Method for detecting fault symptoms of marine air conditioning system based on autoencoder

Abstract

The application provides a marine air conditioning system fault symptom detection method based on an autoencoder, comprising: processing the original physical quantity to obtain a real-time original sliding window matrix; based on the real-time original sliding window matrix, based on a dynamic historical reference similar data training set under different working conditions and an autoencoder, obtaining a training converged complete floating point weight model file and a lightweight complete model weight file; based on the real-time normalized window matrix and the lightweight complete model weight file, obtaining a ship working condition label and obtaining a fault root cause positioning diagnosis result. The application is suitable for a multi-subsystem, multi-condition, high-dynamic ship air conditioning scene, can realize three-level fault symptom positioning at the system level, the subsystem level and the variable level, and can solve the adaptability, accuracy and real-time problems of the existing method in the marine scene through dynamic training sets, hierarchicalization and ship adaptation optimization.

Description

Technical Field

[0001] This invention relates to the field of marine HVAC fault detection technology, and in particular to a method for detecting fault symptoms of marine air conditioning systems based on a self-encoder. Background Technology

[0002] Marine air conditioning systems are core equipment for ensuring the living and working environment of crew members. However, their operating environment is characterized by high salt spray, strong vibration, frequent switching of operating conditions (cooling, heating, ventilation), and close coupling of subsystems (such as energy transfer between chillers, water systems, and air conditioners), which leads to a higher failure rate than land-based systems. Summary of the Invention

[0003] The purpose of this invention is to provide a method for detecting fault symptoms in marine air conditioning systems based on a self-encoder.

[0004] To address the above problems, this invention provides a method for detecting fault symptoms in marine air conditioning systems based on a self-encoder, comprising: Step 1: The industrial data acquisition card collects the raw physical quantities from sensors at various points of the marine air conditioning system, processes the raw physical quantities, and obtains the real-time raw sliding window matrix. ; Step 2, the cabin edge computing server is based on the real-time raw sliding window matrix. Construct a dynamic historical reference similar data training set and a real-time normalized window matrix for different working conditions. ; Step 3: Based on the training set of similar data from the dynamic historical reference of the different working conditions and the autoencoder, obtain the training convergence complete floating-point weight model file and the lightweight complete set of model weight files; Step 4, based on the real-time normalized window matrix Combine the lightweight complete model weight file to obtain the ship's operating condition label; Step 5: Based on the ship's operating condition markings, obtain the fault root cause location and diagnosis results; Step 6: Based on the training and convergence of the complete floating-point weight model file, obtain the dynamically adjusted inference strategy.

[0005] This invention is applicable to multi-subsystem, multi-operating-condition, and highly dynamic marine air conditioning scenarios, enabling three-level fault symptom localization: system-level, subsystem-level (equipment-level and sub-system-level), and variable-level. Through "dynamic training sets, hierarchical structure, and ship-adaptive optimization," this invention addresses the adaptability, accuracy, and real-time issues of existing methods in marine environments. Attached Figure Description

[0006] Figure 1 This is an embodiment of the overall system framework for detecting fault symptoms in a marine air conditioning system according to the present invention; Figure 2 This is a data processing flow according to an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the updated dynamic history reference similar data training set with the original fixed data training set according to one embodiment of this aspect. Figure 4 This is a schematic diagram of the autoencoder subsystem model architecture of the present invention, which mainly includes a three-layer architecture: input layer, intermediate layer, and fusion layer. Figure 5 This is a schematic diagram of the hierarchical structure of the encoder and decoder; Figure 6 This is a flowchart of the fault symptom detection logic of the present invention; Figure 7 This is a diagram illustrating the fault location logic and calculation example of the present invention; Figure 8 This is a schematic diagram of the software interface of the present invention. Figure 1 Operating curve of air conditioner No. 3; Figure 9 This is a schematic diagram of the software interface of the present invention. Figure 2 Air conditioner No. 3 alarm malfunction; Figure 10 This is a schematic diagram of the software interface of the present invention. Figure 3 Operating curve of chiller unit No. 1; Figure 11 This is a schematic diagram of the software interface of the present invention. Figure 4 Operating curves of the computer room water system; Figure 12 This is a schematic diagram of the software interface of the present invention. Figure 5 : Cabin water system operation curve. Detailed Implementation

[0007] 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.

[0008] Existing fault detection methods have the following key shortcomings: (1) Manual inspection is inefficient and delayed: Data from 10 ships of a certain ocean shipping company over the past 3 years shows that the average response time for manual inspection is 2.3 hours, and up to 5.7 hours at night or in bad sea conditions, with a missed inspection rate of 18%. The average labor cost for each manual inspection is 600 yuan, and the average annual downtime loss due to untimely inspection is 320,000 yuan. (2) Single parameter alarm false alarms and missed alarms are serious: According to the sensor data statistics of a bulk carrier over a year, the false alarm rate of single parameter alarms is 32% (such as the false alarm of instantaneous fluctuation of chilled water pressure accounts for 27%), and only 25.3% of the faults can be directly located by single parameter. (3) Poor adaptability of traditional machine learning methods: Three marine air conditioning fault detection methods were selected: Support Vector Machine (SVM), Long Short Term Memory (LSTM) network, and traditional autoencoder. A comparative experiment was conducted with the autoencoder model used in this invention. The results are shown in Table 1: Table 1 It is evident that existing methods do not take into account the multi-condition characteristics of marine systems, and the fixed, historically similar data training set cannot adapt to system aging / environmental changes, resulting in a rapid decline in accuracy.

[0009] Therefore, there is an urgent need for a detection method that can adapt to multiple operating conditions of ships, dynamically update historical reference similar data training sets, and hierarchically locate faults, in order to solve the problems of "slow response, high false alarm rate, and poor adaptability" of existing technologies.

[0010] like Figure 1 As shown, this invention provides a method for detecting fault symptoms in marine air conditioning systems based on a self-encoder, comprising: Step S1: Multi-dimensional sliding window data acquisition, encrypted transmission, and CRC32 verification of the marine air conditioning system to obtain the real-time raw sliding window matrix. The following can be used as input for step S2: the industrial data acquisition card collects the raw physical quantities of sensors at various points of the marine air conditioning system, processes the raw physical quantities, and obtains the real-time raw sliding window matrix. ; Step S1 includes: Sub-step S1-1, Real-time physical quantity acquisition of sensors: Based on the global static hardware configuration (294 variable points, sensor accuracy, 1 sampling cycle / minute), the original physical quantities of each point of the temperature sensor, pressure sensor, flow sensor and vibration sensor of the marine air conditioning system are synchronously acquired to obtain the original m=294-dimensional instantaneous sampling vector at a single moment, and transmitted to the ship's engine room PLC for storage. Here, PT100 stands for Platinum Resistance Thermometer.

[0011] Sub-step S1-2, sliding window timing splicing: The industrial data acquisition card reads the raw data of a single moment stored in the ship's engine room PLC. m =294-dimensional instantaneous sampling vector and fixed window length n =60, contiguous cache n =60 sampling points, which are then stitched together to form the original unencrypted sliding window matrix. ; Sub-step S1-3, AES-128 data packet encryption and encapsulation: The industrial data acquisition card encapsulates the raw, unencrypted sliding window matrix output from sub-step 1-2. The data packet is encapsulated using AES-128 symmetric encryption, and a CRC32 checksum is added to obtain an encrypted data packet with a checksum. This data packet is then transmitted to the edge computing server via Ethernet inside the ship. Sub-steps S1-4, CRC32 integrity verification: The cabin edge computing server performs checksum matching on encrypted data packets with checksums based on the local storage verification benchmark algorithm. If the verification code does not match, the cabin edge computing server sends a retransmission instruction to the industrial data acquisition card, returning to sub-steps 1-3 to repackage. If the checksum matches, the encrypted data packet with the checksum is decrypted to restore the original real-time sliding window matrix. .

[0012] Here, the hardware carriers for this step are the marine air conditioning system, three types of high-precision industrial sensors, industrial data acquisition cards, marine Ethernet transmission modules, and edge computing servers; all acquisition parameters, sensor accuracy, sampling rules, window length, and encryption verification rules are fixed.

[0013] For 16 equipment / subsystems in a marine air conditioning system, consisting of 8 air conditioners, 1 air system, 2 chillers, and 1 water system (including 2 chillers, 2 chilled water pumps, and 2 cooling seawater pumps), high-precision sensors (PT100 temperature sensor ±0.5℃, diffused silicon pressure sensor ±0.2% FS, and electromagnetic flowmeter ±0.5% FS) were installed at key locations. Sliding window data of 294 variables (with 60 continuous sampling points in each window) were collected at a sampling frequency of 1 time per minute. The data was transmitted to the data processing center with AES-128 encryption and CRC32 verification was used to ensure integrity.

[0014] Hardware locations and sensor configuration (all 294 measurement variables): (1) 8 centralized air conditioners (20 measurement variables per unit, totaling 160 variables), including: each air conditioner is equipped with a PT100 temperature sensor (temperature measurement accuracy ±0.5℃), an integrated temperature and humidity sensor, an electromagnetic flow meter (flow accuracy ±0.5% FS), an opening transmitter, and a current / frequency acquisition module; the 20 measurement variables are uniformly: operating mode identifier, blower start / stop status, chilled water flow rate, fan coil inlet water temperature, fan coil outlet water temperature, three-way valve opening, coil inlet air temperature and humidity, coil outlet air temperature and humidity, fresh air temperature and humidity, return air temperature and humidity, supply air temperature, blower outlet air volume, blower operating current, and blower operating frequency; (2) One set of air system (36 measurement variables), including: diffused silicon pressure sensor (pressure accuracy ±0.2% FS), air volume transmitter, to collect the external residual pressure of 8 air conditioners, the supply air temperature and humidity of each compartment / fire station, and the air volume of the main supply duct; (3) Two chiller units (16 variables per unit, 32 in total), including: vibration acceleration / velocity sensor, pressure transmitter, temperature sensor, power acquisition module, to collect compressor start / stop, current and voltage, triaxial vibration, suction and discharge pressure and temperature, cumulative running time, chilled water inlet and outlet water temperature and flow rate, and cooling water return water temperature difference; (4) Two chilled water pumps (12 variables per pump, 24 in total), including: pressure, vibration and current sensors, to collect pump start-up and shutdown, inlet and outlet water pressure, filter pressure difference, triaxial vibration acceleration and velocity; (5) Two cooling seawater pumps (10 variables per pump, 20 in total), including: pressure, vibration and current sensors, to collect pump start-up and shutdown, inlet and outlet water pressure and triaxial vibration parameters; (6) One set of cold water circulation system (22 variables), including: flow rate, temperature and pressure transmitters, to collect the flow rate of the main / branch pipe of the circulating water, the inlet water temperature of the fan coil unit in each compartment and the inlet water pressure; Here, there is one water system, including two chilled water pumps, two cooling seawater pumps, and one cold water circulation system.

[0015] The total number of variables is 160+36+32+24+20+22=294.

[0016] Sampling and sliding window rules: A uniform sampling frequency of 1 time / minute is used via industrial data acquisition cards for continuous data acquisition. =60 time-series sampling points are concatenated into one sliding window data set; single-window data matrix dimension , =60 (number of samples) =294 (total number of measured variables).

[0017] Encrypted transmission and integrity verification: The raw window data output by the acquisition card is transmitted via the ship's Ethernet fiber optic cable. Before transmission, AES-128 symmetric encryption is performed. After receiving the data packet, the edge server performs CRC32 integrity verification. If the verification code matches successfully, the raw sliding window dataset is output. If the verification fails, a retransmission command is sent to the acquisition card until the data is completely received.

[0018] Here, step S1 outputs the real-time original sliding window matrix. This can be used as a data preprocessing step in step S2.

[0019] Step S2, the cabin edge computing server (equipped with a Python data processing library) is based on the real-time raw sliding window matrix. A training set of similar dynamic historical reference data for different working conditions is constructed, which can be used as input for model training in step 3.

