Cold and heat source center intelligent fault diagnosis system and state evaluation method thereof

Abstract

The invention discloses an intelligent fault diagnosis system for a cold and heat source center, and belongs to the technical field of fault judgment of heating, ventilation and air conditioning equipment. The system collects an operation feature vector X (t) in real time, and a state signal database stores data; a signal processing module of the upper computer subsystem adopts empirical mode decomposition X (t) to generate an IMF sequence and a residual sequence rn (t), a characteristic analysis module judges chaotic characteristics by calculating the maximum Lyapunov index of each IMF, and a neural network modeling module selects a modeling path according to chaotic / non-chaotic marks to generate a predicted value. The result integration module fuses the multi-path prediction results, compares the multi-path prediction results with a preset threshold interval, and outputs normal, early warning or fault state marks; and the result display module dynamically visualizes the state curve and the maintenance suggestion. According to the system, real-time judgment and early warning of abnormal operation of the cold and heat source equipment are realized, and decision support is provided for preventive maintenance.

Description

Technical Field

[0001] This invention relates to the field of intelligent equipment fault diagnosis technology. More specifically, this invention relates to an intelligent fault diagnosis system for a cold / heat source center and a method for assessing its condition. Background Technology

[0002] Heating and cooling source equipment plays a crucial role in energy supply and building environment control systems, and its operational stability directly affects system energy efficiency and safety. Current fault diagnosis methods mainly face the following limitations: First, judgment methods based on physical models are difficult to accurately characterize the complex nonlinear dynamic characteristics of equipment operation. Due to the strong coupling and random disturbances of parameters such as pressure, flow rate and temperature of cold and heat source equipment, the sensor signals exhibit highly non-stationary and multi-scale characteristics, and traditional linear models are not adaptable to such complex correlations. Secondly, the regular maintenance method that relies on manual experience analysis has a slow response speed and cannot capture the gradual process of equipment status changes in real time. On-site operators need to make fault judgments by comparing historical data and matching rule bases. The identification of early performance degradation signs is lagging and is prone to misjudgment or missed detection due to subjective experience differences. While existing data-driven methods can partially overcome the limitations of physical models, significant bottlenecks remain when processing equipment signals: time-series data acquired by sensors often contains both chaotic and non-chaotic characteristics. Chaotic signals are sensitive to initial conditions and unpredictable in the long term, while non-chaotic signals exhibit deterministic patterns, with the two overlapping and intermingling in the frequency and time domains. Traditional single signal processing techniques (such as Fourier transform or wavelet analysis) struggle to effectively separate these two characteristics, leading to insufficient feature extraction. For example, while empirical mode decomposition can decompose non-stationary signals, it fails to distinguish the dynamic characteristics of different intrinsic mode functions (IMFs); and conventional neural network models have limited ability to model the long-term evolution of chaotic sequences. This limitation stems from the complexity of multi-parameter interactions during equipment operation—pressure fluctuations may induce chaotic flow oscillations, while current harmonics exhibit a nonlinear correlation with temperature drift, making it difficult for a single analysis method to simultaneously meet the modeling requirements of chaotic and non-chaotic signals.

[0003] Furthermore, existing technologies struggle to balance real-time performance and accuracy: to improve judgment accuracy, some methods require high-dimensional feature extraction or complex model calculations, but in scenarios where hot and cold source centers require minute-level response times, the computational load significantly delays fault warnings. Simultaneously, performance degradation assessment relies on long-term data accumulation, while dynamic changes in equipment operating conditions (such as seasonal load switching and start-up / shutdown transients) cause the mapping relationship between historical data and real-time status to drift, leading to frequent false alarms in assessment methods based on fixed thresholds. These problems are particularly pronounced in systems lacking adaptive modeling mechanisms. For example, when sensor signals transition from a non-chaotic state to a chaotic state due to equipment aging, traditional models cannot autonomously switch modeling strategies, resulting in accumulated prediction biases.

