1)^(1 / 4) = (1 / 135)^(1 / 4) ≈0.39.

[0162] Step 2: For all w i Summation, calculation of normalized weights; formula: Summation: ∑ w i = 1.39 + 3.98 + 0.61 + 0.39 = 6.37.

[0163] Normalized weights: w 1 (B1 weight) = 1.39 / 6.37 ≈ 0.22 (22%); w 2 (B2 weight) = 3.98 / 6.37 ≈ 0.62 (62%); w 3 (B3 weight) = 0.61 / 6.37 ≈ 0.096 (9.6%); w 4 (B4 weight) = 0.39 / 6.37 ≈ 0.063 (6.3%).

[0164] Weight vector: W B = [0.22, 0.62, 0.096, 0.063] T (Verification: 0.22+0.62+0.096+0.063≈1, which meets the requirements).

[0165] 2.3, Weight calculation for the indicator layer (C layer) (taking B1 as an example): For C1 (cooling capacity deviation rate) and C2 (COP value fluctuation) under criterion layer B1 (cooling performance), repeat the above process: 1. Construct the fuzzy judgment matrix for layers B1-C: Experts determined that "the cooling capacity deviation rate (C1) is significantly more important than the COP value fluctuation (C2)," with a scale of 5, and the matrix is ​​as follows: R B1 = [1, 5; 1 / 5, 1] 2. Calculate the geometric mean and normalized weights: w 1'(C1)= (1×5)^(1 / 2) ≈ 2.236; w 2'(C2)= (1 / 5×1)^(1 / 2) ≈ 0.447; Summation: 2.236 + 0.447 ≈ 2.683; Weight: w C1= 2.236 / 2.683≈0.833 (83.3%); w C2 =0.447 / 2.683≈0.167 (16.7%); Weight vector: w C1 = [0.833, 0.167] T .

[0166] Similarly, the index weights of layers B2-C, B3-C, and B4-C can be calculated (e.g., the weight of C3 under B2 is 0.6, the weight of C4 is 0.4; the weight of C5 under B3 is 0.4, the weight of C6 is 0.35, the weight of C7 is 0.25, etc.).

[0167] Third, consistency test to verify the rationality of the judgment matrix: To avoid logical contradictions in experts' subjective judgments, a consistency test is required (CR < 0.1 is qualified), and the steps are as follows: 3.1 Calculate the maximum eigenvalue of the judgment matrix λ max : Taking the criterion layer (A-B) matrix as an example: Calculate the matrix R and the weight vector W B product: R × W B ; First row result: 1×0.22 + 1 / 4×0.62 + 3×0.096 + 5×0.063 ≈ 0.22 + 0.155+ 0.288 + 0.315 ≈ 0.978; Second row result: 4×0.22 + 1×0.62 + 7×0.096 + 9×0.063 ≈ 0.88 + 0.62 +0.672 + 0.567 ≈ 2.739; Third row result: 1 / 3×0.22 + 1 / 7×0.62 + 1×0.096 + 3×0.063 ≈ 0.073 +0.089 + 0.096 + 0.189 ≈ 0.447; Fourth row result: 1 / 5×0.22 + 1 / 9×0.62 + 1 / 3×0.096 + 1×0.063 ≈ 0.044 +0.069 + 0.032 + 0.063 ≈ 0.208.

[0168] Calculate ( R ×W B ) / ( n × w i And calculate the average value. n =4).

[0169] 0.978 / (4×0.22)≈1.114; 2.739 / (4×0.62)≈1.093; 0.447 / (4×0.096)≈1.164; 0.208 / (4×0.063)≈0.825; λ max = (1.114+1.093+1.164+0.825) / 4 ≈ 4.196 / 4 ≈ 1.049.

[0170] 3.2 Calculate the consistency index CI and the random consistency ratio CR: Consistency index: CI = ( λ max ‒ n ) / ( n ‒ 1) = (1.049‒ 4) / (4‒ 1) ≈ (‒ 2.951) / 3 ≈‒ 0.984 (The negative sign indicates that the consistency of the judgment matrix is ​​excellent).

[0171] Random Consistency Index (RI): Refer to the FAHP standard table. n = 4, RI=0.89.

[0172] Random consistency ratio: CR = |CI| / RI ≈ 0.984 / 0.89 ≈ 1.106.

[0173] IV. Health Decision Derivation, Fuzzy Comprehensive Evaluation and Result Output: 4.1 Determine the membership degree of the indicators and the health level: The actual monitoring values ​​of the indicator layer (layer C) are converted into "health level membership degree", and the health level is divided into 5 levels: Level 1 (Excellent): Index deviation ≤ 5%, membership degree 1; deviation 5%-10%, membership degree 0.8.

[0174] Level 2 (Good): Indicator deviation 10%-15%, membership degree 0.6.

[0175] Level 3 (General): Index deviation 15%-20%, membership degree 0.4.

[0176] Level 4 (Poor): Indicator deviation 20%-25%, membership degree 0.2.

[0177] Level 5 (Fault): Index deviation > 25%, membership degree 0.

[0178] Taking C1 (cold load deviation rate) as an example, if the actual deviation is 8%, then the membership vector is: [0.8, 0, 0, 0,0].

[0179] 4.2, Second-level fuzzy comprehensive evaluation (taking B1 as an example): The overall membership degree of B1 = index weight vector × index membership degree matrix.

[0180] Assume C1 has a membership degree of [0.8,0,0,0,0] and C2 (COP deviation 12%) has a membership degree of [0,0.6,0,0,0].

[0181] B1 Overall membership degree = [0.833, 0.167] × [0.8,0,0,0,0; 0,0.6,0,0,0] = [0.666,0.100, 0, 0, 0].

[0182] Take the level corresponding to the highest membership degree: B1 (cooling performance) health level is "excellent".

[0183] 4.3, First-level fuzzy comprehensive evaluation (target layer A): Target layer comprehensive membership degree = Criterion layer weight vector × Criterion layer membership degree matrix

[0184] Assume the overall membership degrees of B1-B4 are [0.666,0.1,0,0,0], [0.7,0,0,0,0], [0,0.5,0.5,0,0], and [0.9,0,0,0,0].

[0185] The overall membership degree of layer A = [0.22, 0.62, 0.096, 0.063] × the above matrix ≈ [0.68, 0.048, 0.048, 0, 0].

[0186] The target layer health level is "Excellent", and the corresponding health assessment decision is "maintain the current operating parameters and extend the maintenance cycle to 6 months".

[0187] 4.4, Output Results: Weight calculation data ( W B , W CThe health level and decision recommendations are saved to a PostgreSQL database. The industrial control all-in-one machine on the device edge displays the following content on the "Health Assessment Decision Interface": Chiller unit subsystem health status: Excellent (overall membership degree 0.68).

[0188] Weights for each criterion layer: Electrical safety 62%, Refrigeration performance 22%, Equipment status 9.6%, System adaptability 6.3%.

[0189] Recommended decision: No downtime for maintenance is required; increase vibration monitoring frequency by 1 time per week.