[0020] Sub-step 2-1, Raida 3σ outlier removal: The cabin edge computing server restores the original sliding window matrix and calculates the mean of the variables column by column. Standard deviation Remove The outliers are identified, and a coarsely cleaned matrix of missing data is obtained to fill in the missing data. ; Sub-step 2-2, linear interpolation to fill in time-series missing values: The cabin edge computing server performs coarse cleaning of the missing matrix to be filled. Linear interpolation is performed between two adjacent points in the time series to fill in the blank sampling points, resulting in a complete cleaning matrix without missing data. ; Sub-steps 2-3, min-max normalization mapping to [0,1]: The cabin edge computing server is based on a complete cleaning matrix without missing data. The real-time normalized window matrix is ​​obtained by applying the following normalization formula. : ; Gongzhong: Complete cleaning matrix without missing parts A specific element data in; for The normalized data, i.e., the normalized window matrix A specific element data in; min( () is a complete cleaning matrix without missing parts Take the minimum observation value for each column (all 60 samples of a single variable); each column has an independent minimum observation value; max( () is a complete cleaning matrix without missing parts Take the maximum observation value of each column (all 60 samples of a single variable); each column has an independent minimum observation value. Sub-steps 2-4, PCA principal component dimensionality reduction operation: cabin edge computing server, based on the normalized window matrix output from sub-step 2-3. This yields the principal component vector of the current window. ]; Sub-step 2-4-1: Calculate the normalized window matrix The vector of sample means of each variable, i.e., the row vector : ; in, For normalized window matrix No. Row vectors, including all The value of the measured variable is the first one. One time-series sampling point; Column-by-column traversal Each measurement variable, for each measurement variable all Sum of row sample values ​​and divide by the number of samples ; Generate dimension 1× row vector Each element in the vector corresponds to a window mean of a column of measured variables; Sub-step 2-4-2, based on row vectors Solve for the sample covariance matrix: ; get covariance matrix Forming a covariance matrix ; in, This is the transpose of the decentralized sample matrix; The mean matrix is ​​composed of the mean values ​​of each feature dimension; Sub-steps 2-4-3 include: Sub-step 2-4-3-1, for the covariance matrix Solving the characteristic equation yields all the results. Each feature value; will all The eigenvalues ​​are rearranged in descending order of value to obtain the complete descending sequence of eigenvalues. Simultaneously extract the feature vector corresponding to each feature value. ; Sub-step 2-4-3-2: Calculate the sum of all eigenvalues: ; Starting from the first-order eigenvalue, the cumulative contribution rate is calculated iteratively by adding values ​​sequentially. ; Cyclic Increasing Order =10, until If the cumulative contribution rate is ≥85%, stop the loop and select the top k sets of feature values ​​and feature vectors that satisfy the condition ≥85%, where k = The first k effective eigenvalue subsets are obtained. , and the subset of eigenvectors corresponding to the first k, such as the 10th order. Among them, single-branch feature vectors The dimension is ×1=294×1; Here, the normalized window matrix Principal component analysis was performed to extract the top contributors with a cumulative contribution rate ≥85%. One principal component (e.g., extract k=10 principal components from 294-dimensional data) to retain the core information of the data.

[0021] This application describes a fixed marine air conditioning system scenario: after calculating all normal sliding windows, the first condition is met. ≥85% is always present =10.294-dimensional marine air conditioning time-series window data (complete set of coupled operation data of 8 air conditioners, air, chilled water and water pumps), after a large number of historical normal samples pre-calculation, the variance distribution of the data under this working condition is stable, and the order of the standard determination in each cycle is stable, etc. 10.

[0022] Sub-step 2-4-3-3: For the real-time window link, the subset of feature vectors corresponding to k=10 is... By concatenating the columns horizontally, we obtain a reduced-dimensional two-dimensional matrix: The input to sub-step 2-4-4-1 is then fed into the projection calculation. After the calculation is complete, only the current session is cached. For offline pre-computation links, the reduced two-dimensional matrix is ​​output. As a globally unified projection matrix Persisting to a PostgreSQL database, all subsequent historical window batch dimensionality reduction sub-steps 2-5-1 will read from the database uniformly. use, = ; Here, the matrix dimension can be =294×10; With real-time temporary The construction rules and dimensions are completely consistent. = =10.

[0023] Sub-step 2-4-4 includes: Sub-step 2-4-4-1: Calculate the principal component subspace score matrix. ; Sub-step 2-4-4-2: Reconstruct the approximate score matrix based on the principal component subspace score matrix. = ; Sub-step 2-4-4-3: Solve for the residual principal space matrix. = Based on the residual principal space matrix The mean square error of the global residuals of the window is obtained. , in, This represents the total number of time-series sampling points within the sliding window. This represents the total dimension of the original sensor features; For the window number The time series sample, the first The reconstruction residual of each measuring point; this index is the two-dimensional joint average error of the entire window time sequence plus all measuring points, used to verify the overall reconstruction effect of the window.

[0024] The residual matrix of the first Line number Column element, representing the first The sampling point, the first Reconstruction bias of each measurement variable; Sub-step 2-4-4-4, if Then proceed to sub-step 2-4-4-5 to extract the principal component vector of the current window. Among them, the global static residual threshold = 0.03; if Then, execution will restart from sub-step 2-1 until... .

[0025] Here, the residual principal space matrix For the residual master space, auxiliary verification of reconstruction error is performed within this stage; Sub-step 2-4-4-5: Obtain the principal component subspace score matrix. The last row (row 60, the complete principal component score at the latest window time) is used as the global performance representation of the window. The numerical combinations of the principal components form a one-dimensional row vector: The current window principal component vector is directly fed into sub-steps 2-5 for PCSF similarity calculation with historical window principal component vectors; the real-time original sliding window matrix that has passed sub-steps 1-4 is then used for the calculation. Normalized window matrix Assign the same unique, non-duplicate history window save number The real-time original sliding window matrix that has passed the verification in sub-steps 1-4 The archive is named Historical Primitive Matrix The normalized window matrix that passed the verification in sub-steps 1-4 The archive is named the History Normalization Matrix The archive is synchronously written with dual time tags, including: the current system time as the archive creation time tag, and the most recent PCSF matching time tag, which is initially blank; the two sets of dual time tags and the history normalization matrix of sub-step 2-4-4-5 are then combined. The operating condition identifier (built into the historical window, 1 for cooling, 2 for heating, and 3 for ventilation) and the principal component vector of the current main window are stored together in the historical reference similar data training set data table of the PostgreSQL database.

[0026] Store the following set of related data in a unified historical reference similar data training set table in a PostgreSQL database: 1) Historical Original Matrix (The real-time raw sliding window matrix constructed by reading PLC data starting from step S2) (Obtained by renaming the archive) 2) Historical Normalization Matrix (Output from sub-step 2-3) (Renamed) 3) This window contains built-in operating condition indicators (1 represents cooling, 2 represents heating, and 3 represents ventilation, which change synchronously with the original measurement points of the PLC). 4) Principal component vector of the current window After being stored in the data table, it serves as the historical principal component vector corresponding to this historical sample. ; 5) Dual timing tags.

[0027] Here, .

[0028] Sub-steps 2-5, based on the global unified projection matrix The PCSF similarity matching records were obtained, including: Sub-step 2-5-1: The cabin edge computing server reads the global uniform projection matrix from the PostgreSQL database. Obtain the set of all valid historical normalized matrices { For each historical normalized matrix Perform dimensionality reduction: Take the principal component score matrix for each historical record. The last line (the (row) refers to the row retrieval operation, where the row contains =10 principal component values, The principal component values ​​are concatenated into a one-dimensional row vector: ], Traverse all history windows For each Perform the above row retrieval operation on all rows to generate the corresponding rows. After summarizing all of them, we get the historical principal component vector set { }

[0029] Here, This is a unique, non-repeating archive number for each history window. This represents historical archived samples stored in the database, distinguishing them from real-time temporary variables. The superscript ' indicates that the result has undergone min-max normalization; the subscript... This indicates the unique archive number for this history window; Sub-step 2-5-2: The cabin edge computing server calculates the principal component similarity factor group by group using the cosine similarity formula. Bind each principal component similarity factor The corresponding historical window contains the original data, operating conditions, and time tags; A unique, non-repeating archive number for each history window; ; Sub-steps 2-5-3 involve the cabin edge computing server filtering and retaining data that meets the requirements. A historical window of 0.85, based on principal component similarity factor. Sort in descending order and extract the top items with the highest similarity in the descending order. A subset of historical windows is selected to obtain a highly similar subset of historical windows with condition identifiers and time tags. For example: =10. The current real-time window is compared with the highest similarity captured... The pairing relationships, similarity scores, and archive creation timestamps of each historical window are used as PCSF similarity matching records, written to the database, and the highest similarity records extracted in that batch are updated synchronously. The most recent PCSF match time stamp corresponding to each history window; where the archive creation time stamp and the most recent PCSF match time stamp are used as dual time stamps.

[0030] Sub-steps 2-6: The cabin edge computing server groups the highly similar historical windows in the PCSF similarity matching records that have built-in operating condition identifiers (1 for cooling, 2 for heating, and 3 for ventilation) to ensure that the data within each group is uniformly distributed. The output is an uncleaned dataset of similar operating conditions. The operating condition identifiers include: cooling, heating, and ventilation; the dual timing tags include: the archive creation time tag and the most recent PCSF matching time tag. Sub-steps 2-7 set the criteria for determining a sample as redundant if it simultaneously meets the conditions of storage duration exceeding 6 months and having no PCSF similarity matching records within the last 30 days; and uses the unique archive number of the historical window within the uncleaned functional similarity dataset. As an index, it retrieves historical reference similar data training sets from the PostgreSQL database (the historical reference similar data training set data table in PostgreSQL, with a unique archive number). The data is stored using two time tags (archive creation time tag and most recent PCSF match time tag) for a complete storage record (primarily a primary key). Based on the two time tags, the sample storage duration and the most recent match interval are calculated. All storage records in the historical reference similar data training set, distinguished by archive number t, are traversed, and those that simultaneously meet the conditions for being judged as redundant samples are selected. , And supporting operating conditions, time series and principal component vector data Batch delete the filtered items , And supporting operating conditions, time series and principal component vector data ; Synchronously update the global historical normalized matrix set { Remove deleted entries. ; Based on the updated global history normalized matrix set { All valid historical window samples corresponding to} are re-collected and grouped according to the working condition identifier bound to each sample, and a non-redundant dynamic historical reference similar data training set for each working condition is reconstructed.

[0031] L: Reconstruction loss for a single static vector / single-time sample, denominator m (number of features), not involving the time window n; E MSE : Complete sliding window two-dimensional global error, denominator n·m (time series points × features), which includes both time and measurement point dimensions; Offline global RMSE global : Denominator K (total number of samples in all independent windows), K is independent of window parameters n and m.

[0032] Step 3, Hierarchical autoencoder network construction: Based on the training set of similar data with dynamic historical reference for different working conditions and the autoencoder, the training convergence complete floating-point weight model file and the lightweight complete set of model weight files are obtained; Sub-step 3-1: Initialize and build the two-level autoencoder network structure to obtain the untrained network skeleton: A lightweight global autoencoder (encoder + symmetric decoder, parameters ≤ 1000) is built on the cabin edge computing server with 28-dimensional global core variable constraints. Based on the 16 devices and the three operating conditions (cooling, heating, and ventilation), an independent sub-encoder is built for each device, generating a total of 48 sub-models of devices and corresponding operating conditions. The mathematical form of the encoder is... The mathematical form of the decoder is Network weight With bias Random initialization; this stage only completes the structure building and does not perform iterative updates. The mathematical description of the encoder is: ,in, This is the encoded feature vector; For activation functions; This is the weight matrix; The historical principal component vector for sub-step 2-4-4-5 During historical sample inference, the historical principal component vectors retrieved from the database are used. Also serves as model input ; This is the bias vector.

[0033] The mathematical description of the decoder is as follows: ; in, for The reconstructed output data; For activation functions; This is the weight matrix; This is the bias vector.

[0034] To address the issues of "multi-subsystem coupling + multi-condition adaptation," a three-level model of "global, subsystem, and device" is designed: 1) Global Model: 28 core variables across subsystems (reflecting the energy interaction of water, wind, and cold source) are selected, and a lightweight structure (parameters ≤ 1000) is adopted to achieve rapid screening of system-level anomalies; 2) Subsystem and equipment models: 16 equipment / subsystems, 48 ​​"equipment and operating condition models" are constructed according to 3 operating conditions (cooling, heating, ventilation). The full variables of the corresponding equipment are input, the "normal" distribution is narrowed (the variance of the training set of similar historical reference data is reduced by 40%), and the sensitivity of small anomalies is improved. 3) Model compression and training: 8-bit quantization + bias removal reduces parameters by 50%, training under different conditions + weighted loss ensures error balance in each subsystem, and Adam optimizer + early stopping strategy accelerates convergence.

[0035] Autoencoder model design: Each autoencoder model contains an encoder and a decoder. The encoder consists of multiple fully connected layers or convolutional layers, which extract features and reduce dimensions of the input data (the input data of a general model, such as the historical reference similar data training set during model training, and the current sliding window data during fault detection). The decoder has a structure symmetrical to the encoder and reconstructs the dimensionality-reduced features to restore the input data.

[0036] Sub-step 3-2, Training the Model for Different Working Conditions (Weighted MSE + Adam + Early Stop): The cabin edge computing server receives the untrained network skeleton output from sub-step 3-1 and the dynamic historical reference similar dataset for different working conditions obtained from sub-step 2-7. First, the training set is split into s (e.g., 4 subsystems), r (e.g., 16 devices), and 3 types of working conditions, where each subsystem includes multiple devices. For each type of device, the corresponding working condition samples are independently divided into hierarchical subsets according to a preset ratio (e.g., 8:2), generating training and validation subsets for each working condition. A 1-bit working condition identifier is appended to all samples in both the training and validation subsets. The training subset is input into the network of the current iteration stage for forward propagation. The first iteration uses the untrained original network structure skeleton output from sub-step 3-1. Subsequent iterations use the network with updated parameters optimized by Adam, employing the basic mean square error as the single-sample reconstruction loss. Calculate the reconstruction error, where the reconstruction error is the total weighted reconstruction loss after introducing the weights of each subsystem item. Total weighted reconstruction loss The network parameters are updated via backpropagation using the target value through the Adam optimizer; the encoder parameters are iteratively updated via backpropagation using the Adam optimizer based on the training subset loss. and Decoder parameters and In each iteration, the validation loss is calculated using a validation subset. An early stopping strategy is enabled to monitor changes in the validation set loss. If the validation loss does not decrease after 5 consecutive preset rounds, the model is considered to have converged and the iteration is stopped, resulting in a fully converged floating-point weighted model file. After training, the variance of the normal sample distribution of each sub-model is reduced by 40%, narrowing the normal data distribution range and improving the sensitivity of identifying minor anomalies. The feature dimension is that of a single principal component vector; this loss is calculated only for a single static principal component sample and does not include the sliding window time dimension n, which is different from the online window global error E. MSE .

