Multi-source information collaborative early warning system for ship key equipment

Abstract

The invention provides a ship key equipment multi-source information collaborative early warning system, and relates to the technical field of ship equipment monitoring, and the system comprises a data collection module which is used for collecting multi-source information of ship key equipment; the data preprocessing module is used for preprocessing the multi-source information, including time alignment processing and standardization processing; the feature extraction module is used for extracting time domain features, frequency domain features and time frequency features from the preprocessed multi-source information; the information fusion module is used for carrying out feature layer fusion and decision-making layer fusion on the features to obtain fused features; the early warning decision module is used for carrying out early warning analysis according to the fusion features to obtain an early warning result; and the man-machine interaction module is used for displaying and transmitting the early warning result. According to the invention, comprehensive monitoring, accurate early warning and intelligent diagnosis of the operation state of the ship key equipment can be realized, so that the accuracy, timeliness and reliability of equipment fault early warning are improved.

Description

Technical Field

[0001] This invention relates to the field of ship equipment monitoring technology, and in particular to a multi-source information collaborative early warning system for key ship equipment. Background Technology

[0002] Key ship equipment is a crucial foundation for ensuring safe navigation, and mainly includes core equipment such as power systems, propulsion systems, and auxiliary machinery. These devices operate for extended periods in the complex marine environment, enduring multiple stresses including high loads, high temperatures, high pressures, vibrations, and corrosion. Their operational status directly affects the ship's navigational safety and economic efficiency; any malfunction could lead to serious safety accidents and economic losses.

[0003] Currently, the condition monitoring of critical ship equipment mainly adopts single-signal-source monitoring methods, such as vibration monitoring, temperature monitoring, and pressure monitoring. These monitoring systems typically operate independently, using fixed warning thresholds for fault diagnosis. Some systems are also equipped with basic data analysis functions to assess the operating status of the equipment and predict potential failures.

[0004] However, existing monitoring and early warning technologies have the following problems: First, monitoring using a single signal source is insufficient to fully reflect the actual operating status of the equipment, which can easily lead to misjudgments or missed judgments; second, the monitoring systems are independent of each other, making it impossible to achieve information sharing and collaborative analysis, thus reducing the accuracy of fault diagnosis; third, fixed early warning thresholds cannot adapt to the dynamic changes of equipment under different operating conditions and environments, affecting the timeliness and reliability of early warnings; and finally, existing systems do not make high use of historical equipment data, making it difficult to achieve intelligent fault prediction. Summary of the Invention

[0005] In view of this, the present invention proposes a multi-source information collaborative early warning system for key ship equipment. By collecting and processing equipment information from multiple signal sources, including vibration, temperature, pressure, current, sound, and oil, and using a dynamic time warping algorithm to achieve signal alignment, combined with multi-level information fusion at the feature layer and decision layer and an adaptive threshold adjustment mechanism, the system aims to achieve comprehensive monitoring, accurate early warning, and intelligent diagnosis of the operating status of key ship equipment, thereby improving the accuracy, timeliness, and reliability of equipment fault early warning.

[0006] The technical solution of this invention is implemented as follows:

[0007] This invention provides a multi-source information collaborative early warning system for critical ship equipment, comprising:

[0008] The data acquisition module is used to collect multi-source information from key ship equipment, including at least two of the following: vibration signal, temperature signal, pressure signal, current signal, sound signal, and oil signal.

[0009] The data preprocessing module is used to preprocess the multi-source information, including time alignment and standardization.

[0010] The feature extraction module is used to extract time-domain features, frequency-domain features, and time-frequency features from preprocessed multi-source information;

[0011] The information fusion module is used to perform feature-level fusion and decision-level fusion on the time-domain features, frequency-domain features, and time-frequency features to obtain fused features;

[0012] The early warning decision module is used to perform early warning analysis based on the fused features and obtain early warning results; the early warning decision module includes an adaptive threshold adjustment mechanism, which constructs and dynamically adjusts a multi-dimensional threshold space based on equipment operating conditions, operating time, and environmental stress.

[0013] The human-computer interaction module is used to display and transmit the warning results.

[0014] Based on the above scheme, preferably, the time alignment process employs a dynamic time warping algorithm, including:

[0015] The signal sequence with the highest sampling frequency is selected as the reference signal R and its timestamp sequence T. r ;

[0016] The remaining signal sequence is denoted as the signal to be aligned {S1, S2, ..., S...} A} and its corresponding timestamp sequence {T s1 ,T s2 ,...,T sA}, where A is the number of signals to be aligned;

[0017] For each signal S to be aligned a Execute in sequence:

[0018] The cumulative distance matrix D(i,j) between the reference signal and the target signal is calculated using the following formula:

[0019] D(i,j)=d(s i ,r j )+min{D(i-1,j-1),D(i-1,j),D(i,j-1)}

[0020] Where i = 1, 2, ..., n, and n is the signal S to be aligned. a The length of the reference signal R; j = 1, 2, ..., m, where m is the length of the reference signal R; s i Signal S to be aligned a At its timestamp t si The amplitude at that point; r jThe reference signal R at its timestamp t rj The amplitude at that point; d(s) i ,r j () represents the Euclidean distance;

[0021] The optimal regularized path W = {w1, w2, ..., w...} is determined based on the minimum cumulative distance. k}, where each w k Contains two time indices (i k ,j k ), i k Indicates the signal S to be aligned a The sampling time sequence number, j k This indicates the sampling time number that best matches the reference signal R, where k is the path length;

[0022] Establish a timestamp mapping relationship M(t) based on the optimal regularization path. si )=t r , where t si t r The signals to be aligned are S a and the timestamp of the reference signal R;

[0023] Based on the mapping relationship M, cubic spline interpolation is used to align the signal S to be aligned. a The timestamp sequence T of the resampled reference signal r The aligned signal sequence S is obtained from the above. a ';

[0024] After time alignment of all signals, a multi-source signal sequence {R,S1',S2',...,S} is formed under a unified time reference. A '}.