[0190] Step S3: Based on the complete algorithm dataset, the device edge side splits and assembles standardized JSON messages with CRC checksum, global message ID, serial number and device identifier, and uploads them in batches to the upper system side; Upper-level system receiving stage: The upper-level system uploads JSON messages based on the device edge side, completes CRC verification, message deduplication and timing alignment, and writes the obtained valid data in batches into the upper-level system's summary timing library, and generates system logs for abnormal messages; The complete algorithm dataset includes: The real-time data stream of the raw measurement variables in step S1 includes: equipment temperature, differential pressure, flow rate, vibration, raw sensor sampling values, equipment on / off status, and actuator feedback; and synchronously generated basic monitoring fields such as NIS residuals, measurement point out-of-limit markers, and sensor anomaly identifiers. The quantization calculation results generated in step S2: 1) Basic indicators obtained from S2-5-1: steady-state operation indicators, failure risk indicators, performance degradation indicators, and sensor reliability indicators; 2) Sub-item scores obtained from S2-5-5: basic score for operational stability, basic score for failure risk, and score for performance degradation trend; 3) S2-5-5 after correction: operational stability score, failure risk score and performance degradation trend score corrected by sensor reliability calculation, as well as independent measurement reliability sub-item score; 4) Overall results: The comprehensive health index of equipment in the 0~1 range obtained from S2-5-6, the four-level health level judgment value obtained from S2-5-6, the short-term performance degradation prediction value obtained from S2-4-3, and the intelligent operation and maintenance strategy label obtained from S2-5-6.

[0191] Here, the network outage caching process is as follows: when the network is out of service, the cache is persistently cached for 72 hours. After the link is restored, a retransmission message is generated and sent to the upper system side for receiving and processing. Step S3 finally outputs a summary time-series database of the entire upper-level system, which provides all the input data for the downstream step S4.

[0192] Specifically, the message encapsulation process is as follows: read the complete algorithm dataset from the PostgreSQL database on the device edge, split the JSON message into 14 subsystems, and attach CRC checksum, global message ID, serial number, and device identifier; use TCP Keep-Alive long connection for batch upload, with each message controlled within 1400 bytes to avoid IP fragmentation.

[0193] Upper-level system receiving and processing: Galaxy Kylin industrial control HTTP listening port continuously receives messages, and sequentially performs CRC verification, message ID deduplication, and multi-message timing alignment; legitimate data is written in batches to the upper-level unified PostgreSQL summary database using Prisma transactions, and duplicate / verification failure messages are directly discarded and recorded in the system log.

[0194] Disconnection caching: A persistent message queue is set up at the edge layer to cache the dataset to be uploaded for 72 hours when the network is disconnected. The data will be automatically resumed after the link is restored, and all aggregated data during the interruption period will be completed.

[0195] Step S4: UDP broadcast message encapsulation and push notification of the visualization layer on the upper-level system side: The upper-level system side divides the data in the aggregated time-series database into two categories for differentiated encapsulation: For the first category of data, i.e., real-time KPI values, no encoding conversion is performed; the data is directly encapsulated into the first UDP broadcast message to meet the high-frequency refresh requirements of the large screen on the centralized control side. For the second category of data, including Chinese text in fault diagnosis, health decision-making, and maintenance work orders, a preset text-to-number mapping table is matched to convert the Chinese text into the corresponding numerical encoding, and then encapsulated into the second UDP broadcast message to avoid the defect of garbled Chinese characters in native UDP transmission. Based on the first and second UDP broadcast messages after classification and differentiated encoding, a standardized UDP single data packet is generated. The UDP broadcast push service on the upper-level system side pushes a stream of UDP broadcast data packets containing standardized UDP single data packets to the global broadcast address 255.255.255.255 at a fixed period; the sending frequency is differentiated according to the event level, and the sending frequency of high-priority messages such as faults and warnings is increased; the fixed broadcast port 8086 is bound, and a multi-threaded asynchronous sending mechanism is used to reduce transmission latency.

[0196] Here, step S4 is deployed on the upper-level system side. All UDP broadcast data packets are actively generated by the upper-level system side and broadcast periodically. The reception and subsequent parsing of data packets are completed by the step S5 module deployed on the central control system side. The overall transmission direction is a one-way downward push from the upper-level system side to the central control system side.

[0197] Data classification and encoding refinement: Data read from the upper-level summary database is divided into two categories: real-time KPI values: directly encapsulated as unencoded UDP broadcast messages for high-frequency refresh of large screens; fault diagnosis, health decision, and maintenance work order text: after converting the numerical encoding through a preset mapping table, UDP messages are encapsulated to solve the defect that UDP cannot transmit Chinese text.

[0198] The broadcast push mechanism is refined: data packets are pushed to the broadcast address 255.255.255.255 at fixed intervals, and the priority of events is distinguished. The sending frequency of fault and early warning messages is increased; the broadcast port is fixed at 8086, and multi-threaded asynchronous packet sending reduces latency.

[0199] like Figure 3 As shown, in step S5, the visualization layer of the central control system performs UDP-to-HTTP fault-tolerant conversion and global persistence: Step S5 is deployed in the visualization layer of the central control system. It receives the UDP broadcast data packet stream sent by the upper system side in step S4 as input. The UDP-to-HTTP data conversion processing module of the central control system completes the full-link fault tolerance processing and finally outputs a global structured operation and maintenance database for use in the visualization rendering of step S6.

[0200] The module binds to port 8086 to receive UDP packets, and sequentially performs CRC verification, message deduplication and timing normalization, data parsing unit conversion, batch cache merging, and Chinese encoding reverse decoding. It has a built-in adaptive fragmentation and retransmission algorithm to adapt to packet loss scenarios in weak ship networks, and converts data into RESTful standard HTTP packets for layered fault-tolerant persistence. It has the ability to cache KPIs and replenish time-series data after the link is restored when the network is interrupted. At the same time, it is equipped with full-link monitoring, security verification and high availability deployment strategies to achieve reliable data transmission and global unified persistent storage across levels from the upper level to the centralized control.

[0201] (1) Network receiving layer (UDP listener): The UDP to HTTP data conversion and processing module on the central control side is bound to the 8086 fixed port to listen to the broadcast address. It receives all UDP packets with high concurrency based on asynchronous I / O, encapsulates the received UDP broadcast data packets into event messages, and stores them in the memory receiving queue to avoid I / O thread blocking and business coupling.

[0202] Specifically, a UDP listening service is deployed on the central control / upper-level system side, binding a pre-configured broadcast / multicast address (optional: 255.255.255.255 or 224.0.0.x) and port (port number is configurable), supporting IPv4 / IPv6.

[0203] Use multithreaded / asynchronous I / O (such as epoll, IOCP) to ensure high concurrent receive performance. Each received UDP packet is immediately encapsulated into a receive event and added to an in-memory receive queue to avoid complex processing in the I / O thread.

[0204] Message format and protocol: Application layer messages can use a unified JSON format, with fields including: message header (version number, sender ID, grandchild system ID, message type), timestamp (UTC), sequence number, data body, checksum (CRC32), and signature / token (optional).

[0205] Strictly limit the size of a single packet (recommended <1400 bytes) to avoid IP fragmentation. If the data volume is large, use an application-layer fragmentation and reassembly scheme (fragment ID, total number of fragments, fragment sequence number, timeout retransmission).

[0206] (2) Verification and preliminary filtering: For event messages in the memory receiving queue, verify the CRC32 check code, device whitelist, protocol version and legal signature packet by packet. Discard illegal messages, tampered messages and unauthorized device messages directly and keep the system alarm log to ensure the legality of the access data and obtain a valid UDP packet stream that has passed the verification. Specifically, upon receiving a message, you can first verify its CRC and digital signature (if enabled). If the verification fails, discard the message and log it.

[0207] Verify the message header version against the sender's whitelist; discard messages from unauthorized senders or with mismatched versions.

[0208] (3) Deduplication and order processing: For valid UDP packet streams that pass the verification, the global message ID is recorded using an LRU hash cache to remove duplicate packets; a 50ms sliding time window buffer is set, and the time-order rearrangement of short-term out-of-order packets is completed based on the timestamp and sequence number to ensure the continuity of time-order data and obtain a deduplicated and ordered standardized time-order packet sequence. Specifically, message ID + sequence number can be used for deduplication: maintain a short-term hash cache (such as LRU) to record the IDs of processed messages, and discard duplicate messages directly or return redundant responses.