[0037] Here, the marine air conditioning system is divided into s subsystems, each of which contains one or more air conditioning-related operating devices; each device independently corresponds to three operating conditions: cooling, heating, and ventilation, forming r×3 sets of equipment-operating condition-specific sub-models.

[0038] Input the training subset samples from all normal operating conditions into the converged global model in the fully converged floating-point weighted model file to calculate the global RMSE in batches and obtain the mean global normal RMSE. Standard deviation And both are persistently stored in the PostgreSQL database as global static threshold parameters used in the first-level global anomaly coarse judgment inference step 4. Normal operating condition refers to a steady-state operation in which, during the ship's navigation, all sensor measurement points, equipment operation coupling relationships, and self-encoder reconstruction errors fall within the normal baseline range of offline statistics for all subsystems and individual equipment of the air conditioning system under the design specifications, heating, and ventilation modes, without exceeding limits, deviations, or cross-system coupling disturbances, and the equipment operates smoothly without any abnormal alarms.

[0039] ; ; ; .

[0040] For each of the 48 equipment-operating condition sub-models, the following steps were performed: Extract all normal training subset samples for the corresponding equipment and operating condition; input these samples into the converged sub-models for the corresponding equipment and operating condition from the fully converged floating-point weighted model file; calculate the reconstruction RMSE for each individual equipment in batches; and statistically analyze the mean normal RMSE for each sub-model specific to each equipment and operating condition. Standard deviation ; will be compared with the average RMSE of 48 normal groups Standard deviation The data is persistently stored in a PostgreSQL database as the top-level static threshold parameter for the secondary device fine-tuning in step 4.

[0041] For each sensor's original measurement point time series, statistical analysis was performed separately on all normal operating condition samples within the training subset where all operating conditions were normal: traversing the variables of each sensor measurement point. Extract the time-series observation set of the measurement point under all normal operating conditions training samples, and calculate the time-series mean of the variable of the sensor measurement point. and time series standard deviation .

[0042] Here, the device's unique m-dimensional measurement point vector is input, and forward reconstruction is completed through the sub-encoder and sub-decoder. Calculate the mean squared error of single-sample reconstruction: ; ; Subscript Representing the first in this group One normal sample; K represents the total number of all normal training samples within the current equipment's operating condition group. ; The weighted loss weights are 0.3, 0.2, 0.3, and 0.2. The total weighted reconstruction loss after introducing the component weights is: The Adam optimizer has a learning rate of 0.001, a maximum of 20 iterations, and an early stopping criterion (no decrease in loss for 5 consecutive iterations). Training converges the complete floating-point weighted model file (global model floating-point parameters + 48 sets of equipment-operating condition sub-model floating-point parameters, including optimized parameters). and , and ).

[0043] The total training subset is first fed into the global model to complete the learning of general features for the entire system; then the training subset is split according to subsystem-equipment-operating condition, and each equipment operating condition sub-model is trained separately. After the global model is trained, the mean global normal RMSE is calculated. Standard deviation This is used for the first-level global coarse anomaly detection in step 4; After each matching sub-model is trained, the average normal RMSE value for that specific device is calculated separately. Standard deviation This is used for precise and detailed inspection of the secondary equipment in step 4; The global model weights and all r×3 sets of matching sub-model weights are packaged together and stored in the same lightweight complete model weight file, and then stored in the database.

[0044] For each sensor's original measurement point time series, statistical analysis was performed separately on all normal operating condition samples within the training subset where all operating conditions were normal: traversing the variables of each sensor measurement point. Extract the time-series observation set of the measurement point under all normal operating conditions training samples, and calculate the time-series mean of the variable of the sensor measurement point. and time series standard deviation ; Here, model training involves classifying the data from the entire historical reference similarity training set into four subsystems (containing 16 devices or sub-systems) and three operating conditions (cooling, heating, and ventilation), and inputting each subsystem into its corresponding sub-model for training. Model parameters are optimized by minimizing reconstruction errors (such as mean squared error), encompassing all learnable variables in the model, enabling it to accurately reconstruct data under normal operating conditions. Model parameters determine the mapping relationship between input data and reconstructed data, including: weights (representing the connection strength between neurons in each layer of the neural network, directly affecting feature extraction and reconstruction results), biases (constant terms added to each neuron to adjust the neuron's activation threshold and enhance the model's ability to fit the data), etc.

[0045] The formula for calculating the overall reconstruction loss (using mean squared error) is as follows: ; L is the loss value, used to measure the deviation between the model's prediction / reconstruction results and the actual original data. The smaller L is, the better the fitting / reconstruction effect. This represents the total number of samples used in the calculation. For the one-dimensional row vector in sub-step 2-4-4-5: In ; for Reconstructed data; reference formula .

[0046] Sub-step 3-3: The cabin edge computing server traverses all network weight connections in the training converged complete floating-point weight model file, deletes weight connections with an absolute value less than 0.01, completes weight pruning, and obtains the pruned sparse floating-point model file. In sub-steps 3-4, the cabin edge computing server linearly maps the 32-bit floating-point weights in the pruned sparse floating-point model file to the 8-bit signed integer range [-128, 127] to complete the quantization, while simultaneously preserving the encoder network weights. and decoder network weights Map the correspondence and output the intermediate quantization model file; Sub-steps 3-5: The cabin edge computing server removes the encoder bias of the quantized intermediate model. and decoder bias During the inference phase, global static compensation values ​​are used to offset the reconstruction bias caused by the missing bias, resulting in a quantized and debiased intermediate model file. All weights of the model are traversed to verify that the sparse structure is not lost, and finally a lightweight complete set of model weight files are generated and stored in a PostgreSQL database as input for real-time inference in step 4.

[0047] Step 4: Layered triggered RMSE reconstruction error calculation, two-level initial fault judgment, and cross-system coupled anomaly identification: based on real-time normalized window matrix. Combine the lightweight complete model weight file to obtain the ship's operating condition label; Sub-step 4-1, Level 1 trigger, global model 60s cycle inference, calculate global... : The cabin edge computing server (equipped with the TensorFlow Lite inference library) obtains the real-time normalized window matrix output from sub-steps 2-3. The system uses real-time normalized data of the 28-dimensional core variables; extracts global lightweight model weights from the lightweight full set of model weights output in sub-steps 3-5; reads the input 28-dimensional core variable real-time normalized data at a fixed interval of 60 seconds, loads the input global lightweight model weights, and calls the global autoencoder to complete the encoding-decoding forward inference reconstruction; and utilizes the root mean square error formula. The global reconstruction error for the current period is calculated. Simultaneously, read the global normal RMSE mean value pre-stored in the PostgreSQL database in sub-step 3-2. Standard deviation Through the formula: , Calculate the global decision threshold ,Will and Simultaneously transmitted to subsequent judgment stages; Sub-step 4-2, Global anomaly threshold determination branch calculation (first-level triggered branch routing): Execution entity: Cabin edge computing server, comparison and Triggering instructions distinguish inference cycles: like If the system is found to be abnormal, a global abnormality flag is output, all 16 devices corresponding to the sub-model are triggered synchronously, and a 10-second short-cycle secondary fine inspection trigger command is initiated, with data flowing into sub-step 4-3. like If the system is deemed to be normal overall, a global normal status flag is output, and device sub-model reasoning is executed in parallel for each device. Then, sub-step 4-3 is entered to determine local minor anomalies. Sub-step 4-3, secondary triggering, sub-models of each device. Calculation and local anomaly localization: Read the average normal RMSE value corresponding to the current device under test from the input. and standard deviation Combined with the given The calculation formula calculates the device's specific anomaly detection threshold in real time. Based on the trigger command issued in sub-step 4-2, the inference cycle is differentiated. For scenarios with global anomaly marking, a 10-second short-cycle fine-grained detection is enabled; for scenarios with system-wide normal marking, a 120-second normal cycle is used. Lightweight weights of the device condition sub-model matching the current device and real-time operating conditions are loaded one by one from the lightweight full set of model weight files in sub-step 3-5. The full real-time normalized variables for a single device (from the real-time normalized window matrix in sub-step 2-3) are input. Complete encoding and decoding reconstruction, and calculate the real-time reconstruction error of a single device. The real-time reconstruction error calculated for each device Compared with the pre-calculated specific anomaly detection threshold By comparing them one by one, the set of abnormal subsystem numbers and / or abnormal device numbers that have reconstruction errors exceeding the threshold are selected. Sub-step 4-4: The cabin edge computing server outputs a real-time normalized window matrix from sub-step 2-3 of step 2. In the process, real-time time-series data of 28 global core variables are read; two sets of time-series measurement points, chilled water flow rate and supply air temperature, are extracted from the 28 global core variable time-series data to construct a coupling coefficient matrix, and the cross-system correlation coupling coefficient is calculated in real time; the calculated cross-system coupling coefficient is compared with the normal reference range of the chilled water flow rate and supply air temperature coupling coefficients pre-stored in the database, such as 0.8~1.2. If the cross-system coupling coefficient deviates from the normal reference range, it is determined that there is a cross-system correlation anomaly that cannot be identified by the traditional single-device model, and the corresponding cross-system coupling anomaly mark is generated; the cross-system coupling anomaly mark is summarized and merged with the set of abnormal subsystem numbers and / or abnormal device numbers output by sub-step 4-3, and uniformly sent to the top level of step 5 as input information for fault root cause location.

[0048] Sub-steps 4-5 integrate all abnormal states from the first three steps to generate a unified operating condition label: If sub-step 4-2 obtains a system-wide normal label and there is no system-wide abnormal label from sub-step 4-2, and no cross-system coupling abnormal label from sub-step 4-4, then a stable navigation label is generated; if sub-step 4-2 obtains a system-wide abnormal label, then it carries the abnormal subsystem number and / or abnormal equipment number obtained from sub-step 4-3; if sub-step 4-4 obtains a cross-system associated abnormality, then it carries the cross-system coupling abnormal label; the stable navigation label, abnormal subsystem number, abnormal equipment number, or cross-system coupling abnormal label are packaged into a ship operating condition label. Here, the ship's operating condition flag serves as the operating condition input for sub-step 6-4, dynamic resource scheduling of computing power.

[0049] To address the balance between real-time performance and positioning accuracy, a three-level triggering logic is designed: Level 1 trigger: The global model performs a 60-second periodic detection; if the error exceeds the threshold, the sub-model is triggered. Level 2 trigger: Subsystem / equipment model performs a detailed inspection every 10 seconds to locate abnormal equipment; Level 3 triggering: Variable-level Jacobian sensitivity analysis, combined with physical weights to locate the root cause of the fault; Cross-system anomaly identification: By using the "chilled water flow rate and supply air temperature coupling coefficient matrix", "correlated anomalies" that cannot be identified by traditional methods (such as the chiller unit is normal but the water system distribution is faulty) can be captured.

[0050] Step 5: Jacobi standardized sensitivity calculation + physical weight correction, screening of fault root cause variables: Based on ship operating condition marking, the fault root cause location diagnosis results are obtained; Sub-step 5-1, based on the set of abnormal subsystem numbers and abnormal device numbers output from sub-step 4-3, and the cross-system coupling anomaly markers output from sub-step 4-4, identifies the abnormal subsystems and abnormal devices to be analyzed; and outputs the real-time normalized window matrix from sub-step 2-3. Extract all measured variable observations from the currently locked anomaly subsystem to be analyzed, and construct a real-time anomaly sample vector. , This represents the total number of variables in this subsystem; read the real-time anomaly sample vector of this anomaly subsystem. The input is the sub-model lightweight weights extracted from the lightweight full set of model weights file in step 3, which match the operating conditions corresponding to the abnormal equipment. This is used to complete the forward inference and generate the autoencoder reconstruction output vector. ; Sub-step 5-2 utilizes real-time anomaly sample vectors Reconstructing the output vector with the autoencoder The overall subsystem reconstruction loss is calculated using the following formula to complete the initial screening of the univariate original residuals and obtain the reconstruction loss. : ; This represents the total number of sensor measurement points in this subsystem; this formula is the average reconstruction error of all measurement points in the subsystem under a single time series sample, and does not involve the time series window dimension. n .

[0051] Here, standardized residuals are calculated and initial screening is performed. For each variable in this subsystem within the outlier samples, the error between its original value and the autoencoder reconstructed value is calculated individually, reflecting the degree of deviation of a single variable. Autoencoder reconstruction error: .

[0052] This represents the total number of sensor measurement points included in this subsystem; this formula only calculates the average of all measurement points sampled in a single time frame at a fixed moment, without traversing the time window. n , with window global E MSE The computational dimensions are completely independent.