[0025] Based on the above scheme, preferably, in the dynamic time warping algorithm, the initial value of the cumulative distance matrix D(i,j) is set as follows:

[0026] When i = 0 or j = 0, D(i,j) takes a preset large constant or infinite value to ensure that the path can only start normally from D(1,1) during actual calculation;

[0027] Euclidean distance d(s) i ,r j The calculation of ) is limited to the square root of the difference between their respective amplitudes, and the maximum allowable step size between sampling points is limited to not exceeding the set threshold L.

[0028] Based on the above scheme, preferably, the time-domain features include mean, variance, peak value, skewness, and kurtosis; the frequency-domain features include power spectral density and band energy ratio; and the time-frequency features include wavelet packet decomposition coefficients.

[0029] Based on the above scheme, preferably, the information fusion module includes:

[0030] The feature layer fusion unit is used to construct a feature importance evaluation matrix for the time-domain features, frequency-domain features and time-frequency features, and calculate the feature weight coefficients using the mutual information entropy criterion to obtain weighted features;

[0031] The decision-level fusion unit is used to construct evidence from the weighted features output by the feature-level fusion. It uses an improved DS evidence theory to handle evidence conflicts and generates fused features through trust allocation and evidence combination.

[0032] Based on the above scheme, preferably, the construction process of the feature importance evaluation matrix is ​​as follows:

[0033] Based on the sampled values ​​at different times for the time-domain features, frequency-domain features, and time-frequency features, an original feature matrix X = [x i,t ]∈R (N×T) , where N represents the total number of features and T represents the number of time segments selected;

[0034] Calculate the temporary importance value 'a' of each feature at different time segments. i,t :

[0035]

[0036] In the formula, τ is the time index variable; dec(·) is the time decay function, dec(t-τ) is used to measure the decay of importance over time interval t-τ; corr(·) is the correlation function, corr(x i,τ ,x i,t (b) is used to measure the feature similarity of the same feature at different times; i,τ This is represented as the time-varying trade-off factor preset for feature i;

[0037] For the importance of the temporary quantity a i,t After performing row and column normalization, the final feature importance evaluation matrix Z = [z i,t ]∈R (N×t) :

[0038]

[0039] in,

[0040] Based on the above scheme, preferably, after the feature layer fusion unit normalizes the importance of each feature according to the feature importance evaluation matrix M, it determines the feature weight coefficient ω using the following calculation formula. i :

[0041]

[0042] In the formula, γ, α, β, and λ are pre-defined parameters; z i,t Δ represents the importance value of the i-th feature at time t; i,t Let represent the variance of the i-th feature at time t; T is the number of time segments selected. Represents the average importance value of the i-th feature across all time segments; |·| represents the absolute value operation; e (·) This indicates exponentiation.

[0043] Based on the above scheme, preferably, when the decision-level fusion unit generates fusion features based on the improved DS evidence theory, it includes the following steps:

[0044] Step 1: Create the initial evidence set and construct the basic probability assignment function:

[0045] Based on a predefined set of fault modes Θ = {θ1, θ2, ..., θ g} and the weighted features from the feature layer fusion output, to establish an evidence set E = {e1, e2, ..., e h}, where each piece of evidence e k Corresponding to the corresponding weighted feature set;

[0046] For each piece of evidence e k According to its relationship with each failure mode θ s Match score (e) k ,θ s Calculate the basic probability assignment function m k (G):2 Θ →[0,1], where G is a subset of failure modes, satisfying the following condition:

[0047]

[0048] If score(e k ,θ s If θ is greater than a preset threshold, then θ s The corresponding single-element set {θ s} is considered a focal element, and other mode sets are considered composite focal elements, according to score(e k ,θ s Weighted allocation is performed;

[0049] Step 2: Define the time-varying conflict correction function:

[0050] Set the conflict metric K t (·), used to measure the degree of conflict between two or more pieces of evidence at time t;

[0051] Before evidence synthesis, a time-varying correction factor δ is introduced for evidence items whose conflict measure exceeds a preset threshold. t Time-segmented error compensation and reallocation are performed, and the expression is as follows:

[0052]

[0053] Where κ is an adjustable parameter, K t The larger δ is, or the smaller t is. t The closer it is to 1, the stronger the redistribution of conflicting evidence at that moment;

[0054] Step 3: Construct improved DS combination rules:

[0055] Regarding the two pieces of evidence e k and e d During synthesis, the synthesized basic probability assignment function m'(G) is expressed as:

[0056]

[0057] Where: κ′=∑[m k (B)·m d (C)·(1-ρ(e k ,e d ))], This indicates the difference between the current time t and the evidence generation time t. b The weights that decay between each other are used to reduce the weight of expired evidence; ρ(e k ,e d ) is used to quantify the independence of two pieces of evidence, if ρ(e k ,e d A smaller value indicates a higher degree of independence of evidence and a weaker impact of conflict.

[0058] Step 4: Normalization and Uncertainty Allocation

[0059] Once all the evidence has been synthesized, the synthesized result is normalized.

[0060] An uncertainty factor u(G) is introduced to measure the balance between confidence in set G and unassigned residual confidence:

[0061] u(G)=1-∑[m′(G)]

[0062] Where m'(G) represents the final basic probability assignment of G, and the closer u(G) is to 1, the lower the degree of certainty about the set G;

[0063] Finally, the synthesized evidence result m'(·) and the uncertainty factor u(·) are output together as the fusion feature.