[0209] For critical data that requires sequentiality (such as time series indicators), the sequence number and timestamp can be used to reorder the data. If the data is temporarily out of order, it can be buffered and reordered. If the timeout occurs, it can be submitted in an approximate order based on the timestamp.

[0210] (4) Parsing and business diversion: For the deduplicated and ordered standardized time-series message sequence, complete the JSON structured parsing and field integrity verification, and uniformly complete the engineering unit conversion (Pa / kPa, L / s / m³ / h); according to the message type, divert it into real-time KPI data channel and fault event and operation and maintenance decision event channel, realize the independent processing of the two types of data, and obtain the diverted real-time KPI dataset and fault operation and maintenance event dataset; Here, messages can be parsed into internal business objects (KPIs, diagnostic results, health decisions, etc.) and field integrity checks and unit conversions can be performed (e.g., pressure Pa→kPa, flow rate L / s→m3 / h).

[0211] Messages are routed based on type: real-time KPIs go through the real-time channel, diagnostics / decision-making go through the event channel, and historical data go through the batch inbound channel.

[0212] (5) Queue caching and batch merging: For real-time KPI datasets and fault operation and maintenance event datasets, rely on 1G ring memory queue to cache business data, trigger batch merging conditions (accumulated 100 records or waiting time 50ms) to automatically aggregate data packets, reduce frequent HTTP requests, improve overall throughput performance, and obtain aggregated business data packets after batch merging.

[0213] To address sudden traffic spikes and HTTP upload latency, a high-performance local queue (such as a memory ring queue or a Kafka / Redis queue) can be introduced. The queue supports persistent writes to prevent power outages (optionally, writes can be performed locally in SQLite / PostgreSQL as a buffer).

[0214] Supports batch merging: Merges small messages within a short time window into a single HTTP batch packet to improve efficiency (controls batch size and maximum latency, such as up to 100 messages or 50ms).

[0215] (6) Encoding Reverse Decoding Stage: For the Chinese text-digit encoding operation completed in the S4 stage to adapt to UDP transmission, reverse decoding is completed through a preset text-digit mapping table to accurately restore the complete Chinese fault description, equipment health assessment text, and maintenance work order suggestion information, completely solving the technical defects of Chinese character encoding in native UDP transmission, and obtaining a decoded dataset containing complete Chinese business information.

[0216] Specifically, the HTTP client module can convert business objects in the queue into the target API format (RESTfulJSON or Protobuf) and carry the necessary authentication headers (Bearer token, client certificate).

[0217] It uses connection pooling and Keep-Alive to support concurrent uploads; it adopts synchronous confirmation mode for critical messages and asynchronous upload and status recording for non-critical messages.

[0218] (7) Adaptive fragmentation retransmission fault tolerance link: For the decoded dataset, the optimal fragmentation size and transmission priority weight are dynamically calculated, and the network quality is judged in combination with the real-time packet loss rate; when the packet loss rate ρ>10%, the exponential backoff retransmission mechanism is started, with a maximum of 5 retries, to ensure the data integrity in weak network and high-latency marine network environments, and to obtain reliable service data packets with well-organized fragmentation and fault tolerance of retransmission. Specifically, failed uploads are retried according to categories: network failures use exponential backoff (e.g., 1s, 2s, 4s), and after reaching the maximum number of retries, the message is written to the database and an alarm is triggered; HTTP 4xx errors are treated as format / permission issues, and are not retried but are recorded.

[0219] When the upstream is unavailable, the queue enters backpressure mode, prioritizing the retention of the latest key KPIs, while historical data can be discarded or restored in batches according to the backspace strategy.

[0220] (8) HTTP protocol encapsulation and uploading process: Convert the fault-tolerant business data into RESTful standard JSON format, carry the Bearer authentication token to complete identity verification; enable HTTP connection pool and Keep-Alive long connection mechanism to support multi-threaded concurrent uploading, adopt synchronous confirmation mode for critical operation and maintenance events, and adopt asynchronous uploading for non-critical KPI data, taking into account both real-time transmission and reliability, and obtain standardized, authenticated HTTP request data packets; Here, after the HTTP upload is successful, the upper-level system and the central control system write to PostgreSQL via API or directly through Prisma ORM. The write uses transactions and batch inserts to ensure the atomic persistence of time-series data and diagnostic results.

[0221] Important events are recorded in both the event table and the audit log table for easy traceability.

[0222] (9) Hierarchical fault-tolerant persistence stage: A hierarchical fault-tolerant data entry strategy is adopted. After the HTTP request data packet is successfully sent, it is written to the global database of the centralized control PostgreSQL through Prisma ORM batch transactions, and the device event table and audit log table are updated synchronously to ensure data traceability. The classification and handling method when the HTTP request data packet fails to be written is as follows: network fluctuation failure starts the exponential backoff retry, and after the maximum number of retries is reached, the data is written to the database and an alarm is triggered; 4xx permission errors and format errors are not retried repeatedly, but are directly cached to the local SQLite buffer database as a backup to avoid data loss and obtain the global structured operation and maintenance data of the centralized control layer after the data is written to the database. Specifically, after writing to the database, a cache refresh or WebSocket / push mechanism is triggered to drive the large screen and industrial control terminal to update in real time.

[0223] For data that needs to be broadcast to other internal modules, internal HTTP callbacks or re-UDP broadcasts are supported (controlled to avoid loops).

[0224] (10) Detailed backup for network anomalies: When a network interruption occurs between the upper system side and the central control overall side, the conversion module automatically caches the latest real-time data of device KPIs to avoid the loss of large screen visualization data; after the network link is restored, the upper system re-investigates the historical encoded messages of the network period in full, and the central control side repeats the full link parsing, fault tolerance, and database entry process to accurately fill in the missing time series data, realize the seamless continuation of time series data, and obtain the complete time series operation and maintenance data after the network interruption is filled in; Specifically, the entire link records metrics such as reception rate, packet loss rate, parsing error rate, HTTP success rate, queue length, and database insertion delay, exposes the Prometheus / monitoring interface, and sets threshold alarms.

[0225] Enable TLS / HTTPS uploading, message signing, source whitelisting, and ACLs to ensure data integrity and security in the maritime network environment.

[0226] (11) Visualized downstream distribution, monitoring and security management: After storing the global structured operation and maintenance data and the complete time-series operation and maintenance data after network outage recovery on the central control side into the database, the cache refresh and WebSocket push are triggered to drive the central control screen and industrial control terminal to refresh in real time; support internal HTTP callback and controlled secondary UDP broadcast to realize multi-module data sharing; full-link statistical data reception rate, packet loss rate, parsing error rate, HTTP success rate, queue backlog, database entry delay and other indicators, open the Prometheus monitoring interface and configure threshold alarms; enable HTTPS encrypted transmission, message signature verification, device whitelist and ACL access control throughout the process to adapt to the complex network security transmission requirements at sea, and obtain the full-domain operation and maintenance data service capabilities that are monitorable, traceable and secure and compliant.

[0227] Specifically, redundant UDP receiving instances and load balancers can be deployed on the central control room's large screen or host computer, enabling hot standby switching.

[0228] Critical channels (diagnostic results, health decisions) should prioritize using dedicated ports and be configured with Quality of Service (QoS) policies to ensure low-latency delivery.

[0229] (12) Performance and deployment strategy: The system adopts redundant UDP receiving instance deployment, coupled with load balancing and hot standby switching mechanism to ensure high service availability; independent port and QoS service quality policy are configured for core business channels such as fault diagnosis and health decision-making to prioritize low latency and high reliability transmission of key data and adapt to the harsh operating environment of ship industrial control system.