[0053] Sub-step 5-3 uses the reconstruction loss obtained in step 5-2. Using the objective function, backpropagation is performed based on the lightweight sub-model weights (i.e., the complete network parameters of the sub-model) extracted from the lightweight full set of model weights output in sub-steps 3-5 and matched with the corresponding operating conditions of the abnormal equipment. Solve for the partial derivatives of the decoder output with respect to the input variables. (Obtained through backpropagation from the decoder); combined with the univariate residuals, the Jacobian matrix elements are calculated variable-wise to quantify the original sensitivity contribution of univariate variables to the overall reconstruction error. Specifically, for a single sample, the Jacobian matrix... J The dimension is 1× m , where the Jacobian gradient of the variable J 1,i Defined as: ; in, It is the output of the autoencoder decoder. For input The partial derivatives need to be calculated through backpropagation (using the network parameters of the autoencoder). Their physical meaning is: The larger the absolute value, the stronger the variable. Small fluctuations can lead to larger changes in the reconstruction error, meaning that the variable is more "sensitive" to the current anomaly.

[0054] Here, taking into account on-site engineering constraints (such as limited computing power of on-site embedded devices), if a subsystem malfunctions, a fine-grained variable contribution analysis (Jacobian sensitivity contribution decomposition) is performed to identify specific devices / variables. For the marine air conditioning subsystem... Variables Its reconstructed value is The reconstruction error is... L ( , By calculating L for each gradient Quantifiable variables Sensitivity to errors.

[0055] Sub-step 5-4: Jacobian gradient for all variables J 1,i Standardize the absolute value to eliminate dimensional differences, and calculate the basic standardization sensitivity. : ; Here, since the gradient values ​​of different variables may not be directly comparable due to differences in units (such as temperature in °C, pressure in kPa) or range of values, standardized sensitivity processing is required: ; in, The standardized variable sensitivity (contribution) satisfies . The larger the value, the higher the relative contribution of the variable to the reconstruction error. Sub-step 5-5: Read the physical importance weights of the global variables pre-stored in the database. The absolute value of the original Jacobian gradient used in sub-step 5-4 is calculated. Using the base case, substitute into the formula to calculate the weighted correction contribution. Complete the correction of the physical importance weights of the equipment: ; Among them, the physical importance weights of global variables pre-stored in the database , , The data is based on industry expert experience and historical failure frequency, and is stored offline. Here, based on the physical processes of marine air conditioning systems, key variables (such as total air volume at the blower outlet, cooling coil outlet air temperature, and cooling coil inlet water temperature) are given higher weights. Correct contribution level: .

[0056] Weight It can be determined through expert experience (such as the degree of influence of the variable on system safety) or historical fault data (such as the frequency of faults caused by the abnormality of the variable).

[0057] Sub-steps 5-6 involve adjusting the weighted contribution. Sort by largest to smallest, and read a fixed threshold. Filter to meet These variables are considered as highly sensitive suspected fault variables, and they are the main suspected factors leading to subsystem anomalies. Sub-steps 5-7 involve reading the pre-stored normal data distribution characteristics of this subsystem in the input database (the time-series mean of the corresponding sensor measurement point variables stored in PostgreSQL after training convergence in sub-step 3-2). and time series standard deviation The high-sensitivity suspected fault variables obtained from sub-steps 5-6 are used as the analysis objects. Combined with cross-system coupling anomaly markers (the set of anomalous subsystem numbers and anomalous device numbers output from sub-steps 4-3, and the cross-system coupling anomaly markers output from sub-steps 4-4), fault types are distinguished. Specifically, if the real-time observed values ​​of the high-sensitivity suspected fault variables... Deviating from the normal range If the cross-system coupling anomaly of a high-sensitivity suspected fault variable is marked as "present", it is determined to be a cross-system associated fault. , This refers to the i-th sensor measurement point variable statistically entered into the database from the training subset based on working conditions during step 3-2. Time series mean and time series standard deviation .

[0058] , ; N is the total number of samples in the entire normal training subset; For the t-th normal training sample, the th Observed values ​​from each sensor measurement point; For the first Average of the normal operating time sequence of each measuring point; For the first Standard deviation of the normal operating time sequence of each measuring point; Here, sub-steps 5-7 output the fault root cause location diagnosis results, including: high-sensitivity suspicious variable number, corresponding abnormal equipment number, abnormal subsystem number; and fault type label: local single equipment fault / chilled water flow rate - supply air temperature cross-system coupling correlation anomaly.

[0059] The normal data distribution characteristics of the subsystem (such as mean, variance, or normal pattern parameters learned by the autoencoder).

[0060] Step 6, Dynamic scheduling of ship embedded computing power + bidirectional integration of Modbus RTU and ship PLC: Based on training and converging complete floating-point weight model file, a dynamic adjustment inference strategy is obtained; Sub-step 6-1, Model Structure Optimization: Based on the baseline complete autoencoder network structure (3-layer encoder + 3-layer decoder) in the training convergence complete floating-point weight model file of sub-step 3-2, under the premise of ensuring that the fault detection accuracy is not lower than the preset accuracy constraint threshold, the number of encoder and decoder network layers and the size of neurons are reduced, simplifying the 3-layer encoder + 3-layer decoder to 2-layer encoder + 2-layer decoder, and outputting a lightweight basic network structure. Among them, the accuracy constraint threshold (the minimum allowable detection accuracy lower limit) is a static constraint configuration parameter preset offline for the needs of ship on-site engineering.

[0061] Sub-step 6-2, model compression techniques (pruning + quantization), includes: Sub-step 6-2-1, Structured Pruning: Load the lightweight basic network structure and matching FP32 raw floating-point weights from sub-step 6-1, use the matching FP32 raw floating-point weights to forward infer batch samples, and calculate the contribution of each network connection to the reconstruction error; compare the contribution of each connection with the reconstruction error sensitivity threshold, and remove neurons and network connections with contributions below the threshold; simultaneously retain the corresponding FP32 weights for the retained network structure, delete the weight parameters of the pruned connections, and obtain the sparse network and corresponding FP32 weights after pruning; Here, the FP32 floating-point original model weights trained in step 3 refer to the full-precision original weights obtained after the model training in sub-step 3-2 converges, which match the lightweight basic network structure in sub-step 6-1. The reconstruction error sensitivity threshold is a preset static parameter in offline engineering, used to determine which neuron connections can be safely pruned without compromising detection accuracy. The TensorFlow Lite compression toolchain is a low-level tool for model compression and inference deployment that comes pre-installed on embedded inference endpoints.

[0062] Model structure optimization: Reduce the number of network layers and neurons in the autoencoder model (e.g., 3 layers for the encoder and 3 layers for the decoder can be simplified to 2 layers for the encoder and 2 layers for the decoder) to reduce the complexity of the model while ensuring detection accuracy.

[0063] Sub-step 6-2-2, Quantization Compression: Import the pruned sparse network structure and corresponding FP32 weights into the TensorFlow Lite offline conversion toolchain; perform quantization mapping calculation based on the global numerical distribution of FP32 weights to obtain the scaling factor and zero offset required for converting floating-point to fixed-point or integer, as the mapping calculation result; based on the mapping calculation result, convert all valid 32-bit floating-point weights into 16-bit fixed-point numbers or 8-bit integers to generate quantized integer weight parameters; during the conversion process, retain the sparse network connection relationship obtained from the pruning in the input, quantize only the valid connection parameters, and remove the pruned invalid connections; combine the pruned sparse network from sub-step 6-2-1 with the quantized integer weight parameters obtained in this step to generate a complete compressed model file; Here, model compression techniques are employed, such as quantization compression (e.g., converting model parameters from 32-bit floating-point numbers to 16-bit fixed-point numbers or 8-bit integers) and structured pruning (e.g., removing neurons or connections that have minimal impact on reconstruction errors), to reduce model storage space and computational complexity, thereby improving model execution speed. Using TensorFlowLite pruning and quantization, inference time was reduced from 0.52 seconds to 0.12 seconds.

[0064] Here, sub-steps 3-5 use a lightweight complete model (3 encoders and 3 decoders) for fault diagnosis, generating operating condition markings and fault root cause location results (dedicated to steps 4 and 5). Sub-step 6-2, the complete compressed model (2 encoders and 2 decoders), is specifically for embedded performance testing. It only provides inference time metrics and assists in sub-step 6-4 in dynamically scheduling the inference strategy; it does not perform fault diagnosis. The underlying model used for subsequent actual fault inference is still the lightweight full set of weight files output from sub-step 3-5.

[0065] Sub-step 6-3 calls the TF Lite embedded inference runtime library in the TensorFlow Lite compression toolchain to load the complete compressed model file from sub-step 6-2-2, completes the model deployment in the ship's embedded hardware, and obtains the deployed lightweight inference model; inputs a unified standard test sample into the deployed lightweight inference model, performs multiple rounds of forward inference, and collects the time consumption using the built-in timing function of the TF Lite embedded inference runtime library, and calculates the actual compressed inference time of the compressed model, such as 0.12s; The baseline time of the original FP32 model is read, such as 0.52s. The measured inference time after compression is compared with the baseline time to obtain the comparison index of the model inference time per sample before and after model compression (original 0.52s, compressed 0.12s). Here, the unified standard test sample is a standardized benchmark sample set that is separately extracted from the dynamic historical reference similar data training set of step 2, retained independently, and not used in the model training and verification of sub-step 3-2. It is not used in network parameter training or threshold statistics throughout the process, but only used in sub-step 6-3 to evaluate the inference time and accuracy of the 2-layer encoding and decoding small compression model.

[0066] The samples meet the three-class coverage requirements: 1) Full coverage of operating conditions: Includes three types of operating conditions: cooling, heating, and ventilation; 2) Full equipment coverage: The coverage system includes all r marine air conditioning units; 3) Full fault coverage: It includes stable and normal samples, samples with minor local faults, and samples with abnormal cross-system coupling.

[0067] Sub-step 6-4 calls the PLC real-time computing power monitoring value (current computing power load data collected in real time by embedded hardware), the model inference time comparison index (speed optimization index output from sub-step 6-3), and the ship operating condition marker (from sub-step 4-5), and dynamically adjusts the inference strategy in conjunction with the complete compressed model file (output from sub-step 6-2-2): If the ship's operating condition is marked as stable navigation, based on the comparison index of model inference time, a scheduling instruction is issued to extend the global loop detection cycle of sub-step 4-1 and reduce the execution frequency of the complete compressed model file inference in sub-step 6-2-2; If the ship's operating condition flag carries an abnormal subsystem number (with an accompanying cross-system coupling abnormal flag), only the lightweight model of the corresponding subsystem in the complete compressed model file group of step 6-2-2 is loaded for inference, and the inference of the other sub-models is suspended; the overall computing power consumption is reduced by 60% through the above dynamic scheduling.

[0068] Here, the present invention uses the dynamic detection and scheduling strategy based on the working condition and the real-time computing power occupancy control instructions as the scheduling logic input of the PLC interface integration link in sub-step 6-5, and is used to encapsulate the communication control instructions.

[0069] Sub-step 6-5 involves establishing a bidirectional communication channel between the embedded inference unit of the engine room edge computing server and the engine room PLC, based on the Modbus RTU hardware communication bus of the ship's engine room PLC and the Modbus RTU protocol. Combining the dynamically adjusted inference strategy from sub-step 6-4, a standardized communication instruction set is encapsulated, and a standardized Modbus RTU interactive interface is generated based on this set. This standardized Modbus RTU interactive interface enables three types of data interaction: 1) The ship's engine room PLC sends control commands to the embedded inference unit in the engine room edge computing server through the Modbus RTU standardized interactive interface to control the start and stop of the fault diagnosis program of the lightweight model in steps 4 and 5, based on the lightweight full set of model weight files in sub-steps 3-5. 2) The embedded inference unit uploads and transmits the fault root cause location and diagnosis results output in step 5 to the engine room PLC through the Modbus RTU standardized interactive interface for alarm display in the engine room monitoring system. 3) When the embedded inference unit detects a major equipment failure, it sends a low-level equipment control command to the ship's engine room PLC through this interface. The ship's engine room PLC then drives the ship's low-level field equipment, such as chilled water units and fans, to perform safety operations such as shutdown and load reduction.

[0070] Here, the Modbus RTU standardized interactive interface serves as the standardized external communication interface for the entire lightweight fault detection system, enabling the issuance of inference scheduling commands, the uploading of fault diagnosis results, and the integrated operation with the ship's engine room PLC.

[0071] The Modbus RTU standardized interactive interface serves only as a communication implementation carrier. Relying on the encapsulated standardized communication instruction set, it maps various business interaction behaviors defined by the inference strategy in sub-step 6-4 to the register allocation, data encapsulation format, and message response rules corresponding to the Modbus RTU bus, thereby realizing the hardware bidirectional transmission of business instructions and diagnostic data.

[0072] The engine room operators / engine room monitoring system operate the ship's engine room PLC. The engine room PLC sends control commands to the embedded inference unit inside the edge server via a Modbus interface, issuing commands to start / stop AI fault detection, query model parameters, and modify detection cycles—essentially controlling the start / stop and parameter reading of the AI ​​inference program on the engine room edge computing server side.