[0064] Based on the above scheme, preferably, the early warning decision module adopts an adaptive threshold adjustment mechanism, which includes:

[0065] The equipment operating environment, historical fault records, and aging status information of key components are mapped onto a multi-dimensional coordinate axis, and a threshold coordinate set is formed according to the influence of different dimensions on threshold sensitivity.

[0066] The initial warning threshold vector is obtained by performing linear weighting or genetic algorithm optimization on the threshold coordinate set;

[0067] During real-time equipment monitoring, the threshold vector is adaptively corrected by combining the dynamic changes of fusion features, so that the warning point can be quickly adjusted according to the fluctuations of equipment operating conditions and environmental stress.

[0068] Based on the above scheme, preferably, the method for adaptively correcting the threshold vector includes:

[0069] Set the early warning threshold adjustment period P and adjustment step size α;

[0070] Within each adjustment period, calculate the deviation between the fused feature and the current threshold;

[0071] When the deviation value exceeds the preset range n times consecutively, the threshold will be adjusted by a step size α in the direction of the deviation.

[0072] Set the upper and lower limits of the threshold change [θ] min ,θ max This ensures that the corrected threshold remains within that range.

[0073] The present invention has the following advantages over the prior art:

[0074] (1) By constructing a multi-source information collaborative early warning system, this invention realizes the collaborative acquisition and fusion analysis of various signals such as vibration, temperature, pressure, current, sound, and oil of key ship equipment. Combined with multi-level information fusion of the feature layer and decision layer and the adaptive threshold adjustment mechanism, it significantly improves the comprehensiveness of equipment status monitoring and the accuracy of early warning, and provides reliable technical support for early detection of equipment failure and preventive maintenance.

[0075] (2) The dynamic time warping algorithm is used for time alignment. By calculating the minimum cumulative distance and the optimal warping path, the precise alignment of signals with different sampling frequencies is achieved, which solves the problem of asynchronous multi-source heterogeneous data in the time dimension.

[0076] (3) Based on the feature layer fusion scheme of feature importance evaluation matrix and mutual information entropy criterion, combined with time-varying trade-off factor and correlation function, adaptive weighting of different features is realized, which effectively improves the accuracy and representativeness of feature extraction and enhances the system's ability to identify abnormal equipment status.

[0077] (4) The improved DS evidence theory introduces a time-varying conflict correction function and an evidence independence assessment mechanism. Through dynamic compensation and redistribution of conflicting evidence, it improves the reliability of multi-source information fusion and effectively reduces the misjudgment rate caused by evidence conflict.

[0078] (5) The adaptive threshold adjustment mechanism realizes the adaptive adjustment of the warning threshold according to the equipment operating conditions and environmental stress through the construction and dynamic optimization of the multi-dimensional threshold space, which improves the system's sensitivity to abnormal equipment conditions and ensures the stability and reliability of the warning results. Attached Figure Description

[0079] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0080] Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0081] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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.

[0082] like Figure 1 As shown, the present invention provides a multi-source information collaborative early warning system for key ship equipment, comprising:

[0083] The data acquisition module is used to collect multi-source information from key ship equipment, including at least two of the following: vibration signal, temperature signal, pressure signal, current signal, sound signal, and oil signal.

[0084] The data preprocessing module is used to preprocess the multi-source information, including time alignment and standardization.

[0085] The feature extraction module is used to extract time-domain features, frequency-domain features, and time-frequency features from preprocessed multi-source information;

[0086] The information fusion module is used to perform feature-level fusion and decision-level fusion on the time-domain features, frequency-domain features, and time-frequency features to obtain fused features;

[0087] The early warning decision module is used to perform early warning analysis based on the fused features and obtain early warning results; the early warning decision module includes an adaptive threshold adjustment mechanism, which constructs and dynamically adjusts a multi-dimensional threshold space based on equipment operating conditions, operating time, and environmental stress.

[0088] The human-computer interaction module is used to display and transmit the warning results.

[0089] Specifically, in one embodiment of the present invention, the data acquisition module adopts a distributed data acquisition architecture, achieving collaborative acquisition of multi-source information by arranging multiple types of sensors at key parts of the host. Specifically, a triaxial accelerometer is installed at the bearing housing of the host, with a sampling frequency of 10kHz, to acquire vibration signals of the equipment; PT100 temperature sensors are installed at key parts such as cylinder liners, bearings, and the engine block, with a sampling frequency of 1Hz, to acquire temperature signals; pressure sensors are installed at the inlet and outlet of the lubricating oil system and fuel system, with a sampling frequency of 100Hz, to acquire pressure signals. Regarding electrical system monitoring, a Hall current sensor is installed on the power supply line of the host, with a sampling frequency of 1kHz, to acquire current signals; an acoustic sensor array is arranged around the host, with a sampling frequency of 44.1kHz, to acquire sound signals; and an online oil monitoring device, including a ferrometer and an oil particle counter, is installed in the lubricating oil circuit, with a sampling frequency of 0.1Hz, to acquire oil characteristic signals. All sensors are connected to the nearest data acquisition unit via shielded cables, and a 24-bit AD converter is used for analog-to-digital conversion to ensure the accuracy of the acquired signals. The data acquisition module uses industrial Ethernet as the communication bus to transmit the distributed data to the central processing unit. Each data acquisition unit is equipped with a local buffer, using a circular buffer to store the raw data from the most recent 30 minutes. Simultaneously, a CRC check mechanism is employed during data transmission to ensure reliability. The raw data received by the central processing unit is tagged with a uniform timestamp and stored in binary format on a solid-state drive, providing data support for subsequent data preprocessing and analysis. During acquisition, the system automatically records the sensor's operating status, and when an anomaly is detected, it issues a prompt to the operator through the human-machine interface module.