[0004] Ultimately, the technical bottlenecks of existing methods can be summarized in three points: Insufficient feature separation: The coupling of chaotic and non-chaotic components in non-stationary signals leads to the masking of effective features; Lack of modeling adaptability: A single model is difficult to reconcile the long-range correlation of chaotic sequences with the deterministic patterns of non-chaotic sequences; Real-time constraints: There is a contradiction between the computational complexity of high-precision modeling and the timeliness requirements of fault diagnosis. Summary of the Invention

[0005] One object of the present invention is to provide an intelligent fault diagnosis system for a cold and heat source center, comprising: The data acquisition subsystem uses a sensor array including temperature, pressure, flow, current, and voltage sensors to collect real-time operating data from the central cold and heat source equipment. The collected data includes real-time temperature parameters of the equipment. T i Pressure parameters P i Flow parameters F i Current parameters I i Voltage parameters V i Composition of running feature vectors X =[ T i , P i , F i , I i , V i ]; The status signal database is connected to the data acquisition subsystem and is used to store the operational data collected by the data acquisition subsystem. The host computer subsystem includes the following components connected in sequence: The signal processing module connects to the status signal database and extracts the operational feature vector of the central cold / heat source equipment at time step t. X(t) The empirical mode decomposition method is used to process the data, and n IMF sequences are output. and residual sequences r n (t) ; The feature analysis module analyzes IMF sequences. Analyze the signal characteristics and output the chaotic / non-chaotic determination result; The neural network modeling module selects a modeling path based on the chaos determination results output by the characteristic analysis module and generates corresponding predicted values. ; The results integration module integrates the prediction results of different modeling paths generated by the neural network modeling module and outputs the fault judgment conclusion. The results display module receives the fault judgment conclusions output by the results integration module and visualizes these conclusions.

[0006] Preferably, the empirical mode decomposition steps of the signal processing module include: 1) For the input signal X(t) Extract extreme points and generate the upper envelope using cubic spline interpolation. and lower envelope ; 2) Calculate the envelope mean line ; 3) Extract the difference component ; 4) If If both conditions of the IMF are met, then it is determined to be... The dual conditions are that the number of extreme points is equal to or differs by at most one from the number of zero-crossing points, and the mean of the upper and lower envelopes at any point is approximately zero. 5) If the dual conditions are not met, then As new input, iteratively execute steps 1) to 3) until the... k Sub-iteration component Satisfying the standard deviation criterion ,in, The length of the original signal is used for output. = ; 6) Calculate the residual signal ; 7) with Repeat steps 1) to 6) as new inputs until the termination condition is met: the residual signal amplitude is lower than a preset threshold or becomes a monotonic function, and the final output is obtained. nIMF sequences { ,…, } and residual sequences Simultaneously satisfying: .

[0007] Preferably, the signal characteristic analysis module analyzes the signal characteristics by including the following steps: For each IMF sequence output by the signal processing module For each IMF sequence output by the signal processing module ,implement: Delay time τ i Calculation: via autocorrelation function Confirmed, among which , is the sequence mean. T Let be the sequence length, and take . The τ corresponding to the first drop to zero is taken as ; Embedding dimension calculate: a. in m Calculate the phase point in 3D phase space. Nearest Neighbor The distance is ; b. When the dimension increases to m When +1, the distance is updated to ; c. Definition about The mean is Define the rate of change of distance, where N is the number of effective phase points; d. Judgment criteria: Gradually increase the number of dimensions. m Calculate the corresponding ,when Follow m When the dimension increases and tends to stabilize, that is, when the dimension is continuously increased... The rate of change is less than the preset threshold α The minimum dimension corresponding to time. m That is, the embedding dimension of the sequence is determined.

[0008] Preferably, the signal characteristics analyzed by the characteristic analysis module further include: S1, Receive And calculate the delay time for each sequence. and embedding dimension Based on τ i and m i Reconstructing phase space trajectory Z i (t) ,in: In the reconstruction Z i (t) The maximum Lyapunov exponent is calculated using a small data set method. λ max : S2, in Z i (t) Select two initial points in the phase space. and ; S3, Calculate the initial distance ; S4. Iterate along the phase space trajectory N times to obtain... ; S5. Calculated using the distance logarithmic growth rate formula λ max : ,in, In time step distance, It is the time step; S6, according to λ max Determining chaotic properties: If λ max >0, mark If the sequence is chaotic, it is marked as chaotic; otherwise, it is marked as non-chaotic. Output the IMF sequence set with chaotic labeling. }