[0230] Step S5 Final Output: The system adopts redundant UDP receiving instance deployment, coupled with load balancing and hot standby switching mechanisms to ensure high service availability; independent port and QoS service quality policies are configured for core business channels such as fault diagnosis and health decision-making to prioritize low-latency and high-reliability transmission of critical data, adapt to the harsh operating environment of ship industrial control systems, and obtain a global structured operation and maintenance database at the centralized control layer, providing a standardized input data source for the downstream step S6.

[0231] Step S6, Real-time Rendering and Display of the Three-Layer Hierarchical Visual Interface: Step S6 is the real-time rendering stage of the three-layer hierarchical visualization. It takes three types of hierarchical storage data as input: the edge local database from Step S2, the upper-level summary time-series database from Step S3, and the central control global operation and maintenance database from Step S5. It relies on three sets of visualization programs—edge, upper-level, and central control—to render the data in layers. It provides differentiated monitoring interfaces for field operators, on-duty engineers, and central control administrators: the edge side displays the detailed operation of a single device and local health alarms; the upper-level side summarizes the collaborative operating conditions and fault comparisons of multiple devices; and the central control screen presents the overall health status of the entire system and highlights high-risk devices for early warning. Finally, it outputs a three-layer hierarchical real-time monitoring view, completing the hierarchical visualization and display of the entire system data.

[0232] Step S6-1: The interface rendering program on the device edge side displays the complete algorithm dataset calculated and persisted in the edge local PostgreSQL database in step S2. This dataset includes the device's original operating parameters, filter standardization features, four types of basic evaluation indicators, sub-item health scores, device warning signs, and other real-time calculation results on the edge side. The device edge side provides a refined real-time monitoring view for a single system, meeting the needs of on-site operators to view equipment status, troubleshoot local faults, and export equipment health reports. Here, the interface on the edge side of the device (such as...) Figure 10(For on-site operators): Edge 1 has 81 interfaces (8 subsystems of air conditioning units + 1 subsystem of air conditioning units, 9 pages per subsystem); Edge 2 has 45 interfaces (2 subsystems of water chillers + 3 subsystems of water chillers, 9 pages per subsystem); 9 types of page details: main interface, operation performance monitoring, steady-state operation judgment, sensor fault, equipment fault alarm, equipment fault early warning, health assessment decision, historical data query, and health report export; The data source is the edge local PostgreSQL written in step S2, which refreshes traffic, temperature and humidity, vibration, health score, and early warning indicators in real time.

[0233] Step S6-2: The interface rendering program on the upper-level system side will persist the upper-level aggregated time-series database after verification, deduplication, and time-series alignment in step S3, and aggregate all time-series operating conditions, faults, and health assessment data uploaded by all edge devices, and display it on the interface on the upper-level system side; the multi-device collaborative aggregated monitoring view on the upper-level system side will support the on-duty engineer to carry out whole system operating condition inspection, fault tracing analysis, and batch operation and maintenance report output; Here, the interface on the upper-level system side (such as...) Figure 12 (For use by on-duty engineers): A total of 98 interfaces, 14 sub-systems, each with 7 pages; 7 types of page details: main interface, basic equipment information, summary of operating performance, summary of fault detection, summary of fault diagnosis, system health assessment, and report printing; The data source is the upper-level summary database in step S3, displaying cross-device collaborative operating conditions and related fault comparisons.

[0234] Step S6-3: The interface rendering program on the central control side displays the global structured operation and maintenance database, which is the most complete and reliable time-series operation and maintenance aggregate dataset for the entire system, after UDP-to-HTTP fault-tolerant conversion, decoding and completion, and global persistence in step S5, on the large screen interface on the central control side. Here, the large screen interface on the central control side (such as...) Figure 13 (For use by the centralized control administrator) A total of 56 interfaces, 14 sub-systems, each with 4 pages; 4 types of page details: global view of running performance, aggregated view of fault detection, global view of fault diagnosis, and health assessment status view; The large screen displays the overall health score and total number of faults of the entire system, with high-risk equipment highlighted in pop-up windows; The data source is the centralized control global database converted in step S5.

[0235] In one embodiment of the present invention, combined with Figure 10As shown in Figures 11-12, the device edge-side (first-layer visualization) software interface of this invention has 9 monitoring interfaces for each sub-system, including: main interface, operational performance monitoring, steady-state operation discrimination, sensor fault, electrical fault, thermal fault, health assessment decision, historical data query, and health assessment report; the upper-level system-side (second-layer visualization) software interface has 7 monitoring interfaces for each sub-system, including: main interface, equipment information, operational performance, fault detection, fault diagnosis, health assessment, and report printing; the central control overall-side (third-layer visualization) software interface has 4 monitoring interfaces for each sub-system, including: operational performance, fault detection, fault diagnosis, and health assessment. This can meet the different information needs of different management and operation personnel. Management and operation personnel can obtain information of different granularities at different levels (equipment operators / field technicians, system administrators / duty engineers, central control room managers).

[0236] (1) The first level (equipment operators / site technicians) needs detailed equipment operation details in order to detect and deal with equipment failures in a timely manner.

[0237] (2) The second level (system administrators / duty engineers) needs to understand the operating status and coordination of each subsystem in order to ensure the stable operation of the entire system.

[0238] (3) The third level (central control room manager) needs to have a macro-level understanding of the entire air conditioning system in order to make overall decisions.

[0239] The visualization at these three levels needs to be able to provide relevant information to meet the needs of different personnel, so as to achieve accurate information delivery and efficient utilization, and improve the efficiency and quality of system management.

[0240] On the one hand, this invention enables data preprocessing, rapid identification of fault symptoms, and local diagnostic decision-making at the edge, reducing communication burden and improving response speed; on the other hand, it enables system-level aggregation, situation display, and decision-making at the upper-level system and centralized control system, meeting the needs of operation and maintenance processes and management.

[0241] In one embodiment of the present invention, as shown in Figure 9, the timing and processing links of the system communication and data flow of the present invention are implemented according to a typical normal process, key message types and abnormal handling, which facilitates engineering implementation and fault tracking.

[0242] (1) Overall timing overview (normal operation): Data collection cycle start ( t 0): Each sensor / PLC collects raw data according to the configured cycle (e.g., 1s / 5s) and sends it to two edge controllers (Edge A / Edge B) through the field gateway; each message contains a timestamp, device ID, serial number and CRC.

[0243] Local reception and caching ( t 0+ δ 1): The edge I / O module receives UDP or Modbus frames, performs Cyclic Redundancy Checksum (CRC), eliminates anomalies, and writes them to the local circular buffer and PostgreSQL temporary table, while simultaneously sending out heartbeat events.

[0244] Preprocessing and feature extraction t 0+ δ 2): The data cleaning module processes the nearest window data in parallel, generating time-domain / frequency-domain feature vectors; missing data is filled with Kalman prediction or previous values, and is recorded and marked.

[0245] Real-time diagnosis ( t 0+ δ 3): Features are fed into the improved backpropagation neural network module to obtain the fault probability vector (classification event); low confidence results are linked to the expert rule module for secondary judgment.

[0246] Forecasting and Decision Making t 0+ δ 4): The Kalman filter provides short-term state predictions, and the FAHP module combines the diagnostic results, predicted values, and operating strategies to calculate the health level and recommended actions.

[0247] Local storage and broadcasting ( t 0+ δ 5): Diagnostic events, health decisions, and key performance indicators (KPIs) are persistently written to PostgreSQL; at the same time, short event packets (including ID, type, priority, and summary) are broadcast to the central control / large screen via UDP.

[0248] HTTP reporting ( t 0+ δ 6): The UDP→HTTP converter on the edge device side sends batch / high-priority events to the upper cloud / central control server via REST API, and uses Keep-Alive and retry queue to ensure reliable delivery; after receiving the event, the upper server confirms it and sends back confirmation response information.