[0090] Specifically, in one embodiment of the present invention, the time alignment process employs a dynamic time warping algorithm, including:

[0091] The signal sequence with the highest sampling frequency is selected as the reference signal R and its timestamp sequence T. r ;

[0092] The remaining signal sequence is denoted as the signal to be aligned {S1, S2, ..., S...} A} and its corresponding timestamp sequence {T s1 ,T s2 ,...,T sA}, where A is the number of signals to be aligned;

[0093] For each signal S to be aligned a Execute in sequence:

[0094] The cumulative distance matrix D(i,j) between the reference signal and the target signal is calculated using the following formula:

[0095] D(i,j)=d(s i ,r j )+min{D(i-1,j-1),D(i-1,j),D(i,j-1)}

[0096] Where i = 1, 2, ..., n, and n is the signal S to be aligned. a The length of the reference signal R; j = 1, 2, ..., m, where m is the length of the reference signal R; s i Signal S to be aligned a At its timestamp t si The amplitude at that point; r j The reference signal R at its timestamp t rj The amplitude at that point; d(s) i ,r j () represents the Euclidean distance;

[0097] The optimal regularized path W = {w1, w2, ..., w...} is determined based on the minimum cumulative distance. k}, where each w k Contains two time indices (i k ,j k ), i k Indicates the signal S to be aligned a The sampling time sequence number, j k This indicates the sampling time number that best matches the reference signal R, where k is the path length;

[0098] Establish a timestamp mapping relationship M(t) based on the optimal regularization path. si )=t r , where t si t r The signals to be aligned are S a and the timestamp of the reference signal R;

[0099] Based on the mapping relationship M, cubic spline interpolation is used to align the signal S to be aligned. a The timestamp sequence T of the resampled reference signal r The aligned signal sequence S is obtained from the above. a ';

[0100] After time alignment of all signals, a multi-source signal sequence {R,S1',S2',...,S} is formed under a unified time reference. A '}.

[0101] In the dynamic time warping algorithm, the initial value of the cumulative distance matrix D(i,j) is set as follows:

[0102] When i = 0 or j = 0, D(i,j) takes a preset large constant or infinite value to ensure that the path can only start normally from D(1,1) during actual calculation;

[0103] Euclidean distance d(s) i ,r j The calculation of ) is limited to the square root of the difference between their respective amplitudes, and the maximum allowable step size between sampling points is limited to not exceeding the set threshold L.

[0104] In this embodiment, a sound signal with a sampling frequency of 44.1 kHz is selected as the reference signal R, and other signals including vibration signal (10 kHz), current signal (1 kHz), pressure signal (100 Hz), temperature signal (1 Hz) and oil signal (0.1 Hz) are selected as the signal sequence to be aligned {S1, S2, ..., S5}.

[0105] When performing the dynamic time warping algorithm, the cumulative distance matrix D(i,j) is first initialized. When i=0 or j=0, D(i,j) is set to 10. 6 This ensures that the path can only start from D(1,1) during actual calculation. When calculating the Euclidean distance d(s... i ,r j When ), the signal to be aligned, S, is used. i At time t si The amplitude of the reference signal R at time t rj The square root of the difference in amplitude. To avoid unreasonable time jumps, the maximum allowable jump step size between sampling points is set to L = 3, that is, the time index difference between adjacent matching points shall not exceed 3 sampling points.

[0106] For each signal to be aligned, the system sequentially performs a dynamic normalization process. Taking vibration signal S1 as an example, the cumulative distance matrix D(i,j) between it and the reference signal R is first calculated. During the calculation, each matrix element D(i,j) is based on the Euclidean distance d(s) of the current point. i ,r jThe minimum value among the three preceding directions (D(i-1,j-1), D(i-1,j), D(i,j-1)) is updated. The optimal regularized path W is determined by backtracking the path with the minimum value in the cumulative distance matrix. This path contains a series of time index pairs (i... k ,j k ), used to establish the mapping relationship M(t) between the sampling time of the vibration signal and the sampling time of the reference signal. si )=t r Finally, based on this mapping relationship, the vibration signal is resampled onto the timestamp sequence of the reference signal using cubic spline interpolation to obtain the aligned signal sequence S1'. Other signals undergo the same processing in sequence, ultimately yielding a multi-source signal sequence {R,S1',S2',S3',S4',S5'} under a unified time reference.

[0107] After time alignment, all signals are standardized using the Z-score standardization method. For each signal sequence X, its mean μ and standard deviation σ are calculated, and then the original data are transformed into standardized data with a mean of 0 and a standard deviation of 1 using the formula Z = (X - μ) / σ.

[0108] Specifically, in one embodiment of the present invention, the time-domain features include mean, variance, peak value, skewness, and kurtosis; the frequency-domain features include power spectral density and band energy ratio; and the time-frequency features include wavelet packet decomposition coefficients.

[0109] In this embodiment, the feature extraction module extracts time-domain, frequency-domain, and time-frequency features from the preprocessed multi-source information. For each signal source, the system extracts a corresponding feature set to comprehensively reflect the operating status characteristics of the equipment. In terms of time-domain feature extraction, the system calculates five statistical features: mean reflects the overall signal level and is used to monitor the stability of the equipment's operating status; variance characterizes the degree of signal fluctuation and can be used to assess the vibration intensity of the equipment; peak value is used to detect instantaneous impacts and has a good indicative effect on abnormal impacts on the equipment; skewness describes the asymmetry of the signal distribution and can reflect the nonlinear characteristics of the equipment's operation; kurtosis characterizes the impulse characteristics of the signal and is more sensitive to early failures of components such as bearings. In terms of frequency-domain feature extraction, the system first performs a Fast Fourier Transform (FFT) on the signal to obtain its spectrum. Then, it calculates the power spectral density to analyze the distribution characteristics of the signal energy in the frequency domain, which can effectively identify the characteristic frequency components of the equipment. Simultaneously, the frequency band is divided into a low-frequency band (0-1 / 3 times the critical speed), a mid-frequency band (1 / 3-2 / 3 times the critical speed), and a high-frequency band (above 2 / 3 times the critical speed), and the ratio of energy in each frequency band to the total energy is calculated, i.e., the frequency band energy ratio. For time-frequency feature extraction, wavelet packet decomposition is used for multi-scale analysis of the signal. First, a suitable wavelet basis function (such as the db4 wavelet) is selected to decompose the signal to the third level, obtaining wavelet packet coefficients for eight frequency bands. These coefficients can simultaneously reflect the local characteristics of the signal in both the time and frequency domains. For different signal sources, the system selects appropriate feature combinations based on their physical characteristics. For example, for vibration signals, all time-domain features, frequency-domain features, and time-frequency features are extracted; for temperature signals, the focus is mainly on time-domain features such as mean and variance; for oil signals, the emphasis is on extracting statistical features reflecting particle characteristics.