[0073] Once the AI ​​inference on the edge computing server side of the engine room detects faults and abnormal risks, the embedded inference unit sends the underlying equipment action instructions, such as equipment shutdown, load reduction, and valve adjustment, to the ship's engine room PLC. After receiving the instructions, the ship's engine room PLC drives the underlying actuators such as fans and chilled water units.

[0074] Preferably, in terms of data read and write management, this application can use PostgreSQL and Prisma tools for data read and write design. Database Design: A PostgreSQL database is used to store collected runtime data, historical reference and similar data training set data, model parameters, fault records, and other information. Based on the data type and purpose, the database table structure includes: a current sliding window runtime data table, a historical reference and similar data training set data table, a model parameter table, and a fault record table. Object-Relational Mapping Framework Application: Prisma is used as the object-relational mapping (ORM) framework to achieve interaction with the PostgreSQL database. The data model is defined through Prisma Schema, generating the corresponding database table structure and providing a concise Application Programming Interface (API) for data query, insertion, update, and deletion operations, simplifying the data management development process. Preferably, in terms of data visualization, this application can design a corresponding interface and use Python's Matplotlib, Plotly library, or front-end ECharts for data visualization.

[0075] Interface design: The designed interface includes the overall system operation status interface, the operation parameter interface of each subsystem, the fault alarm interface, and the historical data query interface. Display content: The overall system operation status interface displays the overall operation of the system, such as the number of normally operating devices, the number of faulty devices, and the operation status of each subsystem; the subsystem operation parameter interface displays the detailed operation parameters of the corresponding subsystem, such as real-time values ​​and historical trend curves of temperature, pressure, flow rate, etc.; the fault alarm interface displays fault information in real time, including the fault occurrence time, faulty device, fault variables (or symptom rules), etc., and provides audible and visual alarms; the historical data query interface allows operators to query operation data and fault records over a period of time, and displays them in the form of charts and graphs.

[0076] This invention is applicable to multi-subsystem, multi-operating-condition, and highly dynamic marine air conditioning scenarios, enabling three-level fault symptom localization: system-level, subsystem-level (equipment-level and sub-system-level), and variable-level. Through "dynamic training sets, hierarchical structure, and ship-adaptive optimization," this invention addresses the adaptability, accuracy, and real-time issues of existing methods in marine environments.

[0077] Compared with the prior art, the beneficial effects of the present invention are as follows: 1) Improved adaptability to operating conditions: The accuracy of the sub-operating condition model is ≥92% under cooling / heating / ventilation conditions, which is 25%~40% higher than the existing methods; 2) Adaptability of historical reference similar data training set: The dynamically updated historical reference similar data training set resulted in an accuracy decay rate of only 3% after 3 months, compared to a 18%~20% decrease compared to the fixed historical reference similar data training set; 3) Fault location accuracy: Level 3 positioning can pinpoint the specific variable within 10 seconds, improving efficiency by 20% compared to manual inspection; 4) Ship deployment adaptation: The model size is compressed to 38.5MB, the inference time is ≤0.12 seconds, and it is adapted to embedded devices in the engine room.

[0078] The following is combined Figures 1-12 The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. In this embodiment, combined with Figure 1 As shown, the marine air conditioning system includes: 8 centralized air conditioners, 8 air systems, 2 chillers, and 1 water system (including: 2 chilled water pumps, 2 cooling seawater pumps, and 1 chilled water circulation system). Each time, 294 measurement variables are collected in a sliding window (20 variables per centralized air conditioner, totaling 160 variables for 8 units; 36 variables per air system; 16 variables per chiller, totaling 32 variables for 2 units; 12 variables per chilled water pump, totaling 24 variables for 2 units; 10 variables per cooling seawater pump, totaling 20 variables for 2 units; 22 variables per chilled water circulation system; totaling 294 variables).

[0079] The 294 measurement variables include: (1) 20 variables per centralized air conditioner (160 variables in total for 8 units): air conditioner operating mode (1 cooling; 2 heating; 3 ventilation), fan start / stop status (0 off, 1 on), chilled water flow rate (m³ / h).3 / h), Fan coil unit inlet water temperature (°C), Fan coil unit outlet water temperature (°C), Three-way water valve opening (%), Fan coil unit inlet air temperature (°C), Fan coil unit inlet air relative humidity (%), Fan coil unit outlet air temperature (°C), Fan coil unit outlet air relative humidity (%), Fresh air temperature (°C), Fresh air relative humidity (%), Return air temperature (°C), Return air relative humidity (%), Supply air temperature (°C), Supply air relative humidity (%), Supply air temperature setpoint (°C), Total air volume at the supply fan outlet (m³ / h) 3 / h), blower operating current (A), blower operating frequency (Hz).

[0080] (2) 36 variables for 8 air systems: external static pressure of centralized air conditioners No. 1-8 (Pa), supply air temperature of fire station No. 1 (°C), relative humidity of fire station No. 1 (%), and air volume of the main supply duct of fire station No. 1 (m³). 3 / h), air supply temperature of compartments 1-3 (°C), relative humidity of compartments 1-3 (%), air volume of main air supply duct of compartments 1-3 (m³ / h) 3 / h), supply air temperature (°C) of compartments 1-6, relative humidity (%) of compartments 1-6, and air volume of the main supply duct of compartments 1-6 (m³ / h). 3 / h), No. 2 fire station supply air temperature (°C), No. 2 fire station relative humidity (%), No. 2 fire station main air supply duct air volume (m³) 3 / h), supply air temperature (°C) of compartments 2-5, relative humidity (%) of compartments 2-5, and air volume of the main supply duct of compartments 2-5 (m³ / h). 3 / h), Captain's cabin supply air temperature (°C), Captain's cabin relative humidity (%), Captain's cabin main supply air volume (m³ / h) 3 / h), medical cabin air supply temperature (°C), medical cabin relative humidity (%), medical cabin main air supply duct air volume (m³ / h) 3 / h), cockpit air supply temperature (°C), cockpit relative humidity (%), cockpit main air supply duct air volume (m³ / h), 3 / h), isolation chamber supply air temperature (°C), isolation chamber relative humidity (%), isolation chamber main air volume (m³ / h). 3 / h).

[0081] (3) 16 variables per chiller unit (32 variables in total for 2 units): compressor start / stop status (0 shutdown, 1 start), compressor operating current (A), chiller unit input voltage (V), compressor vibration acceleration in the X direction (dB), compressor vibration acceleration in the Y direction (dB), compressor vibration acceleration in the Z direction (dB), compressor vibration velocity in the X direction (mm / s), compressor vibration velocity in the Y direction (mm / s), compressor vibration velocity in the Z direction (mm / s), compressor suction pressure (kPa), compressor discharge pressure (kPa), compressor discharge temperature (°C), compressor cumulative running time (h), chilled water outlet temperature (°C), chilled water return temperature (°C), chilled water flow rate (m³ / h). 3 / h), Cooling water inlet and outlet temperature difference (°C).

[0082] (4) 12 variables for each chilled water pump (24 variables in total for 2 pumps): chilled water pump start-stop status (0 stopped, 1 started), chilled water pump operating current (A), chilled water pump inlet pressure (MPa), chilled water pump outlet pressure (MPa), Y-type filter inlet pressure (MPa), Y-type filter outlet pressure (MPa), chilled water pump vibration acceleration X direction (dB), chilled water pump vibration acceleration Y direction (dB), chilled water pump vibration acceleration Z direction (dB), chilled water pump vibration velocity X direction (mm / s), chilled water pump vibration velocity Y direction (mm / s), chilled water pump vibration velocity Z direction (mm / s).

[0083] (5) 10 variables for each cooling seawater pump (20 variables in total for 2 pumps): seawater pump start-stop status (0 stops, 1 starts), seawater pump operating current (A), seawater pump inlet pressure (MPa), seawater pump outlet pressure (MPa), seawater pump vibration acceleration X direction (dB), seawater pump vibration acceleration Y direction (dB), seawater pump vibration acceleration Z direction (dB), seawater pump vibration velocity X direction (mm / s), seawater pump vibration velocity Y direction (mm / s), seawater pump vibration velocity Z direction (mm / s).

[0084] (6) 22 variables in one cold water circulation system: main water supply flow rate of the No. 1 air conditioning circulating water front zone (m³) 3 / h), water supply flow rate of the first branch pipe in the No. 1 air conditioner circulating water front area (m³) 3 / h), the main water supply flow rate of the No. 1 air conditioning circulating water front area (m³) 3 / h), No. 1 fire station fan coil unit inlet water temperature (°C), No. 1 fire station fan coil unit inlet water pressure (MPa), 1-6 compartment fan coil unit inlet water temperature (°C), 1-6 compartment fan coil unit inlet water pressure (MPa), No. 2 air conditioning circulating water rear zone main water supply flow rate (m³ / h). 3 / h), water supply flow rate of the middle branch pipe in the rear zone of air conditioner No. 2 (m³)3 / h), flow rate of the stern branch pipe in the rear zone of air conditioning circulating water (m³ / h) 3 / h), No. 2 fire station fan coil unit inlet water temperature (°C), No. 2 fire station fan coil unit inlet water pressure (MPa), 2~5 compartment fan coil unit inlet water temperature (°C), 2~5 compartment fan coil unit inlet water pressure (MPa), captain's cabin fan coil unit inlet water temperature (°C), captain's cabin fan coil unit inlet water pressure (MPa), medical cabin fan coil unit inlet water temperature (°C), medical cabin fan coil unit inlet water pressure (MPa), bridge fan coil unit inlet water temperature (°C), bridge fan coil unit inlet water pressure (MPa), isolation compartment fan coil unit inlet water temperature (°C), isolation compartment fan coil unit inlet water pressure (MPa).

[0085] In this embodiment, combined with Figure 1 As shown, in a marine air conditioning system comprising 2 chiller units, 2 chilled water pumps, 2 cooling seawater pumps, 8 centralized air conditioners, and 8 corresponding air systems, 294 key measurement variable acquisition points are distributed across the aforementioned 16 devices (or subsystems) and pipelines. Corresponding sensors are installed at each acquisition point: PT100 temperature sensors with an accuracy of ±0.5℃; diffused silicon pressure sensors with an accuracy of ±0.2% FS; electromagnetic flow meters with an accuracy of ±0.5% FS; and Hall effect speed sensors with an accuracy of ±1 r / min.

[0086] The data acquisition module uses an industrial-grade data acquisition card, connected to each sensor via shielded cables, with a sampling frequency set to 1 time per minute. The acquired data is transmitted via the ship's internal Ethernet communication module (e.g., fiber optic transmission at a rate of 100Mbps) to the server in the data processing center (e.g., configured with an Intel Xeon E5 processor, 16GB of RAM, and a 1TB solid-state drive). During transmission, the data is encrypted using the AES-128 encryption algorithm to ensure data transmission security. Upon receiving the data, the server first performs data verification, using a CRC32 checksum to verify data integrity. If verification fails, retransmission is required.

[0087] In this embodiment, the data processing flow is combined with Figure 2 As shown, data acquisition involves obtaining real-time data through sensors and data acquisition cards. Data transmission and verification: the data is encrypted and transmitted to the server via Ethernet. The server uses a CRC32 checksum to verify data integrity; if verification fails, retransmission is required. After successful verification, the data enters the "data preprocessing" stage, which includes two sub-steps: data cleaning and normalization. Data cleaning uses the Laida criterion to remove outliers and linear interpolation to fill in missing values; normalization uses the min-max standardization method to map the data to the [0,1] interval.

[0088] The principal component similarity factor between the current sliding window data and the original historical window data in the historical reference big data set is calculated according to the formula. Then, the historical reference similar data training set is constructed and updated through "similar data filtering" (with a threshold of 0.85). At the same time, the historical data training set cleanup program is automatically started every Sunday at 0:00. After the historical data training set is ready, the "autoencoder model training" is started. Data is classified by subsystem and operating condition and input into the corresponding sub-model for training. After training is completed, the "real-time fault detection" is started. The reconstruction error is calculated and compared with the threshold to determine whether a fault exists. If a fault is found, "fault location" is performed to determine the subsystem, equipment and variable where the fault is located.

[0089] In this embodiment, combined with Figure 3 As shown, in the data preprocessing stage: after receiving the data, the server first performs data cleaning. The Laida criterion is used to calculate the mean of the historical data for each variable. μ and standard deviation σ When a certain data point x satisfy When a value is identified as outlier, it is removed. For missing values, linear interpolation is used to fill in the missing values. x i Missing data is identified by its preceding and following valid data points. x i-1 and x i+1 The calculation yielded: ; In the formula: This represents the time corresponding to the data point.

[0090] Using formula Normalization is performed using the min-max standardization method. For example, in historical data of a certain temperature variable, min( x ) = 10℃, max( x If 25℃ = 30℃, then the normalized value of 25℃ is (25-10) / (30-10) = 0.75.

[0091] Principal Component Analysis (PCA) Stage: The PCA module of Python's scikit-learn library is used to perform PCA on the preprocessed 294-dimensional data. A principal component contribution rate threshold of 85% is set. For example, if the cumulative contribution rate of the first 10 principal components reaches 85%, then... k =10.

[0092] Principal component similarity factor calculation: The principal component vectors of the current running data are p = [ p 1,p 2, ..., p 10 The principal component vector of historical operating data is q = [ q 1, q 2, ..., q 10 Calculate the similarity factor according to the following formula.