[0249] Visualization and Operations Interaction ( t 0+ δ 7): The upper-level system displays alarms and supports manual confirmation. Manual feedback (confirmation / annotation) is sent back to the edge for use as a subsequent training sample.

[0250] (2) Message type and priority flow: KPI stream (low latency, periodic): sampling, local processing, local storage, batch HTTP reporting (tolerable for low latency).

[0251] Diagnostic event (high priority): Immediate UDP broadcast + synchronous HTTP reporting and wait for acknowledgment response; if no acknowledgment response is received, enter the local persistent queue and retry.

[0252] Health Decision / Work Order (Important): Write to the database via HTTP transaction and push to the work order system, triggering manual SMS / email alerts when necessary.

[0253] Model update / control command (management channel): The center issues model signature packages and pushes them to the edge via HTTPS. The edge performs integrity verification and canary deployment.

[0254] (3) Exception and fault tolerance process: Network interruption: When HTTP upload fails, data is entered into a local queue and exponential backoff is enabled; if the queue overflows, the oldest historical data is discarded according to a policy, while the latest critical events are retained. UDP broadcasts are not acknowledged, therefore critical data must also pass through the reliable HTTP channel.

[0255] Message loss and duplication: The receiving end uses message ID + sequence number to perform deduplication and reordering; fragmented messages are reassembled according to ID and sequence number, and timeout retransmission requests are initiated by the edge.

[0256] Model drift and false alarms: When the diagnostic confidence decreases or the false alarm rate exceeds the threshold, the higher-level system triggers a rollback to expert rule mode and issues an alarm; relevant samples are synchronously sent back to the cloud for retraining.

[0257] Edge fault switching: Dual-machine hot standby. When the primary edge Docker / service fails, the secondary machine takes over the IP and port to restore the data collection and broadcasting capabilities; the switching event is written to the audit log.

[0258] (4) Security and Audit: All HTTP / management channels employ Transport Layer Security (TLS) and certificate authentication; the message body contains a signature / token to prevent tampering.

[0259] Critical operations (model updates, work order triggering) require multi-level signatures and manual confirmation; the system records timestamps, serial numbers and processing times throughout the entire process, facilitating traceability and performance optimization.

[0260] (5) Timing constraints and performance objectives: End-to-end diagnostic latency target: KPI collection → local diagnosis no more than 200–500ms; HTTP reporting to the cloud no more than 2s (normal network).

[0261] Availability goals: 99.9% availability for edge services, and local queue persistence to ensure data is saved for at least 72 hours when the upstream is unavailable.

[0262] In one embodiment of the present invention, the data transmission algorithm design of the present invention adopts an adaptive fragmentation and retransmission algorithm, and the core logic is as follows.

[0263] (1) Dynamic partitioning. Assume the total data size is S (Unit: bytes), real-time network bandwidth is B (Unit: bytes / second), Calculation of optimal fragment size F: ; In the formula: α ∈[0.3,0.7]——Bandwidth adaptive coefficient (smaller value when network fluctuations are large, larger value when stable); β ∈[0.3,0.7] ——Data volume correction factor (takes the larger value when transferring large files); log2( S +1) — Used to avoid calculation errors when the data volume is 0.

[0264] (2) Transmission priority weight. Assume the data importance level is... P (1≤ P (≤5, 5 is the highest), for each increase of importance level, the transmission priority increases by 20%. Transmission priority weight. W for: ; (3) Retransmission mechanism. When the packet loss rate... ρ When the percentage is greater than 10%, exponential backoff retransmission is triggered. Assume the base retransmission interval is... T 0 (Unit: seconds), number of retransmissions is n ( n ≥1), then the first n Second retransmission interval T n for: (When packet loss rate) ρ (When >10%) The formula for calculating packet loss rate is: .

[0265] This invention discloses a three-layer visualization architecture for fault diagnosis and health assessment of marine air conditioning systems, designed to meet the stringent requirements of real-time performance, reliability, and maintainability in marine operating environments. The architecture divides the entire air conditioning system into four subsystems (air conditioning unit, air system, chiller unit, and water system), further subdivided into 14 sub-systems, covering 294 measurement variables (temperature, humidity, flow rate, pressure / differential pressure, electrical quantities, frequency, vibration, operating status, etc.). The architecture consists of the equipment edge side (first-layer visualization), the upper-level system side (second-layer visualization), and the centralized control overall side (third-layer visualization). Through layered visualization design, comprehensive monitoring, fault diagnosis, and health assessment of the marine air conditioning system are achieved. A "core calculation program" is set up at the equipment edge side (first-layer visualization), including data cleaning, operation performance fault symptom detection based on expert rules, fault diagnosis based on an improved backpropagation neural network, equipment status prediction based on Kalman filtering, and health assessment decision-making based on fuzzy hierarchical analysis. On the equipment edge side (first-level visualization), two industrial control computers provide 81 and 45 monitoring interfaces respectively, covering the main interface, operational performance, steady-state judgment, sensor and equipment fault alarms / early warnings, health decisions, historical queries and reports, etc. On the upper-level system side (second-level visualization), this is aggregated into 98 system-level interfaces. On the centralized control overall side (third-level visualization), key performance indicators, fault aggregation views, and health status are displayed on a large screen. This system balances on-site autonomy and centralized management, and communication uses a hybrid strategy of HTTP and UDP (User Datagram Protocol) to balance reliability and low latency.

[0266] This invention addresses four subsystems (14 subsystems and 294 measurement variables) of a ship's centralized air conditioning unit, air system, chiller unit, and cabin water circulation system. It proposes a three-layer distributed architecture: edge-side, upper-level system, and centralized control system. This architecture integrates real-time data acquisition, PostgreSQL / Prisma database management, HTTP-POST and UDP broadcast transmission protocols, and the Galaxy Kylin industrial control terminal. The edge-side embeds core computing programs, including data cleaning, expert rule-based fault symptom detection, improved backpropagation (BP) neural network fault diagnosis, Kalman filter state prediction, and fuzzy hierarchical analysis (FAHP)-based health decision-making, enabling local autonomous diagnosis, state prediction, and hierarchical summary visualization. This technology is suitable for monitoring, operation and maintenance decisions, and lifespan management of ship air conditioning systems under special requirements such as real-time performance, reliability, bandwidth constraints, and electromagnetic compatibility.

[0267] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0268] Obviously, those skilled in the art can make various modifications and variations to the invention without departing from the spirit and scope of the invention. Therefore, if these modifications and variations fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A visualization system for fault diagnosis and health assessment of a ship's air conditioning system, characterized in that, include: On the device edge side, it is used to acquire the real-time data stream of the original measurement variables corresponding to the measurement points of each subsystem of the ship's air conditioning system and the corresponding control parameters of the actuators issued by the ship's air conditioning automatic control system; and based on the real-time data stream of the original measurement variables, it obtains a complete algorithm dataset; through the interface rendering program on the device edge side, it displays the complete algorithm dataset on the interface on the device edge side; and based on the complete algorithm dataset, it generates a standardized JSON message and uploads it to the upper system side. On the upper-level system side, it is used to upload JSON messages based on the device edge side to obtain the aggregated timing library; the aggregated timing library is displayed on the interface of the upper-level system side through the interface rendering program on the upper-level system side. The data in the aggregated time-series database is divided into two categories and encapsulated differently: the first category, real-time KPI values, is directly encapsulated into a first UDP broadcast message; the second category, Chinese text in fault diagnosis, health decision-making, and maintenance work orders, is matched with a preset text-to-number mapping table to convert the Chinese text into corresponding numerical codes, and then encapsulated into a second UDP broadcast message; based on the first and second UDP broadcast messages, standardized UDP single data packets are generated; and the UDP broadcast push service on the upper-level system side pushes a stream of UDP broadcast data packets containing standardized UDP single data packets to the global broadcast address at a fixed period. The centralized control system receives UDP broadcast data packet streams; the UDP to HTTP data conversion and processing module on the centralized control system completes the end-to-end fault tolerance processing of the UDP broadcast data packet streams, resulting in a global structured operation and maintenance database. The global structured operation and maintenance database is displayed on the interface of the central control system through the interface rendering program of the central control system. The edge side of the device includes: The data cleaning and normalization module is used to clean and normalize the real-time data stream of the original measurement variables to obtain real-time standardized feature vectors, which are then output to the expert rule fault symptom detection module. The expert rule-based fault symptom detection module is used to obtain the corresponding real-time standardized feature vector and the corresponding symptom list based on the real-time standardized feature vector, and output them to the BP neural network module. The BP neural network module is used to generate a lightweight optimal model weight file based on historical offline datasets; and to generate a deterministic fault probability distribution vector based on real-time standardized feature vectors, corresponding symptom lists, and the lightweight optimal model weight file. The Kalman time-series state prediction module is used to obtain short-term performance prediction values, NIS anomaly identifiers, and filtered and smoothed standardized feature vectors based on the real-time data stream of the original measurement variables. The FAHP fuzzy hierarchical comprehensive health assessment module is used to obtain the health level of equipment status and the corresponding intelligent operation and maintenance strategy based on the deterministic fault probability distribution vector, short-term performance prediction value, NIS anomaly identifier and filtered and smoothed standardized feature vector.