[0110] Specifically, in one embodiment of the present invention, the information fusion module includes:

[0111] The feature layer fusion unit is used to construct a feature importance evaluation matrix for the time-domain features, frequency-domain features and time-frequency features, and calculate the feature weight coefficients using the mutual information entropy criterion to obtain weighted features;

[0112] The decision-level fusion unit is used to construct evidence from the weighted features output by the feature-level fusion. It uses an improved DS evidence theory to handle evidence conflicts and generates fused features through trust allocation and evidence combination.

[0113] In this embodiment, the construction process of the feature importance evaluation matrix is ​​as follows:

[0114] Based on the sampled values ​​at different times for the time-domain features, frequency-domain features, and time-frequency features, an original feature matrix X = [x i,t ]∈R (N×T), where N represents the total number of features and T represents the number of time segments selected.

[0115] Specifically, the original feature matrix X is constructed from the sampled values ​​of N=24 features (including the time domain, frequency domain, and time-frequency features of each signal source) over T=60 time segments.

[0116] Calculate the temporary importance value 'a' of each feature at different time segments. i,t :

[0117]

[0118] In the formula, τ is the time index variable; dec(·) is the time decay function, dec(t-τ) is used to measure the decay of importance over time interval t-τ; corr(·) is the correlation function, corr(x i,τ ,x i,t (b) is used to measure the feature similarity of the same feature at different times; i,τ This is represented as the time-varying trade-off factor preset for feature i.

[0119] Specifically, the time decay function dec(·) can be expressed in an exponential decay form: dec(t-τ)=e -η(t-τ) Where η is the attenuation coefficient, with a value range of [0.1, 0.3], and in this embodiment, η = 0.2 is used. The correlation function corr(·) is calculated using the Pearson correlation coefficient method; the time-varying tradeoff factor b i,τ Adopting an adaptive adjustment method based on equipment operating conditions: b i,τ =0.4 + 0.4·load τ , where load τ This refers to the equipment load rate.

[0120] For the importance of the temporary quantity a i,t After performing row and column normalization, the final feature importance evaluation matrix Z = [z i,t ]∈R (N×t) :

[0121]

[0122] in,

[0123] The feature layer fusion unit normalizes the importance of each feature according to the feature importance evaluation matrix M, and then determines the feature weight coefficient ω using the following formula. i :

[0124]

[0125] In the formula, γ, α, β, and λ are pre-defined parameters; zi,t Δ represents the importance value of the i-th feature at time t; i, t represents the variance of the i-th feature at time t; T is the number of time segments selected; Represents the average importance value of the i-th feature across all time segments; |·| represents the absolute value operation; e (·) This indicates exponentiation.

[0126] Specifically, γ∈[0.5,1.5] is used to adjust the sensitivity of importance, and in this embodiment, γ = 1.0; α∈[0.3,0.7] is used to balance the effect of variance, and α = 0.5; β∈[1.0,2.0] is used to control the degree of nonlinearity, and β = 1.5; λ∈[0.1,0.3] is used to adjust the effect of periodicity, and λ = 0.2. These parameters can be optimized and determined on historical data using cross-validation methods.

[0127] In this embodiment, when the decision-level fusion unit generates fusion features based on the improved DS evidence theory, it includes the following steps:

[0128] Step 1: Create an initial evidence set and construct a basic probability assignment function:

[0129] Based on a predefined set of fault modes Θ = {θ1, θ2, ..., θ g} and the weighted features from the feature layer fusion output, to establish an evidence set E = {e1, e2, ..., e h}, where each piece of evidence e k Corresponding to the corresponding weighted feature set.

[0130] Specifically, in one example, the set of failure modes is defined as Θ = {normal, mildly abnormal, severely abnormal}.

[0131] For each piece of evidence e k According to its relationship with each failure mode θ s Match score (e) k ,θ s Calculate the basic probability assignment function m k (G):2 Θ →[0,1], where G is a subset of failure modes, satisfying the following condition:

[0132]

[0133] If score(e k ,θ s If θ is greater than a preset threshold, then θ s The corresponding single-element set {θ s} is considered a focal element, and other mode sets are considered composite focal elements, according to score(ek ,θ s Weighted allocation is performed.

[0134] Specifically, the matching score (e k ,θ s The calculation formula is:

[0135]

[0136] In the formula, d(e k ,θ s σ is the Mahalanobis distance from the evidence vector to the class center. s This represents the within-class sample standard deviation.

[0137] Basic probability assignment function m k The calculation can be performed using a distance-based soft allocation method: m k (θ s ) = score(e k ,θ s ) / ∑score(e k ,θ s For a composite focal element F, its fundamental probability assignment is calculated through the combination of its constituent focal elements: m k (F)=∏m k (θ s ),θ s ∈F.