[0009] Preferably, the neural network modeling module is configured as follows: The input preprocessing unit, connected to the output of the feature analysis module, is used to receive the IMF sequence set {c} with chaotic labels. i (t)}, and divide the sequence into two types of data channels according to the chaotic label: First channel reception λ max Non-chaotic sequences ≤0 and residual sequences r n (t) , constitute the input vector , ,..., ,r(t)] ,in, k is the number of non-chaotic IMF sequences, and is the time step; Second channel reception λ max A chaotic sequence with a value greater than 0 is transformed into a phase space reconstruction unit. ; The fuzzy logic layer uses Gaussian membership functions. Process the input vector, where, For the input corresponding to the first k The membership degree of a fuzzy set. and These are the center and width parameters of the membership function, respectively, obtained from historical data training; Wavelet transform layer affects activation intensity Perform discrete wavelet transform and output ,in, For the first k Rule on scale a and displacement b The wavelet coefficients below, For the complex conjugate of wavelet basis functions, a For scale parameters, b These are translation parameters; neural network layer input It outputs through the hidden layer. ,in, For the hidden layer m The output of each neuron For the input layer k The node is connected to the hidden layer. m The weight of each node, For the hidden layer m Bias of each neuron As the activation function, then through Output, where, For the model at time step t The output, To get from the hidden layer m The weights from each node to the output layer For output layer bias, H This represents the number of neurons in the hidden layer. When the output vector is At that time, the output of the neural network layer is ; When the output vector is , the output of the neural network layer is . .

[0010] Preferably, the results integration module is configured as follows: Receive prediction results and prediction results ; The prediction results from the same modeling path are fused using a weighted formula: Among them, the weighting coefficient and The prediction accuracy of each model is set based on historical data; Will Compared with the preset fault threshold range [ θ 1, θ 2] Comparison: when ∈[0, θ 1 Output a normal status flag; when ∈( θ 1, θ 2] Output warning status flag; when > θ 2. Output fault status flags; The results display module dynamically displays the device status change curve and real-time markers.

[0011] Preferably, the results display module is configured as follows: The device status flags output by the result integration module; When a fault condition is identified, a list of components to be repaired is generated based on the chaotic markers of the IMF sequence output by the characteristic analysis module. The list maps physical components according to the sensor type of the chaotic sequence source. The chaotic sequence related to the temperature sensor corresponds to the maintenance items for the condenser or evaporator. The chaotic sequence related to the pressure sensor corresponds to the compressor or valve maintenance items; The chaotic sequence related to the flow sensor corresponds to the maintenance items for the water pump or pipeline. When marked as alarm status, a system checklist is generated that includes sensor group calibration items and associated device test items; When marked as normal, a performance evaluation report is output, which includes the output of the neural network modeling module. and / or ; The device's real-time operating curves, historical status change trends, and current maintenance recommendations are displayed synchronously through a visual interface.