[0093] ; For example, if p = [0.2, 0.3, 0.1, 0.4, 0.2, 0.1, 0.3, 0.2, 0.1, 0.3], q = [0.25, 0.28, 0.12, 0.38, 0.22, 0.11, 0.29, 0.21, 0.12, 0.28], calculated S =0.92.

[0094] Similar data filtering: Set the threshold to 0.85, when... S When the value is ≥ 0.85, the original window data from the historical reference big data set is added to the historical reference similar data training set. The historical reference similar data training set cleanup program is automatically started every Sunday at 00:00, deleting historical data that has been stored for more than 6 months and has not been matched within the last 30 days.

[0095] In this embodiment, combined with Figure 4 As shown, the global model consists of 28 core variables: 1. Air Conditioner Subsystem (8 variables): Supply air temperature (°C) of 8 centralized air conditioners, one for each, for a total of 8 variables (physical meaning: core output parameters of terminal equipment, directly reflecting the air conditioning effect and system coordination).

[0096] 2. Air system (6 variables): Average external static pressure of air conditioners 1-8 (Pa) (1 variable), air volume of main air supply pipe of fire station 1 and fire station 2 (m³ / h) (2 variables), air supply temperature of cockpit, captain's cabin and medical cabin (°C) (3 variables, physical meaning: reflecting the overall resistance characteristics of the air system and the environmental control effect of key areas).

[0097] 3. Chiller unit subsystem (6 variables): chilled water outlet temperature (°C) of 2 chiller units (2 variables), chilled water return temperature (°C) of 2 chiller units (2 variables), chilled water flow rate (m³ / h) of 2 chiller units (2 variables).

[0098] 4. Water System (8 variables): outlet pressure of 2 chilled water pumps (MPa) (2 variables), outlet pressure of 2 cooling seawater pumps (MPa) (2 variables), main water supply flow rate of the front zone of the chilled water circulation system (m³ / h) (1 variable), main water supply flow rate of the rear zone of the chilled water circulation system (m³ / h) (1 variable), inlet water temperature of fan coil units of fire stations No. 1 and No. 2 (°C) (2 variables, physical meaning: reflecting the dynamic characteristics and cooling capacity distribution balance of the water system).

[0099] In this embodiment, combined with Figure 4 The model parameters for each system are shown below: (1) Global / system-level lightweight autoencoder model: 28 neurons in the input layer (corresponding to 28 core variables selected from 294 variables in 4 subsystems), 28→18→10 neurons in the encoder layer, 5 neurons in the latent layer, 10→18→28 neurons in the decoder layer, ReLU activation function for all, and Sigmoid function for the output layer. The total number of model parameters is controlled within 1000.

[0100] (2) Each air conditioner subsystem: 20 neurons in the input layer (corresponding to 20 variable data after screening of 8 air conditioners), 20→12→6 neurons in the encoder layer, 6→12→20 neurons in the decoder layer, ReLU activation function for all, and Sigmoid for the output layer.

[0101] (3) Wind system Sun system: 3 neurons in the input layer, the number of encoder neurons is 36→24→12→6, the number of decoder neurons is 6→12→24→36, and the activation function is the same as above. (4) Each chiller unit's subsystem: 16 neurons in the input layer, 16→10→5 neurons in the encoder, 5→10→16 neurons in the decoder, and the activation function is the same as above.

[0102] (5) Chilled water pump system: 12 neurons in the input layer, 12→8→4 neurons in the encoder, 4→8→12 neurons in the decoder, and the activation function is the same as above.

[0103] (6) Cooling seawater pump system: 10 neurons in the input layer, 10→6→3 neurons in the encoder, 3→6→10 neurons in the decoder, and the activation function is the same as above.

[0104] (7) Cold water circulation system: 22 neurons in the input layer, 22→14→8→4 neurons in the encoder, 4→8→14→22 neurons in the decoder, and the activation function is the same as above.

[0105] In this embodiment, combined with Figure 5As shown, taking a lightweight autoencoder model at the global / system level under cooling conditions as an example, the encoder part contains three fully connected layers, and the input layer receives... n Wei (I) n The global / system-level cooling condition-related variable data (taking 28 as an example). The activation function of the output layer of the first layer (hidden layer 1: 18 neurons) is: ; In the formula: ReLU is the activation function of the output layer in a neural network-based autoencoder; W 1 is a 18×28 weight matrix; b 1 is an 18-dimensional bias vector.

[0106] The output of the second layer (hidden layer 2: 10 neurons) is: ; The output of the third layer (potential layer: 5 neurons) is: ; That is, the encoded feature vector.

[0107] The decoder is symmetrical to the encoder. The output of the first layer (latent layer 1: 10 neurons) is: ; The output of the second layer (latent layer 2: 18 neurons) is: ; The output of the third layer (output layer: 28 neurons) uses the Sigmoid activation function, as follows: ; That is, to reconstruct the data.

[0108] In this embodiment, combined with Figure 4 As shown, when integrating the various subsystems in the global / system-level lightweight autoencoder model, the fusion weights of each subsystem model are determined through grid search: Air conditioner subsystem w 1=0.3, wind system w 2=0.2, chiller unit subsystem w 3=0.3, water system w 4 = 0.2, the total model output is the reconstructed data: ; , , , These represent the output results (i.e., reconstructed data) of the autoencoder models for the air conditioner subsystem (8 variables), air system (6 variables), chiller unit subsystem (6 variables), and water system (8 variables), respectively.

[0109] Model training was performed using the Keras API in TensorFlow, with the training and validation sets divided in an 8:2 ratio using historical, similar data. The Adam optimizer was used with a learning rate of 0.001, a batch size of 32, and 50 training epochs. Training stopped when the validation set loss did not decrease for five consecutive epochs. After training, the model achieved a mean squared error of 0.002 on the validation set.

[0110] In this embodiment, combined with Figure 6 As shown, "Input Real-Time Preprocessed Data" involves inputting the processed real-time data into the trained autoencoder model. Next, "Calculate Reconstruction Errors for Each Subsystem and Variable" is performed, obtaining the RMSE for the air conditioner subsystem, chiller subsystem, air system, water system, and each device and variable. Then, a "Global / System-Level Overall Reconstruction Error Comparison with Threshold" is conducted. If the reconstruction error is ≤ the threshold, it is determined to be "No Fault"; if the reconstruction error is > the threshold, it proceeds to "Subsystem Reconstruction Error Comparison".

[0111] In the "Subsystem Reconfiguration Error Comparison," examine the RMSE of each of the four subsystems and identify the subsystem with the highest RMSE. For example, the air conditioner subsystem has an RMSE of 0.26, which indicates the subsystem where the fault occurs. Next, proceed to "Subsystem Equipment Reconfiguration Error Comparison" to examine the RMSE of all equipment within that subsystem and identify the equipment with the highest RMSE, such as the third air conditioner. Then, perform "Equipment Variable Reconfiguration Error Comparison" to examine the RMSE of all variables for that equipment and identify the variable with the highest RMSE, such as the supply and return air temperature variable with an RMSE of 0.32.

[0112] Finally, "output fault information" clearly identifies the subsystem, device, and variable where the fault occurs, as well as the corresponding fault description, such as "the reconstructing error of the supply and return air temperature variables of the third air conditioner in the air conditioner subsystem exceeds the standard".

[0113] In this embodiment, combined with Figure 7 As shown, real-time data, after preprocessing, is input into the autoencoder model of the corresponding subsystem under the corresponding operating condition to calculate the reconstruction error of each subsystem and variable. The reconstruction error statistics during normal system operation show... μ =0.05, σ =0.03, the RMSE threshold can be obtained. T = 0.05 + 3 × 0.03 = 0.14. For example, if the root mean square error (RMSE) is 0.15 ≥ 0.14, it is considered abnormal (or "symptoms of failure").

[0114] In a certain calculation example, (1) in "Level 1 positioning", the RMSE of the air conditioner subsystem is 0.26 (the air conditioner subsystem includes the RMSE of each air conditioner). Since 0.26>=0.14, the air conditioner subsystem is judged to be "abnormal". In addition, since the RMSE of the chiller subsystem is 0.06, the RMSE of the air system is 0.07, and the RMSE of the water system is 0.05, they are all judged to be "normal".

[0115] The steps for calculating the RMSE of the air conditioning subsystem are as follows: 1) Data Input: The pre-processed real-time operating data belonging to the air conditioning subsystem is input into the autoencoder submodel of the air conditioning subsystem under the corresponding operating condition. The "corresponding operating condition" here should be determined in combination with the current actual operating status of the marine air conditioning system (cooling, heating or ventilation condition), and the submodel matching the operating condition should be selected.

[0116] 2) Model operation: The autoencoder sub-model first encodes (extracts features) the input air conditioner subsystem data, and then decodes (reconstructs data) to obtain the reconstructed data (i.e. the output data that the model attempts to restore and is close to the original input).

[0117] 3) Calculate RMSE: Based on the original input data and the reconstructed data from the model output, calculate the difference between the two using the Root Mean Square Error (RMSE) formula. The formula is: ; In the formula: It is the original input data. This involves reconstructing the data, where 'm' represents the data dimension. Calculations show that an RMSE of 0.26 for the air conditioner subsystem is abnormal, while an RMSE < 0.14 for other subsystems is normal.

[0118] (2) In the “secondary positioning”, after the data of the third air conditioner (equipment level and subsystem level) is input into the air conditioning equipment self-encoder model, the reconstruction error is obtained, and its RMSE is calculated to be 0.18, which is greater than the threshold. T =0.14. Therefore, the third air conditioner is judged as "abnormal", while the other subsystems and equipment are considered normal if RMSE < 0.14.

[0119] (3) In the “three-level positioning”, for example: the RMSE calculated from the reconstruction error of the supply and return air temperature variables is 0.32, which is greater than the threshold. T If RMSE = 0.14, the component corresponding to that variable (such as a temperature sensor or return air valve) is considered faulty and abnormal; other variables with RMSE < 0.14 are considered normal.

[0120] The fault conclusion is: the supply and return air temperature sensor of air conditioner A3 is faulty.

[0121] Actual deployment optimization: Model pruning was performed using the TensorFlow Model Optimization Toolkit, removing connections with absolute weight values ​​less than 0.01. After pruning, the number of model parameters decreased by 30%, and the inference speed improved by 40%. After using INT8 quantization, the model size was reduced from 220MB to 38.5MB, and the inference time on embedded devices was shortened from 0.52 seconds to 0.12 seconds. A comparison of the model algorithm before and after simplification is shown in Table 2.

[0122] Table 2 After data processing simplification, variables with small fluctuations, such as pressure and rotational speed, are normalized every 5 minutes; variables with large fluctuations, such as temperature and humidity, are still normalized every 1 minute. The interface adopts the Modbus RTU protocol and connects to the PLC via an RS485 bus. The PLC can send commands such as "start detection," "stop detection," and "parameter query," and the command format conforms to the Modbus protocol specification.

[0123] Data read / write management: The PostgreSQL database is installed on the server's Ubuntu 20.04 system, using the InnoDB storage engine, and the database character set is set to UTF-8. The data tables are partitioned by time, with one partition per hour, to improve query efficiency.

[0124] A fault log table model can be defined by the Prisma schema; fault logs for the most recent 24 hours can be queried through the Prisma Client.

[0125] In this embodiment, combined with Figures 8-12 As shown, the overall system operation status interface uses blue as the main color. The top displays the system name and the current time, the top left shows the system health status (green indicates normal, red indicates fault), and below that are the operating modes (cooling, heating and ventilation) and descriptions of key operating parameters.

[0126] Figure 8 The Value in the table represents the real-time monitoring value (measured physical quantity) of the corresponding sensor measurement point; for example: real-time sampling data of supply air temperature, water flow rate, total outlet air volume, and supply fan current. The horizontal axis HH:MM = Hour:Minute represents the sampling time (time axis).

[0127] Figure 10 The Value in the figure represents the measurement point value, that is, the vertical axis represents the real-time observed value of the physical quantity corresponding to the sensor (time-series observation of air supply temperature, water flow rate, air supply fan current, etc.).

[0128] Figure 11 The Value in the table represents the real-time time-series sensor observations of the chilled water temperature, chilled water flow rate, chiller voltage, and chiller current in the computer room water system.

[0129] Figure 12 The Value in the table represents the real-time time-series sensor observations of the chilled water temperature, chilled water flow rate, chiller voltage, and chiller current of the cabin water system.

[0130] The operating parameter interfaces for each subsystem display curves showing the changes in key operating performance parameters over time. When the fault alarm interface pops up, the "System Health Status" text turns red, and the alarm box is located at the top center of the interface. In the air conditioner equipment operating interface, the alarm box displays "Fault Subsystem: Air Conditioner; Fault Equipment: Air Conditioner 3; Fault Variable: Supply and Return Air Temperature; Fault Description: Reconstruction Error 0.3 > Threshold 0.14". Clicking the "Confirm" button will disable the alarm.

[0131] The historical data query interface provides a calendar control for selecting a time range. Query results are displayed in tables (showing time, device, variables, and values) and bar charts (averaging hourly statistics). Tables can be exported to Excel format. .xlsx).