[0138] Step 2: Define the time-varying conflict correction function:

[0139] Set the conflict metric K t (·), used to measure the degree of conflict between two or more pieces of evidence at time t;

[0140] Before evidence synthesis, a time-varying correction factor δ is introduced for evidence items whose conflict measure exceeds a preset threshold. t Time-segmented error compensation and reallocation are performed, and the expression is as follows:

[0141]

[0142] Where k is an adjustable parameter, K t The larger δ is, or the smaller t is. t The closer it is to 1, the stronger the redistribution of conflicting evidence at that moment.

[0143] Specifically, the value of κ ranges from [0.5, 1.5], and in this embodiment, κ = 1.0 is used.

[0144] Step 3: Construct improved DS combination rules:

[0145] Regarding the two pieces of evidence e kand e d During synthesis, the synthesized basic probability assignment function m'(G) is expressed as:

[0146]

[0147] Where: κ′=∑[m k (B)·m d (C)·(1-ρ(e k ,e d ))], This indicates the difference between the current time t and the evidence generation time t. b The weights that decay between each other are used to reduce the weight of expired evidence; ρ(e k ,e d ) is used to quantify the independence of two pieces of evidence, if ρ(e k ,e d A smaller value indicates a higher degree of independence of evidence and a weaker impact of conflict.

[0148] Specifically, Exponential decay can be used: Where μ is the time decay coefficient, with a value range of [0.05, 0.15], and in this embodiment, μ = 0.1. ρ(e k ,e d The calculation method is based on mutual information and conditional entropy: ρ(e k ,e d )=1-I(e k ;e d ) / min{H(e k ),H(e d )}, where I(e k ;e d H(·) represents mutual information, and H(·) represents information entropy.

[0149] Step 4: Normalization and Uncertainty Allocation

[0150] Once all the evidence has been synthesized, the synthesized result is normalized.

[0151] An uncertainty factor u(G) is introduced to measure the balance between confidence in set G and unassigned residual confidence:

[0152] u(G)=1-∑[m′(G)]

[0153] Where m'(G) represents the final basic probability assignment of G, and the closer u(G) is to 1, the lower the degree of certainty about the set G;

[0154] Finally, the synthesized evidence result m'(·) and the uncertainty factor u(·) are output together as the fusion feature. When the uncertainty u(G) exceeds a preset threshold (e.g., 0.3), the system will trigger a data quality check mechanism, requiring the relevant data to be re-collected or verified.

[0155] Specifically, in one embodiment of the present invention, the early warning decision module employs an adaptive threshold adjustment mechanism, which includes:

[0156] The equipment operating environment, historical fault records, and aging status information of key components are mapped onto a multi-dimensional coordinate axis, and a threshold coordinate set is formed according to the influence of different dimensions on threshold sensitivity.

[0157] The initial warning threshold vector is obtained by performing linear weighting or genetic algorithm optimization on the threshold coordinate set;

[0158] During real-time equipment monitoring, the threshold vector is adaptively corrected by combining the dynamic changes of fusion features, so that the warning point can be quickly adjusted according to the fluctuations of equipment operating conditions and environmental stress.

[0159] The method for adaptively correcting the threshold vector includes:

[0160] Set the early warning threshold adjustment period P and adjustment step size α;

[0161] Within each adjustment period, calculate the deviation between the fused feature and the current threshold;

[0162] When the deviation value exceeds the preset range n times consecutively, the threshold will be adjusted by a step size α in the direction of the deviation.

[0163] Set the upper and lower limits of the threshold change [θ] min ,θ max This ensures that the corrected threshold remains within that range.

[0164] In a specific example, a five-dimensional coordinate system is first constructed, including the dimensions of equipment operating time (T), ambient temperature (E), load rate (L), vibration intensity (V), and oil contamination level (O). Specifically, equipment operating time (T) is normalized by dividing the cumulative operating hours of the equipment by the designed lifespan hours, with a value range of [0,1]. Ambient temperature (E) is mapped to the [0,1] interval using the min-max normalization method, with a reference temperature range of -10℃ to 45℃. Load rate (L) is directly calculated as the ratio of engine output power to rated power. Vibration intensity (V) is obtained by dividing the root mean square of the RMS values ​​at each measuring point of the equipment by the reference value (2.5 mm / s). Oil contamination level (O) is mapped to the [0,1] interval using particle count results according to the NAS1638 classification standard.

[0165] In the process of forming the threshold coordinate set, the influence function of each dimension on the threshold is first established: f(T) = 1 + 0.3T 2 f(E) = 1 + 0.2|E - 0.5|, reflecting the effect of temperature deviation from the optimal operating range on the threshold; f(L) = 1 + 0.25L, reflecting the effect of load on the threshold; f(V) = 1 + 0.4V, reflecting the effect of vibration intensity on the threshold; f(O) = 1 + 0.35O, reflecting the effect of oil condition on the threshold.

[0166] A genetic algorithm was used to optimize the weights of each dimension. The population size was set to 100, the number of iterations was 200, the crossover probability was 0.8, and the mutation probability was 0.1. The fitness function was based on the early warning accuracy of historical fault data. The optimized weight coefficients for each dimension were obtained: w T =0.25, w E =0.15, w L =0.20, w V =0.25, w O =0.15. The initial warning threshold vector θ0 is calculated using the following formula:

[0167] θ0=θ base ·∑(w i ·f(i)); where θ base The baseline threshold is determined by the 3σ principle based on the normal operating data of the equipment.

[0168] During the adaptive correction process, the early warning threshold adjustment period is set to P = 24h, and the basic adjustment step size is α = 0.02. For the fused feature F(t), its deviation from the current threshold θ(t) is calculated:

[0169] δ(t)=[F(t)-θ(t)] / θ(t);

[0170] When |δ(t)| exceeds the preset range [-0.1, 0.1] for n = 5 consecutive times, the threshold is adjusted as follows:

[0171] θ(t+1)=θ(t)·(1±α·|δ(t)|); where, when δ(t)>0, the plus sign is taken, and when δ(t)<0, the minus sign is taken.