[0012] A condition assessment method for an intelligent fault diagnosis system for a cold / heat source center is provided, comprising the following steps: Data acquisition steps: Real-time acquisition of operating data of the central equipment of the cold and heat source through sensor group. The sensor group includes temperature sensor, pressure sensor, flow sensor, current sensor and voltage sensor. The data types acquired include temperature parameter T, pressure parameter P, flow parameter F, current parameter I and voltage parameter V, which form an operating feature vector X=[T,P,F,I,V], where T, P, F, I and V correspond to the real-time measurement values ​​of the equipment. Data storage steps: The collected operational data is stored in the status signal database, which provides a data source for subsequent processing; Signal processing steps: Extract running feature vectors from the state signal database X(t) t is the time step, and the empirical mode decomposition method is used for processing. X(t) Output n IMF sequences and residual sequences r n (t) Empirical mode decomposition includes the following sub-steps: 1) For the input signal X(t) Extract extreme points and generate the upper envelope using cubic spline interpolation. and lower envelope ; 2) Calculate the envelope mean line ; 3) Extract the difference component ; 4) If If both conditions of the IMF are met, then it is determined to be... The dual conditions are that the number of extreme points is equal to or differs by at most one from the number of zero-crossing points, and the mean of the upper and lower envelopes at any point is approximately zero. 5) If the dual conditions are not met, then As new input, iteratively execute steps 1) to 3) until the... k Sub-iteration component Satisfying the standard deviation criterion ,in, Given the sequence length, the output will be... = ; 6) Calculate the residual signal ; 7) with Repeat steps 1) to 6) as new inputs until the termination condition is met: the residual signal amplitude is lower than a preset threshold or becomes a monotonic function, and the final output is obtained. n IMF sequences { ,…, } and residual sequences Simultaneously satisfying: ; Characteristic analysis step: For each IMF sequence output from the signal processing step Analyzing signal characteristics includes the following sub-steps: Calculate the delay time τ i Through autocorrelation function Confirmed, among which , is the sequence mean. T Let be the sequence length, and take . The τ corresponding to the first drop to zero is taken as ; Calculate the embedding dimension ,include: a. in m Calculate the phase point in 3D phase space. Nearest Neighbor The distance is ; b. When the dimension increases to m When +1, the distance is updated to ; c. Definition about The mean is , Define the rate of change of distance Where N is the number of effective phase points; d. Judgment criteria: Gradually increase the number of dimensions. m Calculate the corresponding ,when Follow m When the dimension increases and tends to stabilize, that is, when the dimension is continuously increased... The rate of change is less than the preset threshold α The minimum dimension corresponding to time. m That is, to determine the embedding dimension of the sequence. ; Based on τ i and m i Reconstructing phase space trajectory ; In reconstruction Z i (t) The maximum Lyapunov exponent is calculated using a small data set method. λ max : exist Z i (t) Select two initial points in the phase space. and ; Calculate the initial distance ; Iterate N times along the phase space trajectory to obtain ; Calculated using the distance logarithmic growth rate formula λ max : ,in, In time step distance, It is the time step; according to λ max Determining chaotic properties: If λ max >0, mark If the sequence is chaotic, it is marked as chaotic; otherwise, it is marked as non-chaotic. Output the IMF sequence set with chaotic labeling. }; Neural network modeling steps: Based on the chaos markers output from the feature analysis step, a modeling path is selected, including the following sub-steps: Input preprocessing categorizes IMF sequences into two types: non-chaotic and chaotic sequences. Non-chaotic sequences include... λ max,i ≤0 and residual sequences r n (t) , constitute the input vector , ,..., ,r(t)] , k The number of non-chaotic IMF sequences; chaotic sequences include... λ max >0 Transformed through phase space reconstruction ; The fuzzy logic layer processes the input vector. Process the input vector, where, For the input corresponding to the first k The membership degree of a fuzzy set. and These are the center and width parameters of the membership function, respectively, obtained from historical data training; Wavelet transform layer affects activation intensity Perform discrete wavelet transform and output ,in, For the first k Rule on scale a and displacement b The wavelet coefficients below, For the complex conjugate of wavelet basis functions, a For scale parameters, b These are translation parameters; neural network layer input It outputs through the hidden layer. ,in, For the hidden layer m The output of each neuron For the input layer k The node is connected to the hidden layer. m The weight of each node, For the hidden layer m Bias of each neuron As the activation function, then through Output, where, For the model at time step t The output, To get from the hidden layer m The weights from each node to the output layer For output layer bias, H This represents the number of neurons in the hidden layer. When the output vector is At that time, the output of the neural network layer is ; When the output vector is At that time, the output of the neural network layer is ; Result integration step: Receive the output of the neural network modeling step. and Through weighted formula Fusion prediction results ， Weighting coefficient and The prediction accuracy of each model is set based on historical data; Will Compared with the preset fault threshold range [ θ 1, θ 2] Comparison: when ∈[0, θ 1 Output a normal status flag; when ∈( θ 1, θ 2] Output warning status flag; when > θ 2. Output fault status flags; The results display module steps are as follows: Receive the device status flags output by the results integration module; When a fault condition is identified, a list of components to be repaired is generated based on the chaotic markers of the IMF sequence output by the characteristic analysis module. The list maps physical components according to the sensor type of the chaotic sequence source. The chaotic sequence related to the temperature sensor corresponds to the maintenance items for the condenser or evaporator. The chaotic sequence related to the pressure sensor corresponds to the compressor or valve maintenance items; The chaotic sequence related to the flow sensor corresponds to the maintenance items for the water pump or pipeline. When marked as alarm status, a system checklist is generated that includes sensor group calibration items and associated device test items; When marked as normal, a performance evaluation report is output, which includes the output of the neural network modeling module. and / or ; The device's real-time operating curves, historical status change trends, and current maintenance recommendations are displayed synchronously through a visual interface.