[0132] 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.

[0133] 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 method for detecting fault symptoms in a marine air conditioning system based on a self-encoder, characterized in that, include: Step 1: The industrial data acquisition card collects the raw physical quantities from sensors at various points of the marine air conditioning system, processes the raw physical quantities, and obtains the real-time raw sliding window matrix. ; Step 2, the cabin edge computing server is based on the real-time raw sliding window matrix. Construct a training set of dynamic historical reference similar data for different working conditions and real-time Normalized window matrix ; Step 3: Based on the training set of similar data from the dynamic historical reference of the different working conditions and the autoencoder, obtain the training convergence complete floating-point weight model file and the lightweight complete set of model weight files; Step 4: Based on the real-time normalized window matrix and the lightweight complete set of model weight files, obtain the ship's operating condition label; Step 5: Based on the ship's operating condition markings, obtain the fault root cause location and diagnosis results; Step 6: Based on the training and convergence of the complete floating-point weight model file, obtain the dynamically adjusted inference strategy; Step S1: The industrial data acquisition card collects the raw physical quantities from sensors at various points of the marine air conditioning system, processes the raw physical quantities, and obtains the real-time raw sliding window matrix. ,include: Sub-step 1-1: Based on the global static hardware configuration, the temperature sensor, pressure sensor, flow sensor and vibration sensor of the marine air conditioning system synchronously collect the original physical quantities at each point to obtain the original m-dimensional instantaneous sampling vector at a single moment, and transmit it to the ship's engine room PLC for storage. Sub-steps 1-2: The industrial data acquisition card reads the raw data of a single moment stored in the ship's engine room PLC. m 3D instantaneous sampling vector and fixed window length n Continuous caching n Each sampling point is used to construct the original unencrypted sliding window matrix. ; Sub-steps 1-3: The industrial data acquisition card outputs the raw, unencrypted sliding window matrix from sub-step 1-2. The data packet is encapsulated using AES-128 symmetric encryption, and a CRC32 checksum is added to obtain an encrypted data packet with a checksum. This data packet is then transmitted to the edge computing server via Ethernet inside the ship. Sub-steps 1-4 involve the cabin edge computing server performing checksum matching on encrypted data packets with checksums based on a local storage verification benchmark algorithm. If the verification code does not match, the cabin edge computing server sends a retransmission instruction to the industrial data acquisition card, returning to sub-steps 1-3 to repackage. If the checksum matches, the encrypted data packet with the checksum is decrypted to restore the original real-time sliding window matrix. ; Step 2, the cabin edge computing server is based on the real-time raw sliding window matrix. Construct a dynamic historical reference similar data training set and a real-time normalized window matrix for different working conditions. ,include: Sub-step 2-1: The cabin edge computing server calculates the mean of the variables column by column for the restored original sliding window matrix. Standard deviation Remove The outliers are identified, and a coarsely cleaned matrix of missing data is obtained to fill in the missing data. ; Sub-step 2-2: The cabin edge computing server performs coarse cleaning of the missing matrix to be filled. Linear interpolation is performed between two adjacent points in the time series to fill in the blank sampling points, resulting in a complete cleaning matrix without missing data. ; Sub-steps 2-3: Cabin edge computing server, based on a complete cleaning matrix without missing parts. The real-time normalized window matrix is ​​obtained by applying the following normalization formula. : ； in: Complete cleaning matrix without missing parts A specific element data in; for The normalized data, i.e., the normalized window matrix A specific element data in; min( () is a complete cleaning matrix without missing parts Take the minimum observation value of each column individually; each column has an independent minimum observation value; max( () is a complete cleaning matrix without missing parts Take the maximum observation value for each column individually; each column has an independent minimum observation value. Sub-steps 2-4: The cabin edge computing server, based on the normalized window matrix output from sub-step 2-3. The global uniform projection matrix is ​​obtained. and the principal component vector of the current window ]; Sub-steps 2-5, based on the global unified projection matrix This yields PCSF similarity matching records; Sub-steps 2-6: The cabin edge computing server groups the highly similar historical windows in the PCSF similarity matching records that have built-in operating condition identifiers into groups based on the historical window's built-in operating condition identifiers, ensuring that the data within each group is uniformly distributed. It then outputs an uncleaned, sub-operating condition similarity dataset. The operating condition identifiers include: cooling, heating, and ventilation; the dual timing tags include: an archive creation time tag and the most recent PCSF matching time tag. Sub-steps 2-7 set the criteria for determining a sample as redundant if it simultaneously meets the conditions of storage duration exceeding 6 months and within the last 30 days, and has no PCSF similarity matching records; the unique archive number of the historical window within the uncleaned sub-work condition similar dataset is used. As an index, it retrieves the corresponding unique archive number from the historical reference similar data training set data table within the PostgreSQL database. The primary key is used to generate a complete storage record with dual timing tags. Based on the dual timing tags, the sample storage duration and the most recent matching interval are calculated. All storage records in the historical reference similar data training set, distinguished by archive number t, are traversed, and those that meet the conditions for being judged as redundant samples are selected. , And supporting operating conditions, time series and principal component vector data Batch delete the filtered items , And supporting operating conditions, time series and principal component vector data ; Synchronously update the global historical normalized matrix set { Remove deleted entries. ; Based on the updated global history normalized matrix set { All valid historical window samples corresponding to} are re-collected and grouped according to the working condition identifier bound to each sample, and a non-redundant dynamic historical reference similar data training set for each working condition is reconstructed. Sub-steps 2-4: The cabin edge computing server, based on the normalized window matrix output from sub-step 2-3. This yields the principal component vector of the current window. ],include: Sub-step 2-4-1: Calculate the normalized window matrix The mean vector of each variable sample, i.e., the row vector : ； in, For normalized window matrix No. Row vectors, including all The value of the measured variable is the first one. One time-series sampling point; Sub-step 2-4-2, based on row vectors Solve for the sample covariance matrix: ； get covariance matrix Forming a covariance matrix ; in, This is the transpose of the decentralized sample matrix; The mean matrix is ​​composed of the mean values ​​of each feature dimension; Sub-steps 2-4-3 include: Sub-step 2-4-3-1, for the covariance matrix Solving the characteristic equation yields all the results. Each feature value; will all The eigenvalues ​​are rearranged in descending order of value to obtain the complete descending sequence of eigenvalues. Simultaneously extract the feature vector corresponding to each feature value. ; Sub-step 2-4-3-2: Calculate the sum of all eigenvalues: ; Starting from the first-order eigenvalue, the cumulative contribution rate is calculated iteratively by accumulating the values ​​at each order: ; Cyclic Increasing Order until If the cumulative contribution rate is ≥85%, stop the loop and select the top k sets of feature values ​​and feature vectors that satisfy the condition ≥85%, where k = This yields a subset of the first k effective eigenvalues. , and the subset of eigenvectors corresponding to the first k orders Among them, single-branch feature vectors The dimension is ×1; Sub-step 2-4-3-3: For the real-time window link, the feature vector subsets corresponding to the first k orders are... By concatenating the columns horizontally, we obtain a reduced-dimensional two-dimensional matrix: The input to sub-step 2-4-4-1 is then fed into the projection calculation. After the calculation is completed, only the current session is cached. For offline pre-computation links, the dimension-reduced two-dimensional matrix is ​​used. As a globally unified projection matrix Persisting to a PostgreSQL database, all subsequent historical window batch dimensionality reduction sub-steps 2-5-1 will read from the database uniformly. use; Sub-step 2-4-4 includes: Sub-step 2-4-4-1: Calculate the principal component subspace score matrix. ; Sub-step 2-4-4-2: Reconstruct the approximate score matrix based on the principal component subspace score matrix. = ; Sub-step 2-4-4-3: Solve for the residual principal space matrix. = Based on the residual principal space matrix The mean square error of the global residuals of the window is obtained. , Among them, the residual matrix is ​​the first Line number Column element, representing the first The sampling point, the first Reconstruction bias of each measurement variable; Sub-step 2-4-4-4, if Then proceed to sub-step 2-4-4-5 to extract the principal component vector of the current window. Among them, the global static residual threshold = 0.03; if Then, execution will restart from sub-step 2-1 until... ; Sub-step 2-4-4-5: Obtain the principal component subspace score matrix. The last line serves as the global operating condition representation for the window. The numerical combinations of the principal components form a one-dimensional row vector: [ ], that is, the current window principal component vector, which is then fed into sub-steps 2-5 for PCSF similarity calculation with the historical window principal component vectors; the real-time original sliding window matrix that has passed the verification in sub-steps 1-4. Normalized window matrix Assign the same unique, non-duplicate save file number. The real-time original sliding window matrix that has passed the verification in sub-steps 1-4 The archive is named Historical Primitive Matrix The normalized window matrix that passed the verification in sub-steps 1-4 The archive is named the History Normalization Matrix The archive is synchronously written with dual time tags, including: the current system time as the archive creation time tag and the most recent PCSF matching time tag, which is initially blank; the two sets of dual time tags and the history normalization matrix of sub-step 2-4-4-5 are then combined. Operating condition identifier and current window principal component vector Together with the historical reference similar data training set data table in the PostgreSQL database; operating condition identifiers, including: cooling, heating, and ventilation; where, the current window principal component vector After being stored in the data table, it serves as the historical principal component vector corresponding to this historical sample. ; Sub-steps 2-5, based on the global unified projection matrix The PCSF similarity matching records were obtained, including: Sub-step 2-5-1: The cabin edge computing server obtains the global unified projection matrix. Obtain the global history normalized matrix set { }, for the global history normalized matrix set { Each history normalization matrix in} Perform dimensionality reduction: Take the principal component score matrix for each historical record. The last line is based on the The row is a row retrieval operation, and the row contains The principal component values ​​will The principal component values ​​are concatenated into a one-dimensional row vector: ], Traverse all history windows For each Perform the row extraction operation on all rows to generate the corresponding historical principal component vectors. After summarizing all of them, we get the historical principal component vector set { }; Sub-step 2-5-2: The cabin edge computing server calculates the principal component similarity factor group by group using the cosine similarity formula. Bind each principal component similarity factor The corresponding historical window contains the original data, operating conditions, and time tags; ； Sub-steps 2-5-3 involve the cabin edge computing server filtering and retaining data that meets the requirements. A historical window of 0.85, based on principal component similarity factor. Sort in descending order and extract the top items with the highest similarity in the descending order. A subset of historical windows is selected to obtain a highly similar subset of historical windows with working condition identifiers and time tags; the current real-time window is then compared with the extracted window with the highest similarity. The pairing relationships, similarity scores, and archive creation timestamps of each historical window are used as PCSF similarity matching records, written to the database, and the highest similarity records extracted in that batch are updated synchronously. The most recent PCSF match time tag corresponding to each history window, where the archive creation time tag and the most recent PCSF match time tag are used as dual time tags; Step 3: Based on the training set of similar data from the dynamic historical reference for the different working conditions and the autoencoder, obtain the fully converged floating-point weight model file and the lightweight complete set of model weight files, including: Sub-step 3-1: The cabin edge computing server constructs a lightweight global autoencoder based on preset global core variable constraints to obtain the untrained network skeleton; wherein, according to The system classifies equipment into three operating conditions: cooling, heating, and ventilation. An independent sub-encoder is constructed for each piece of equipment in the marine air conditioning system, generating a total of [number missing] encoders. A set of equipment and corresponding operating conditions sub-models; the encoder's mathematical form is... The mathematical form of the decoder is Network weight , With bias , Random initialization; this stage only completes the structure building and does not perform iterative updates. in, This is the encoded feature vector; For activation functions; This is the weight matrix; The historical principal component vector for sub-step 2-4-4-5 During historical sample inference, the historical principal component vectors retrieved from the database are used. Also serves as model input ; , It is the bias vector; for The reconstructed output data; For activation functions; This is the weight matrix; Sub-step 3-2: The cabin edge computing server receives the untrained network skeleton output from sub-step 3-1 and the dynamic historical reference similar dataset for different work conditions obtained from sub-step 2-7. It first sorts the dynamic historical reference similar dataset for different work conditions according to... The subsystem has a total of The system is divided into equipment and operating conditions, with each subsystem comprising multiple devices. For each type of equipment, the corresponding operating condition samples are independently stratified according to a preset ratio, generating training and validation subsets for each operating condition. A 1-bit operating condition identifier is appended to all samples in both the training and validation subsets. The training subset is input into the network for the current iteration to perform forward propagation. The first iteration uses the untrained original network structure skeleton output from sub-step 3-1. Subsequent iterations use a network with updated parameters optimized by Adam, employing the basic mean squared error as the single-sample reconstruction loss. Calculate the reconstruction error, where the reconstruction error is the total weighted reconstruction loss after introducing the weights of each subsystem item. Total weighted reconstruction loss The network parameters are updated via backpropagation using the target value through the Adam optimizer; the encoder parameters are iteratively updated via backpropagation using the Adam optimizer based on the training subset loss. and Decoder parameters and In each iteration, the validation loss is calculated using a subset of validation data. An early stopping strategy is enabled to monitor changes in the validation loss. If the validation loss does not decrease for several consecutive preset rounds, the model is considered to have converged, the iteration is stopped, and a fully converged floating-point weighted model file is obtained. Input the samples from the training subset that operates under all normal conditions into the converged global model in the fully converged floating-point weight model file to calculate the global RMSE in batches, and then calculate the mean global normal RMSE. and standard deviation And persist both to a PostgreSQL database; right The sub-models for each set of equipment and corresponding operating conditions are processed separately: Samples from the entire normal training subset for the corresponding equipment and operating condition are extracted and input into the converged sub-models for the corresponding equipment and operating conditions in the fully converged floating-point weighted model file. The reconstruction RMSE of a single device is calculated in batches, and the mean normal RMSE for each set of equipment and corresponding operating condition sub-model is statistically analyzed in groups. and standard deviation ;Will Group normal RMSE mean and standard deviation Persistently store data in PostgreSQL; For each sensor's original measurement point time series, statistical analysis was performed separately on all normal operating condition samples within the training subset where all operating conditions were normal: traversing the variables of each sensor measurement point. Extract the time-series observation set of the sensor measurement point under all normal operating conditions training samples, and calculate the time-series mean of the variable of the sensor measurement point. and time series standard deviation ; Sub-step 3-3: The cabin edge computing server traverses all network weight connections in the training converged complete floating-point weight model file, deletes weight connections with absolute values ​​less than a preset threshold, completes weight pruning, and obtains the pruned sparse floating-point model file. In sub-steps 3-4, the cabin edge computing server linearly maps the 32-bit floating-point weights in the pruned sparse floating-point model file to the 8-bit signed integer range [-128, 127] to complete the quantization, while simultaneously preserving the encoder network weights. and decoder network weights Map the correspondence and output the intermediate quantization model file; Sub-steps 3-5: The cabin edge computing server removes the encoder bias of the quantized intermediate model. and decoder bias During the inference phase, global static compensation values ​​are used to offset the reconstruction bias caused by the lack of bias, resulting in quantized and debiased intermediate model files. All weights in the quantized and debiased intermediate model files are traversed to verify that the sparse structure is not lost. Finally, a lightweight complete set of model weight files is generated and stored in a PostgreSQL database. Step 4, based on the real-time normalized window matrix Together with the lightweight complete model weight file, the ship's operating condition labels are obtained, including: Sub-step 4-1: The cabin edge computing server obtains the real-time normalized window matrix output from sub-step 2-3. The preset core variable real-time normalized data; the global lightweight model weights extracted from the lightweight complete model weight file output by sub-steps 3-5; the input preset core variable real-time normalized data read at fixed intervals, the global lightweight model weights are loaded, and the global autoencoder is called to complete the encoding and decoding forward inference reconstruction; the root mean square error formula is used. The global reconstruction error for the current period is calculated. Simultaneously, read the global normal RMSE mean from sub-step 3-2. and standard deviation Through the formula: ， Calculate the global decision threshold ,Will and Simultaneously transmitted to subsequent judgment stages; Sub-step 4-2, if If the system is found to be abnormal, a global abnormality flag is output, all corresponding sub-models of all devices are triggered synchronously, and a trigger command for a preset short-cycle secondary fine inspection is initiated, with data flowing into sub-step 4-3. like If the system is deemed to be normal overall, a global normal status flag is output, and device sub-model reasoning is executed in parallel for each device. Then, sub-step 4-3 is entered to determine local minor anomalies. Sub-step 4-3: Read the average normal RMSE value corresponding to the current device under test from the input. and standard deviation Combined with the given The calculation formula calculates the device's specific anomaly detection threshold in real time. Based on the trigger command issued in sub-step 4-2, distinguish the inference cycle. Use short-cycle fine-grained detection for scenarios with global anomaly marking, and use the regular cycle for scenarios with system-wide normal marking. Load the lightweight weights of the device condition sub-model that match the current device and real-time operating conditions from the lightweight full set of model weight files in sub-step 3-5 for each device, and input the real-time normalized window matrix from sub-step 2-3. The single device performs encoding and decoding reconstruction using all real-time normalized variables, and calculates the real-time reconstruction error of the single device. The real-time reconstruction error calculated for each device Compared with the pre-calculated specific anomaly detection threshold By comparing them one by one, the set of abnormal subsystem numbers and / or abnormal device numbers that have reconstruction errors exceeding the threshold are selected. Sub-step 4-4: The cabin edge computing server outputs a real-time normalized window matrix from sub-step 2-3. In the process, real-time time-series data of a preset set of global core variables are read; two sets of time-series measurement points, chilled water flow rate and supply air temperature, are extracted from the preset set of global core variable time-series data to construct a coupling coefficient matrix and calculate the cross-system correlation coupling coefficient in real time; the calculated cross-system coupling coefficient is compared with the normal reference range of the chilled water flow rate and supply air temperature coupling coefficients pre-stored in the database. If the cross-system coupling coefficient deviates from the normal reference range, it is determined that there is a cross-system correlation anomaly that cannot be identified by the traditional single-device model, and a corresponding cross-system coupling anomaly marker is generated. In sub-steps 4-5, if sub-step 4-2 obtains a global normal flag and there is no global abnormal flag in sub-step 4-2, and no cross-system coupling abnormal flag in sub-step 4-4, then a stable navigation flag is generated; if sub-step 4-2 obtains a global abnormal flag, then the abnormal subsystem number and / or abnormal device number obtained in sub-step 4-3 are carried; if sub-step 4-4 obtains a cross-system associated abnormality, then the cross-system coupling abnormal flag is carried; the stable navigation flag, the set of abnormal subsystem numbers, the set of abnormal device numbers, or the cross-system coupling abnormal flag are packaged into a ship condition flag.