2. The visualization system for fault diagnosis and health assessment of ship air conditioning systems as described in claim 1, characterized in that, The data cleaning and normalization module is used to process missing values, filter outliers, and denoise the original real-time data stream of measurement variables to obtain a cleaned real-time data stream. It normalizes each feature in the cleaned real-time data stream to obtain a real-time standardized feature vector and outputs it to the expert rule fault symptom detection module at the device edge.

3. The visualization system for fault diagnosis and health assessment of ship air conditioning systems as described in claim 1, characterized in that, The expert rule-based fault symptom detection module is used to retrieve the built-in expert threshold rule library specific to marine air conditioning, and match the real-time standardized feature vectors one by one with the expert threshold rule library to identify fault types in real time. Calculate the fault confidence score for each fault type and generate a symptom list. The symptom list includes: the actual fault type, the fault confidence score, and the fault trigger time. The real-time standardized feature vectors corresponding to the real-time data stream of the original measurement variables and the corresponding symptom list are output to the BP neural network module.

4. The visualization system for fault diagnosis and health assessment of ship air conditioning systems as described in claim 1, characterized in that, The BP neural network module is used to obtain a lightweight optimal model weight file based on historical offline datasets, including: Obtain historical offline datasets, which include: measurement variables of measurement points of each subsystem of the ship's air conditioning system throughout the entire cycle, with manually labeled fault and operating condition labels; Iterate through all measurement point samples in the real-time historical offline dataset, calculate the global mean and global standard deviation for each feature dimension, encapsulate the global mean and global standard deviation of all features and bind them to the Z-score normalization operation logic to generate a reusable global normalization operator; based on the global normalization operator, perform data cleaning and normalization processing on the historical offline dataset to obtain the corresponding historical normalized feature vector. Based on the equipment steady-state operating condition label or equipment transient operating condition label in the operating condition label, the historical standardized feature vector is split into three types of training samples: steady-state subset, transient load variation subset, and high and low temperature marine environment subset. The steady-state subset, transient load subset, and high- and low-temperature marine environment subset are respectively bound and paired with the corresponding manually labeled fault tags to obtain the steady-state subset, transient load subset, and high- and low-temperature marine environment subset and the corresponding manually labeled fault tags; The dimensionality of historical standardized feature vectors is used as the number of input layer nodes to construct a multi-layer feedforward neural network with two hidden layers. ReLU is used as the activation function for the hidden layers, and Softmax is used for the output layer to implement multi-fault classification. The Xavier method is used in conjunction with the input dimension to initialize the network weights, and an adaptive learning rate is preset. Preset adaptive momentum factor The L2 regularization coefficient; the built-in improved weight update calculation formula and cross-entropy loss function are used to obtain the initial neural network model with completed structure building and parameter initialization; The steady-state subset, transient variable load subset, and high / low temperature marine environment subset are globally randomly shuffled, and then divided into a training subset and a validation subset. The training subset is divided into multiple mini-batch standardized offline feature vectors according to a preset batch size, and each mini-batch is synchronously bound to a corresponding manually labeled fault label. The validation subset is stored separately and includes: validation subset features and corresponding manually labeled fault labels. Forward propagation: The number of input layer nodes in the initial neural network model is read, and it is verified that the number of input layer nodes matches the dimension of the mini-batch standardized offline feature vector. If the verification matches, the mini-batch standardized feature vector is fed into the input layer of the initial neural network model. Each component in the mini-batch standardized feature vector represents the dimension of the corresponding input neuron. Perform the first-layer matrix weighting operation, and input the result into the ReLU activation function. The ReLU function outputs the first-layer intermediate activation values. These first-layer intermediate activation values ​​are then fed into the second-layer matrix weighting operation, passing through the ReLU activation function again to generate the second-layer intermediate activation values. These second-layer intermediate activation values ​​are then fed into the output layer matrix operation, passing through the Softmax activation function to output the predicted probabilities of various fault types in a single batch. The corresponding manually labeled fault labels are retrieved, and the cross-entropy loss value for this batch is calculated jointly using the manually labeled fault labels and the predicted probabilities of various fault types in a single batch. Finally, the first-layer intermediate activation values, the second-layer intermediate activation values, the predicted probabilities of various fault types in a single batch, and the batch cross-entropy loss value are summarized. Backpropagation: Using batch cross-entropy loss as the total loss benchmark, the error of each neuron is derived layer by layer from the output layer to the input layer based on the intermediate activation values ​​of the first and second layers, combined with the deviation between the predicted probability of various faults in a single batch and the manually labeled faults. Then, by combining the weight matrices of each layer of the network, the gradients of all network parameters can be solved. ; Obtain the error of all neurons and gradients of all network parameters ; , among which, The error term of the j-th layer neuron, The symbol for partial differential represents the rate of change of a multivariate function with respect to only one variable. Loss, or batch cross-entropy loss, measures the difference between the model's fault prediction results and manually labeled results. Represents the connection weights in a neural network, representing the weight coefficients from the i-th input neuron to the j-th output neuron; This represents the input features of the i-th neuron; Improved weight update process: Based on the error of all neurons Network parameter gradient Preset iterative adaptive learning rate Let the adaptive momentum factor be set. Input characteristics of this round of forward propagation Previous iteration network weights The network weights in the previous two iterations , The network weights for this iteration are calculated using the following improved weight update formula. Iterative updates: ； The network weights in this round of iteration Write back to the initial neural network model to obtain the complete neural network with updated weights; these serve as the basic parameters for the next iteration. The iteration termination determination step involves feeding the validation subset features into the weighted complete neural network and performing forward propagation to calculate the validation loss. It then determines whether the validation loss has converged for a preset number of iterations or whether the number of iterations has reached a preset maximum value. If either condition is met, all training loops are terminated, resulting in a converged complete neural network. If not, the iteration returns to the forward propagation stage to begin a new iteration until a converged complete neural network is obtained. The converged complete neural network is then subjected to lightweight compression and encapsulation to generate a lightweight optimal model weight file. This lightweight optimal model weight file internally encapsulates the network carrier matching the feature dimensions, the weights completed in each iteration layer, the ReLU and Softmax activation function configurations, and the offline global normalization operator.