[0013] The present invention has at least the following beneficial effects: This invention achieves full automation from data acquisition to fault diagnosis output through modular collaborative processing. The signal processing module utilizes EMD to decompose non-stationary signals, the feature analysis module accurately distinguishes between chaotic and non-chaotic sequences, and the neural network modeling module adaptively selects between chaotic and non-chaotic modeling paths, thus solving the problem of insufficient feature extraction from mixed signals. The system response time is reduced to the minute level, the fault identification accuracy is improved, and parallel processing of multi-sensor data is supported.

[0014] This invention standardizes the EMD termination condition (standard deviation criterion + residual signal amplitude threshold), avoiding mode aliasing or over-decomposition caused by insufficient iterations, enhancing the clarity of the physical meaning of IMF components, and improving the residual sequence. r n (t) It effectively characterizes the overall performance trend of the equipment, providing stable input for subsequent analysis.

[0015] This invention achieves rapid convergence of the delay time τ by combining the autocorrelation function with the improved spurious nearest neighbor method. i and embedding dimension Calculating the Lyapunov exponent using a small amount of data λ max,i The error rate is reduced to less than 5%. Compared with traditional methods, the phase space reconstruction time is reduced by 30%, and the robustness to noise interference is enhanced, ensuring the reliability of chaos determination.

[0016] This invention adapts to signals with multiple characteristics by employing a dual-channel modeling strategy (FWNN for non-chaotic sequences and Chaos-FWNN for chaotic sequences). The fuzzy logic layer handles uncertainty through Gaussian membership functions, the wavelet transform layer achieves time-frequency localization feature extraction, and the neural network layer integrates the advantages of both, reducing prediction errors by 15%~22% (compared to a single model), and significantly improving early warning capabilities for chaotic faults such as compressor surge.

[0017] This invention dynamically sets the weighting coefficients based on historical prediction accuracy. , , to make the fusion result More realistic. Preset fault threshold range [ θ 1, θ 2] Supports online calibration, adapting to seasonal load fluctuations and operating condition drift caused by equipment aging, reducing the false alarm rate by approximately 10%. Status markers (normal / warning / fault) are visualized synchronously with curves, assisting maintenance personnel in responding quickly.

[0018] This invention shortens fault location time by 40% through intelligent mapping of chaotic labels to physical components (e.g., chaotic temperature sequence → evaporator maintenance). Early warning status triggers a sensor calibration checklist, while normal status outputs a performance evaluation report (including predicted values), forming a closed loop of "judgment-maintenance-evaluation," extending the average fault-free operating time of the equipment by 17%.

[0019] This invention integrates data acquisition, signal decomposition, chaotic analysis, dual-path modeling, and result output through a unified methodology framework, eliminating delays and errors caused by multi-platform operations. The entire process takes less than 5 minutes, meeting the real-time monitoring needs of hot and cold source centers and reducing operation and maintenance costs by approximately 25%.

[0020] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0021] Figure 1 This is a hardware structure topology diagram of the system according to one of the technical solutions of the present invention; Figure 2 This is a flowchart illustrating the dual-path modeling process of the neural network modeling module according to one of the technical solutions of the present invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0023] It should be noted that, unless otherwise specified, the experimental methods described in the following embodiments are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified. In the description of this invention, the orientation or positional relationship indicated by the terms is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing this invention and simplifying the description. It does not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0024] like Figures 1 - 2 As shown, the present invention provides an intelligent fault diagnosis system for cold and heat source centers, comprising: The data acquisition subsystem uses a sensor array including temperature, pressure, flow, current, and voltage sensors to collect real-time operating data from the central cold and heat source equipment. The collected data includes real-time temperature parameters of the equipment. T i Pressure parameters P i Flow parameters F i Current parameters I i Voltage parameters V i Composition of running feature vectors X =[ T i , P i , F i , I i , V i ]; The status signal database is connected to the data acquisition subsystem and is used to store the operational data collected by the data acquisition subsystem. The host computer subsystem includes the following components connected in sequence: The signal processing module connects to the status signal database and extracts the operational feature vector of the central cold / heat source equipment at time step t. X(t) The empirical mode decomposition method is used to process the data, and n IMF sequences are output. and residual sequences r n (t) ; The feature analysis module analyzes IMF sequences. Analyze the signal characteristics and output the chaotic / non-chaotic determination result; The neural network modeling module selects a modeling path based on the chaos determination results output by the characteristic analysis module and generates corresponding predicted values. ; The results integration module integrates the prediction results of different modeling paths generated by the neural network modeling module and outputs the fault judgment conclusion. The results display module receives the fault judgment conclusions output by the results integration module and visualizes these conclusions.