[0172] The upper and lower limits of the threshold change are determined using historical operating data of the equipment:

[0173] θ min =θ base ·0.7; θ max =θ base ·1.5.

[0174] Meanwhile, a threshold change rate limit is introduced: |θ(t+1)-θ(t)| / θ(t)≤0.05.

[0175] Specifically, in one embodiment of the present invention, the human-computer interaction module adopts a B / S architecture design, enabling remote access and control of the system through a web browser. This module mainly includes four functional units: data display, early warning management, system configuration, and report management.

[0176] The data display unit adopts a partitioned layout, simultaneously displaying real-time data and trend charts from multiple sources on the main interface. The left area displays basic equipment information and operating status, including key parameters such as cumulative operating time, current load rate, and ambient temperature; the central area uses a multi-pane layout to display real-time waveforms of multiple sources such as vibration, temperature, pressure, current, sound, and oil, supporting interactive operations such as waveform zooming and data cursor; the right area displays trend charts for each signal and supports viewing historical data within a custom time range.

[0177] The early warning management unit employs a hierarchical display mechanism. When the system generates an early warning, a prominent warning message will be displayed in the status bar at the top of the interface. Different levels of warnings are distinguished by different colors: red indicates a severe anomaly, yellow indicates a minor anomaly, and green indicates a normal state. Clicking on the warning information will display detailed information such as the warning reason, trigger time, and relevant parameters. The system supports warning confirmation, allowing operators to add handling suggestions. All warning records are archived for subsequent analysis.

[0178] The system configuration unit provides a parameter setting interface, including functions such as sampling parameter configuration, early warning threshold adjustment, and algorithm parameter optimization. To prevent accidental operation, modification of key parameters requires advanced permission verification. Modification history is automatically recorded after parameter changes, and parameter rollback is supported. This unit also provides a system self-check function, which can diagnose system health conditions such as sensor status and communication link quality.

[0179] The report management unit supports data export and report generation in multiple formats. Data can be filtered by time range, equipment type, data type, and other criteria to generate equipment operation reports, early warning statistics reports, maintenance suggestion reports, etc. Reports can be exported in PDF, Excel, and other formats and can be automatically sent to relevant personnel via email.

[0180] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-source information collaborative early warning system for key ship equipment, characterized in that, include: The data acquisition module is used to collect multi-source information from key ship equipment, including at least two of the following: vibration signal, temperature signal, pressure signal, current signal, sound signal, and oil signal. The data preprocessing module is used to preprocess the multi-source information, including time alignment and standardization. The feature extraction module is used to extract time-domain features, frequency-domain features, and time-frequency features from preprocessed multi-source information; The information fusion module is used to perform feature-level fusion and decision-level fusion on the time-domain features, frequency-domain features, and time-frequency features to obtain fused features; The early warning decision module is used to perform early warning analysis based on the fused features and obtain early warning results. The early warning decision module includes an adaptive threshold adjustment mechanism, which constructs and dynamically adjusts a multi-dimensional threshold space based on equipment operating conditions, running time, and environmental stress. The human-computer interaction module is used to display and transmit the warning results.

2. The multi-source information collaborative early warning system for key ship equipment as described in claim 1, characterized in that, The time alignment process employs a dynamic time warping algorithm, including: The signal sequence with the highest sampling frequency is selected as the reference signal R and its timestamp sequence T. r ; The remaining signal sequence is denoted as the signal to be aligned {S1, S2, ..., S...} A } and its corresponding timestamp sequence {T s1 ,T s2 ,...,T sA }, where A is the number of signals to be aligned; For each signal S to be aligned a Execute in sequence: The cumulative distance matrix D(i,j) between the reference signal and the target signal is calculated using the following formula: D(i,j)=d(s i ,r j )+min{D(i-1,j-1),D(i-1,j),D(i,j-1)} Where i = 1, 2, ..., n, and n is the signal S to be aligned. a The length of the reference signal R; j = 1, 2, ..., m, where m is the length of the reference signal R; s i Signal S to be aligned a At its timestamp t si The amplitude at that point; r j The reference signal R at its timestamp t rj The amplitude at that point; d(s) i ,r j () represents the Euclidean distance; The optimal regularized path W = {w1, w2, ..., w...} is determined based on the minimum cumulative distance. k }, where each w k Contains two time indices (i k ,j k ), i k Indicates the signal S to be aligned a The sampling time sequence number, j k This indicates the sampling time number that best matches the reference signal R, where k is the path length; Establish a timestamp mapping relationship M(t) based on the optimal regularization path. si )=t r , where t si t r The signals to be aligned are S a and the timestamp of the reference signal R; Based on the mapping relationship M, cubic spline interpolation is used to align the signal S to be aligned. a The timestamp sequence T of the resampled reference signal r The aligned signal sequence S is obtained from the above. a '; After time alignment of all signals, a multi-source signal sequence {R,S1',S2',...,S} is formed under a unified time reference. A '}.

3. The multi-source information collaborative early warning system for key ship equipment as described in claim 2, characterized in that, In the dynamic time warping algorithm, the initial value of the cumulative distance matrix D(i,j) is set as follows: When i = 0 or j = 0, D(i,j) takes a preset large constant or infinite value to ensure that the path can only start normally from D(1,1) during actual calculation; Euclidean distance d(s) i ,r j The calculation of ) is limited to the square root of the difference between their respective amplitudes, and the maximum allowable step size between sampling points is limited to not exceeding the set threshold L.

4. The multi-source information collaborative early warning system for key ship equipment as described in claim 1, characterized in that, The time-domain features include mean, variance, peak value, skewness, and kurtosis; the frequency-domain features include power spectral density and band energy ratio; and the time-frequency features include wavelet packet decomposition coefficients.