5. The visualization system for fault diagnosis and health assessment of ship air conditioning systems as described in claim 4, characterized in that, The BP neural network module is used to obtain a deterministic fault probability distribution vector based on the real-time standardized feature vector, the corresponding symptom list, and the lightweight optimal model weight file, including: Read the feature dimension values ​​of the real-time standardized feature vector and perform a consistency check with the historical sample standard dimensions built into the offline global normalization operator. If the dimensions do not match, output a dimension anomaly alarm and terminate the current inference process. If the dimensions match, call the offline global normalization operator to perform scale alignment check and scale correction on the real-time standardized feature vector to obtain the real-time standardized feature vector for inference and the corresponding feature dimension values. Read the number of network input layer nodes built into the lightweight optimal model weight file, and verify that the feature dimension value of the inference-specific real-time standardized feature vector is consistent with the number of network input layer nodes in the lightweight optimal model weight file, in conjunction with the feature dimension value corresponding to the inference-specific real-time standardized feature vector. If the verification is consistent, send the inference-specific real-time standardized feature vector into the network input layer of the lightweight optimal model weight file. Forward inference using real-time standardized feature vectors for inference: Load the weight parameters of each layer stored in the lightweight optimal model weight file, and sequentially perform matrix weighting and ReLU activation on the first layer to obtain the intermediate activation values ​​of the first layer; input the intermediate activation values ​​of the first layer into matrix weighting and ReLU activation on the second layer to obtain the intermediate activation values ​​of the second layer; send the intermediate activation values ​​of the second layer into the output layer matrix operation, and output the original predicted probability distribution vector of the real-time sample after Softmax activation operation; calculate the baseline diagnostic total confidence based on the probability dispersion and maximum predicted probability amplitude of the original predicted probability distribution vector of the real-time sample; The original predicted probabilities of all faults in the real-time sample original predicted probability distribution vector are sorted, and the two fault types with the highest probabilities are selected. These two fault types with the highest probabilities are then matched and compared with the actual fault types recorded in the symptom list. If the actual fault type in the symptom list is included in the two most probable fault types, the prediction result is considered to match the actual symptoms. The corresponding fault confidence value is used to corroborate the validity of the abnormal operating condition. The matching correction coefficient is set to 1.0, and the original predicted probability distribution vector of the real-time sample is retained. The baseline diagnostic total confidence is used and remains unchanged. If the fault type in the matched symptom list belongs to a sub-fault type within the same fault category as the two most probable fault types, the model's overall diagnosis is accurate, but the sub-faults have biases. Based on the corresponding fault confidence value, a correction coefficient is adaptively selected within the range of 1.05 to 1.

1. The higher the fault confidence, the closer the correction coefficient is to the upper limit of the range. The predicted probabilities of all sub-faults within this fault category are slightly weighted and increased. After weighting, global normalization is performed on all fault probabilities to ensure that the sum of all predicted fault probabilities is 1, resulting in an auxiliary-verified predicted probability distribution vector. Simultaneously, correction coefficients greater than 1 are adjusted by a preset small threshold to increase the baseline diagnostic total confidence of the current sample, generating an updated diagnostic total confidence. If the fault types in the symptom list do not have any major or sub-category matching relationship with the two most probable fault types, it is determined that the model prediction deviates significantly from the actual operating conditions. Based on the corresponding fault confidence value, an attenuation correction coefficient less than 1 is adaptively selected. The higher the fault confidence and the more significant the abnormal operating conditions, the greater the attenuation force. The baseline overall diagnostic confidence of this sample is lowered, an updated overall diagnostic confidence is generated, and the current diagnostic result is marked as an abnormal and questionable state. The original predicted probability distribution vector of the real-time sample is retained. Based on the intermediate activation values ​​of the first and second layers of the network generated by forward inference using the real-time standardized feature vector specifically for this inference, and the real-time standardized feature vector specifically for this inference, combined with the network weight parameters in the lightweight optimal model weight file, the gradient attribution algorithm is used to calculate the contribution ratio of various fault probabilities in the original predicted probability distribution vector of the real-time sample dimension by dimension, and the core measurement point feature set with the highest contribution ratio is selected, finally obtaining the key fault contribution features corresponding to this diagnosis; the predicted probability distribution vector after symptom list auxiliary verification, and the updated total diagnostic confidence and key fault contribution features are output; the predicted probability distribution vector after symptom list auxiliary verification includes: the original predicted probability distribution vector or the predicted probability distribution vector after auxiliary verification. Extract the fault category with the largest value from the predicted probability distribution vector after symptom list verification, and take it as the current main fault of the equipment; If the updated overall confidence level of the diagnosis is lower than the preset threshold, or if there are any abnormal or questionable markers in this diagnosis, the expert rule secondary review mechanism will be activated. Based on the identified current main fault of the device, the current main fault of the device will be reviewed and corrected in conjunction with the symptom list. The current main fault of the device and the corresponding predicted probability distribution vector after review and correction will be updated simultaneously. At the same time, the difficult time series sample carrying the updated label of the current main fault of the device will be stored in the low confidence sample cache library. If the updated overall diagnostic confidence is higher than the preset threshold, the predicted probability distribution vector after symptom list-assisted verification and the corresponding current device main fault determination result will be directly adopted. The predicted probability distribution vector, after review and correction or direct acceptance, is encapsulated to generate a deterministic fault probability distribution vector, which carries a tag for the current main fault of the equipment corresponding to this diagnosis.

6. The visualization system for fault diagnosis and health assessment of ship air conditioning systems as described in claim 5, characterized in that, The Kalman time-series state prediction module is used to construct a discrete-time state vector of the device based on each component of the real-time standardized feature vector. Among them, each component of the real-time standardized feature vector corresponds to all monitoring points of the ship's air conditioning system; Based on the real-time data stream of the original measured variables, a sensor observation vector is constructed. Based on the actuator's control parameters, the control input is constituted. ; Sampling time; Construct a discrete linear time-series state model for the equipment, which includes state equations and observation equations: ； This is the state transition matrix; To control the input matrix; This is process noise; for Real-time raw sensor observation vectors are used to correct and predict the state, while retaining the true physical dimensions. The observation matrix; To observe noise; By taking the expected value of the state equation and eliminating zero-mean process noise, the state prediction equation is obtained. By expanding the state prediction error covariance and removing irrelevant cross terms, the covariance prediction equation is obtained. The following system of time update equations is constructed simultaneously: ； Relying on historical status With control input Complete one-step prior prediction, and continuously extrapolate and iterate in multiple steps based on the prior state to generate short-term performance prediction values; In the formula, For based on The state obtained at time step The prior estimate of the state at a given time is used, and based on this formula, multiple future time series are extrapolated iteratively to obtain short-term performance predictions of key parameters such as evaporation temperature, compressor load, and pump head. for Prior covariance matrix at time step; The process noise covariance matrix; , They are respectively The posterior state estimate at time t and the posterior error covariance matrix; Perform a measurement update, using the following set of equations to complete the entire process of residual calculation, Kalman gain solution, and state and covariance correction: ； In the formula, for The innovation vector is observed at all times, representing the residual between the original measured observations and the model predicted observations; To innovate the covariance matrix; To observe the noise covariance matrix; for The optimal Kalman gain at any given time is determined by adaptively assigning confidence weights between model predictions and sensor measurements. , After the update The posterior optimal state estimate and posterior error covariance matrix at time t; It is the identity matrix; The innovation vector is obtained by solving the first two equations of the measurement update equation system. Innovation covariance matrix ,calculate : ； This is a scalar statistic, following a chi-square distribution that matches the observed dimension. A chi-square threshold is pre-set based on the confidence level. If the calculated... If the value exceeds the preset chi-square threshold range, the current measurement point is marked as abnormal and an NIS abnormality identifier is generated. After the abnormality occurs, robust processing measures such as reducing sensor weight, skipping the measurement update of this frame, and amplifying the state covariance are taken, and the sensor fault diagnosis subroutine is triggered. The state optimization update is completed using the following three equations after measurement update: Prediction bias is corrected by adaptive Kalman gain weighting to obtain a high-precision posterior state. Based on the global normalization operator, the posterior state of physical quantities is transformed into a standardized statistical scale to generate a filtered and smoothed standardized feature vector; simultaneously, the posterior covariance is updated. This provides an uncertainty benchmark for the next iteration. ； If hardware communication messages are lost, sensor observation vectors cannot be obtained. Then it cannot be solved. and It directly skips the measurement update equations and NIS discrimination calculations for this set of measurements, relying solely on the synchronously acquired control inputs. Execution time prediction; where the time prediction output is The prior error covariance matrix is ​​used to characterize the uncertainty of pure time-series inference state estimation. When a message is lost, the diagonal elements of the prior error covariance matrix are numerically amplified, indicating that there is no actual observation correction in this frame, and the reliability of state prediction decreases. The amplified prior error covariance matrix will be used as the iterative input for the next sampling period, allowing subsequent filtering to incorporate more sensor measured data to compensate for prediction drift.