2. The method for detecting fault symptoms of marine air conditioning systems based on self-encoders as described in claim 1, characterized in that, Step 5: Based on the ship's operating condition markings, obtain the root cause location and diagnosis results of the fault, including: Sub-step 5-1, based on the set of abnormal subsystem numbers and abnormal device numbers output from sub-step 4-3, and the cross-system coupling anomaly markers output from sub-step 4-4, identifies the abnormal subsystems and abnormal devices to be analyzed; and outputs the real-time normalized window matrix from sub-step 2-3. Extract all measured variable observations from the currently locked anomaly subsystem to be analyzed, and construct a real-time anomaly sample vector. , m This represents the total number of variables in this subsystem; read the real-time anomaly sample vector of this anomaly subsystem. From the lightweight complete model weight file output in sub-steps 3-5, extract the lightweight sub-model weights that match the corresponding operating conditions of the locked abnormal equipment, complete the forward inference, and generate the autoencoder reconstruction output vector. ; Sub-step 5-2 utilizes real-time anomaly sample vectors Reconstructing the output vector with the autoencoder The overall subsystem reconstruction loss is calculated using the following formula to complete the initial screening of the univariate original residuals and obtain the reconstruction loss. : ; Sub-step 5-3 uses the reconstruction loss obtained in step 5-2. Using the objective function, backpropagation is performed based on the lightweight sub-model weights extracted from the lightweight full set of model weight files output from sub-steps 3-5, which match the operating conditions corresponding to the abnormal equipment. Solve for the partial derivatives of the decoder output with respect to the input variables. By combining the univariate residuals and calculating the Jacobian matrix elements variable by variable, the original sensitivity contribution of a single variable to the overall reconstruction error is quantified. Specifically, for a single sample, the Jacobian matrix... J The dimension is 1× m , where the Jacobian gradient of the variable J 1,i Defined as: = ； Sub-step 5-4: Jacobian gradient for all variables J 1,i Standardize the absolute value to eliminate dimensional differences, and calculate the basic standardization sensitivity. : ； in, For the standardized variable sensitivity, satisfy =1; in, It is the output of the autoencoder decoder. For input The partial derivatives; Sub-step 5-5 calculates the absolute value of the original Jacobian gradient used in sub-step 5-4. Using the base case as the calculation basis, substitute into the following formula to calculate the weighted correction contribution. Complete the correction of the physical importance weights of the equipment: ; Among them, the physical importance weights of global variables pre-stored in the database , , ; Sub-steps 5-6 involve adjusting the weighted contribution. Sort by size from largest to smallest, and read the preset fixed threshold. Filter to meet The variables are used as highly sensitive suspected fault variables; Sub-step 5-7: Read the time-series mean values ​​of the sensor measurement point variables corresponding to the currently locked anomaly subsystem to be analyzed, as described in sub-step 3-2. and time series standard deviation Using the highly sensitive suspected fault variables obtained from sub-steps 5-6 as the analysis object, and combining them with the cross-system coupling anomaly markers in the ship's operating condition markings, the fault types are distinguished. Specifically, if the real-time observed values ​​of the highly sensitive suspected fault variables... Deviating from the normal range If the cross-system coupling anomaly of a highly sensitive suspected fault variable is marked as "present", it is determined to be a cross-system associated fault.

3. The method for detecting fault symptoms of marine air conditioning systems based on self-encoders as described in claim 2, characterized in that, Step 6: Train and converge the complete floating-point weight model file to obtain the dynamically adjusted inference strategy, including: Sub-step 6-1, Model structure optimization: Based on the baseline complete autoencoder network structure in the training converged complete floating-point weight model file of sub-step 3-2, under the premise that the fault detection accuracy is not lower than the preset accuracy constraint threshold, the number of encoder and decoder network layers and the size of neurons are reduced, and the 3-layer encoder and 3-layer decoder are simplified to 2-layer encoder and 2-layer decoder, and a lightweight basic network structure is output. Sub-step 6-2 includes: Sub-step 6-2-1: Load the lightweight basic network structure and matching FP32 raw floating-point weights from sub-step 6-1. Use the matching FP32 raw floating-point weights to forward infer a batch of samples and calculate the contribution of each network connection to the reconstruction error. Compare the contribution of each connection with the reconstruction error sensitivity threshold and remove neurons and network connections with contributions below the threshold. Simultaneously retain the corresponding FP32 weights for the retained network structure and delete the weight parameters of the pruned connections to obtain the sparse network and corresponding FP32 weights after pruning. Sub-step 6-2-2 imports the pruned sparse network structure and corresponding FP32 weights into the TensorFlow Lite offline conversion toolchain; performs quantization mapping calculation based on the global numerical distribution of FP32 weights to obtain the scaling factor and zero offset required for converting floating-point to fixed-point or integer, as the mapping calculation result; based on the mapping calculation result, converts all valid 32-bit floating-point weights into 16-bit fixed-point numbers or 8-bit integers to generate quantized integer weight parameters; during the conversion process, the sparse network connection relationship obtained from the pruning in the input is preserved, and only the valid connection parameters are quantized, while the pruned invalid connections are removed; combines the pruned sparse network from sub-step 6-2-1 with the quantized integer weight parameters obtained in this step to generate a complete compressed model file; Sub-step 6-3 calls the TF Lite embedded inference runtime library in the TensorFlow Lite compression toolchain to load the complete compressed model file from sub-step 6-2-2, and completes the model deployment in the ship's embedded hardware to obtain the deployed lightweight inference model; inputs a unified standard test sample into the deployed lightweight inference model, performs multiple rounds of forward inference, and collects the time consumption using the built-in timing function of the TF Lite embedded inference runtime library to calculate the measured inference time of the compressed model; reads the baseline time of the original FP32 model, compares the measured inference time of the compressed model with the baseline time, and obtains the comparison index of the model inference time of a single sample before and after model compression; Sub-step 6-4 calls the PLC real-time computing power monitoring value, the model inference time comparison index output from sub-step 6-3, and the ship operating condition markers from step 4-5, and combines them with the complete compressed model file output from sub-step 6-2-2 to dynamically adjust the inference strategy, including: If the ship's operating condition is marked as stable navigation, based on the comparison index of model inference time, a scheduling instruction is issued to extend the global loop detection cycle of sub-step 4-1 and reduce the execution frequency of the complete compressed model file inference in sub-step 6-2-2; If the ship condition marker carries the abnormal subsystem number, only the lightweight model of the corresponding subsystem in the complete compressed model file group of sub-step 6-2-2 is loaded for inference, and the inference of the other sub-models is paused. Sub-step 6-5 involves establishing a bidirectional communication channel between the embedded inference unit of the engine room edge computing server and the engine room PLC, based on the Modbus RTU hardware communication bus of the ship's engine room PLC and the Modbus RTU protocol. Combining the dynamically adjusted inference strategy from sub-step 6-4, a standardized communication instruction set is encapsulated, and a standardized Modbus RTU interaction interface is generated based on this set. This standardized Modbus RTU interaction interface enables three types of data interaction: The ship's engine room PLC sends control commands to the embedded inference unit in the engine room edge computing server through the Modbus RTU standardized interactive interface. This is used to control the start and stop of the fault diagnosis program of the lightweight model in the lightweight full set of model weight file based on sub-steps 3-5 in steps 4 and 5. The embedded inference unit uploads the fault root cause location diagnosis results output in step 5 to the ship's engine room PLC through the Modbus RTU standardized interaction interface. The ship's engine room PLC then pushes the results to the engine room monitoring system to complete the visual display of the fault alarm. When the embedded inference unit detects a major equipment failure, it sends low-level equipment control commands to the ship's engine room PLC through the Modbus RTU standardized interactive interface. The ship's engine room PLC then drives the field equipment of the marine air conditioning system to perform safe operations.

Images