5. The multi-source information collaborative early warning system for key ship equipment as described in claim 4, characterized in that, The information fusion module includes: The feature layer fusion unit is used to construct a feature importance evaluation matrix for the time-domain features, frequency-domain features and time-frequency features, and calculate the feature weight coefficients using the mutual information entropy criterion to obtain weighted features; The decision-level fusion unit is used to construct evidence from the weighted features output by the feature-level fusion. It uses an improved DS evidence theory to handle evidence conflicts and generates fused features through trust allocation and evidence combination.

6. The multi-source information collaborative early warning system for key ship equipment as described in claim 5, characterized in that, The construction process of the feature importance evaluation matrix is ​​as follows: Based on the sampled values ​​at different times for the time-domain features, frequency-domain features, and time-frequency features, an original feature matrix X = [x i,t ]∈R (N×T) , where N represents the total number of features and T represents the number of time segments selected; Calculate the temporary importance value 'a' of each feature at different time segments. i,t : In the formula, τ is the time index variable; dec(·) is the time decay function, and dec(t-τ) is used to measure the decay of importance over time interval t-τ; corr(·) is the correlation function, and corr(x) is the correlation function. i,τ ,x i,t This is used to measure the feature similarity of the same feature at different times; b i,τ This is represented as the time-varying trade-off factor preset for feature i; For the importance of the temporary quantity a i,t After performing row and column normalization, the final feature importance evaluation matrix Z = [z i,t ]∈R (N ×t) : in, 7. The multi-source information collaborative early warning system for key ship equipment as described in claim 6, characterized in that, The feature layer fusion unit normalizes the importance of each feature according to the feature importance evaluation matrix M, and then determines the feature weight coefficient ω using the following formula. i : In the formula, γ, α, β, and λ are preset parameters; z i,t Δ represents the importance value of the i-th feature at time t; i, t represents the variance of the i-th feature at time t; T is the number of time segments selected; Represents the average importance value of the i-th feature across all time segments; |·| represents the absolute value operation; e (·) This indicates exponentiation.

8. The multi-source information collaborative early warning system for key ship equipment as described in claim 5, characterized in that, When the decision-level fusion unit generates fusion features based on the improved DS evidence theory, it includes the following steps: Step 1: Create the initial evidence set and construct the basic probability assignment function: Based on a predefined set of fault modes Θ = {θ1, θ2, ..., θ g } and the weighted features from the feature layer fusion output, to establish an evidence set E = {e1, e2, ..., e h }, where each piece of evidence e k Corresponding to the corresponding weighted feature set; For each piece of evidence e k According to its relationship with each failure mode θ s Match score (e) k ,θ s Calculate the basic probability assignment function m k (G):2 Θ →[0,1], where G is a subset of failure modes, satisfying the following condition: If score(e k ,θ s If θ is greater than a preset threshold, then θ s The corresponding single-element set {θ s } is considered a focal element, and other mode sets are considered composite focal elements, according to score(e k ,θ s Weighted allocation is performed; Step 2: Define the time-varying conflict correction function: Set the conflict metric K t (·), used to measure the degree of conflict between two or more pieces of evidence at time t; Before evidence synthesis, a time-varying correction factor δ is introduced for evidence items whose conflict measure exceeds a preset threshold. t Time-segmented error compensation and reallocation are performed, and the expression is as follows: Where κ is an adjustable parameter, K t The larger δ is, or the smaller t is. t The closer it is to 1, the stronger the redistribution of conflicting evidence at that moment; Step 3: Construct improved DS combination rules: Regarding the two pieces of evidence e k and e d During synthesis, the synthesized basic probability assignment function m'(G) is expressed as: Where: κ′=∑[m k (B)·m d (C)·(1-ρ(e k ,e d ))], This indicates the difference between the current time t and the evidence generation time t. b The weights that decay between these values ​​are used to deweight expired evidence; ρ(e k ,e d ) is used to quantify the independence of two pieces of evidence, if ρ(e k ,e d A smaller value indicates a higher degree of independence of evidence and a weaker impact of conflict; Step 4: Normalization and Uncertainty Allocation Once all the evidence has been synthesized, the synthesized result is normalized. An uncertainty factor u(G) is introduced to measure the balance between confidence in set G and unassigned residual confidence: u(G)=1-∑[m′(G)] Where m'(G) represents the final basic probability assignment of G, and the closer u(G) is to 1, the lower the degree of certainty about the set G; Finally, the synthesized evidence result m'(·) and the uncertainty factor u(·) are output together as the fusion feature.

9. The multi-source information collaborative early warning system for key ship equipment as described in claim 8, characterized in that, The early warning decision module employs an adaptive threshold adjustment mechanism, which includes: The equipment operating environment, historical fault records, and aging status information of key components are mapped onto a multi-dimensional coordinate axis, and a threshold coordinate set is formed according to the influence of different dimensions on threshold sensitivity. The initial warning threshold vector is obtained by performing linear weighting or genetic algorithm optimization on the threshold coordinate set; During real-time equipment monitoring, the threshold vector is adaptively corrected by combining the dynamic changes of fusion features, so that the warning point can be quickly adjusted according to the fluctuations of equipment operating conditions and environmental stress.

10. The multi-source information collaborative early warning system for key ship equipment as described in claim 9, characterized in that, The method for adaptively correcting the threshold vector includes: Set the early warning threshold adjustment period P and adjustment step size α; Within each adjustment period, calculate the deviation between the fused feature and the current threshold; When the deviation value exceeds the preset range n times consecutively, the threshold will be adjusted by a step size α in the direction of the deviation. Set the upper and lower limits of the threshold change [θ] min ,θ max This ensures that the corrected threshold remains within that range.

Images