7. The visualization system for fault diagnosis and health assessment of ship air conditioning systems as described in claim 6, characterized in that, The FAHP fuzzy hierarchical comprehensive health assessment module is used to extract steady-state operating parameters of the ship's air conditioning system based on the filtered and smoothed standardized feature vector, construct and quantify steady-state operating indicators, and avoid original sensor noise and jump interference. Based on the deterministic failure probability distribution vector, the occurrence probability of each failure mode is analyzed, and after normalization, a quantitative failure risk index is constructed to characterize the current failure risk level of the equipment. Based on the deviation residual between the predicted short-term performance value and the rated standard operating condition, the degree of future operational degradation of the equipment is quantified, and a performance degradation index is constructed to characterize the short-term performance degradation trend of the equipment. Based on the NIS anomaly identifier, a sensor reliability index is constructed to correct the credibility of three types of indicators: steady-state operation index, fault risk index, and performance degradation index. If sensor drift or distortion is determined, the weights of the steady-state operation index, fault risk index, and performance degradation index of the corresponding measurement point will be reduced simultaneously. Health assessment hierarchy construction: Based on the aforementioned steady-state operation indicators, failure risk indicators, performance degradation indicators, and sensor reliability indicators, a three-layer FAHP health assessment hierarchy is constructed, resulting in a complete hierarchical health assessment architecture comprising an objective layer, a criterion layer, and an indicator layer, wherein: The target layer is used for comprehensive health status assessment of ship air conditioning equipment; The criteria layer establishes four major evaluation criteria: operational stability, failure risk, performance trend, and measurement reliability. The indicator layer is used to embed the steady-state operation indicators, failure risk indicators, performance degradation indicators and sensor reliability indicators into the four evaluation criteria of the criterion layer, forming a structured, weighted, hierarchical evaluation model. Based on the experience of ship air conditioning equipment operation and maintenance experts, pairwise fuzzy comparison of importance and matrix construction were completed in two stages: Criterion layer comparison: The importance of the four criteria indicators of operational stability, failure risk, performance trend and measurement reliability are compared pairwise. The relative importance between criteria is quantified by triangular fuzzy semantics, and a fuzzy judgment matrix of criteria layer is constructed. Indicator layer comparison: For the four underlying indicators of the indicator layer, namely steady-state operation indicator, fault risk indicator, performance degradation indicator and sensor reliability indicator, the importance of each indicator is compared in pairs. The relative importance between each indicator is quantified by triangular fuzzy semantics, and a fuzzy judgment matrix of the indicator layer is constructed. For the fuzzy judgment matrix of the criterion layer and the fuzzy judgment matrix of the index layer, matrix synthesis, triangular fuzzy number defuzzification and geometric mean calculation are performed respectively to obtain the initial weights corresponding to the four criterion indices and the four underlying indices. The initial weights of the four criteria indicators and the four underlying indicators are normalized to obtain the standard weight coefficients of the four underlying indicators of steady-state operation, fault risk, performance degradation and sensor reliability, and the standard weight coefficients of the four criteria indicators. Consistency checks are performed independently on the fuzzy judgment matrix at the criterion layer and the fuzzy judgment matrix at the index layer: the maximum eigenvalue of each matrix is ​​calculated separately. Consistency index CI and consistency ratio CR; the consistency threshold is CR≤0.

1. If any fuzzy judgment matrix CR>0.1 in the fuzzy judgment matrix of the criterion layer and the fuzzy judgment matrix of the index layer, it means that there is a subjective logical contradiction in the pairwise importance scores of the corresponding fuzzy judgment matrix. Then, the expert scores of the corresponding fuzzy judgment matrix are readjusted and the matrix elements are corrected until the consistency requirements are met. After consistency verification and correction, a vector set of optimal weights for the four underlying indicators and the four criterion indicators is obtained, wherein the optimal weights are the corresponding standard weight coefficients after consistency verification. Based on the optimal weights of the four underlying indicators, the basic scores for each item are calculated by weighting: the steady-state operation indicator is weighted by the optimal weights of the underlying indicators corresponding to the steady-state operation indicator to obtain the basic score for the real-time operation stability of the equipment; the fault risk indicator is weighted by the optimal weights of the underlying indicators corresponding to the fault risk indicator to obtain the basic score for the current fault risk of the equipment; and the performance degradation indicator is weighted by the optimal weights of the underlying indicators corresponding to the performance degradation indicator to obtain the basic score for the future performance degradation trend of the equipment. The sensor reliability index is introduced to simultaneously make a secondary correction to the above three basic scores: the sensor reliability value in the sensor reliability index is used. As a conversion factor, if the sensor reliability is too low, the corresponding sub-item score conversion factor is reduced to shield the evaluation bias caused by abnormal measurement points, and the operation stability score, fault risk score and performance degradation trend score are obtained after sensor reliability conversion correction. Based on the optimal weights of the four criteria indicators, a secondary criterion weighted fusion is carried out: the corrected operational stability score, fault risk score, and performance degradation trend score, and the measurement reliability sub-score obtained by direct mapping from the sensor reliability index are multiplied by the optimal weights of the four criteria indicators of operational stability, fault risk, performance trend, and measurement reliability, respectively, and summed to finally synthesize a comprehensive equipment health index that is continuously quantified in the range of 0 to 1. Based on the preset health grading thresholds and the comprehensive health index of the equipment, and combined with the four-dimensional information linkage judgment of steady-state operation level, fault risk level, performance degradation trend and sensor reliability status, the equipment status is divided into four health levels: healthy, sub-healthy, warning and faulty. According to different health levels, corresponding intelligent operation and maintenance strategies are matched, including: routine monitoring, regular inspection, spare parts warning and shutdown maintenance. Summarize the complete algorithm dataset.

8. The visualization system for fault diagnosis and health assessment of ship air conditioning systems as described in claim 7, characterized in that, The complete algorithm dataset includes: Real-time data stream of raw measurement variables, including: equipment temperature, differential pressure, flow rate, vibration, raw sensor sampling values, equipment on / off status, and actuator feedback; synchronously generated NIS residuals, measurement point out-of-limit markers, sensor anomaly indicators, and other basic monitoring fields; The quantitative calculation results include: The basic indicators include: steady-state operation indicators, failure risk indicators, performance degradation indicators, and sensor reliability indicators; The sub-scoring includes: basic operational stability score, basic failure risk score, and performance degradation trend score; The revised sub-items include: operational stability score, failure risk score, and performance degradation trend score adjusted by sensor reliability calculation, as well as an independent measurement reliability sub-item score; The comprehensive results include: the overall health index of equipment in the range of 0 to 1, the judgment values ​​of four health levels, the predicted value of short-term performance degradation, and the labeling of intelligent operation and maintenance strategies.

Images