[0025] Specifically, the data acquisition subsystem collects equipment operating data in real time through a sensor array. Temperature sensors, with a range of -50℃ to 150℃, are installed on the inlet and outlet pipes of the condenser; pressure sensors, with a range of 0 to 10MPa, are installed near the compressor or valves; flow sensors, with a range of 0 to 1000L / min, are installed on the straight section of the water pump pipeline; current sensors, with a range of 0 to 100A, are installed on the motor power line; and voltage sensors, with a range of 0 to 500V, are installed inside the distribution cabinet. The sensor housings are made of stainless steel, and signals are transmitted via copper-core insulated cables. The status signal database uses a time-series database structure and is deployed on a local server. The sensors collect parameters at a frequency of 1Hz to form an operating feature vector. X The data is processed by the analog-to-digital conversion module, packaged and stored every 5 minutes, and functional tests simulate different working conditions to verify data integrity. The experiment shows that the data loss rate is <0.1%, and the technical effect is to provide a foundation for high-precision real-time data. Signal processing module extraction X(t) Empirical mode decomposition and standard deviation criterion threshold are used. ε The value is set to 0.25, the sequence length is 600 points, the module runs on an industrial control computer, cubic spline interpolation generates the envelope, iterative decomposition is performed until the termination condition is met, the characteristic analysis module calculates the delay time and embedding dimension, reconstructs the phase space trajectory, the delay time judgment threshold is 0.05, the embedding dimension judgment threshold α=0.01, the chaotic characteristic judgment threshold is 0, and sequences greater than 0 are marked as chaotic. The module decomposes the signal into intrinsic mode function sequence and residual sequence, analyzes the characteristics of each sequence, and verifies the function using a standard chaotic dataset. The experiment shows that the judgment error rate is <5%, and the technical effect is to accurately separate signal components and identify chaotic characteristics. The neural network modeling module selects a path based on chaotic labeling: non-chaotic sequences are input into the fuzzy logic layer, and chaotic sequences are reconstructed in phase space and input into the wavelet transform layer. The hidden layer has 50 neurons with ReLU activation. The result integration module weights and fuses the predicted values, and a preset fault threshold is applied. θ 1 =0.5, θ 2 =1.0, the results display module displays status curves and maintenance suggestions on the LCD screen in the monitoring room, fault markers generate component inspection lists; early warning markers generate system inspection lists; normal markers output performance reports, the module generates predicted values ​​and then fuses and compares them with thresholds to output visual status markers, the functional test simulates compressor surge events to verify the timeliness of early warnings, the technical effect is to realize the automation of fault judgment and provide visual decision support.

[0026] Comparative Example 1 Comparative Example 1 uses the closest existing technology: a traditional fault diagnosis system (such as a single-model system based on Fourier transform and wavelet analysis). This system includes sensor data acquisition, direct input to a wavelet transform module to extract features, matching thresholds through a preset rule base (such as an expert system) to output fault conclusions, data storage in a relational database, and results displayed on a basic display screen. During operation, signals are not distinguished by characteristics, a single model processes all sequences, thresholds are fixed and cannot be adjusted, and maintenance suggestions rely on human experience to generate.

[0027] Compared with traditional systems, this invention solves the problem of insufficient feature extraction of mixed signals by modular processing (EMD decomposition, chaos analysis, and dual-path modeling). Existing technologies do not distinguish between chaotic and non-chaotic signals, leading to feature confusion. The adaptive modeling of the embodiment (such as weighted fusion) improves prediction accuracy, while the single model of existing technologies has a large error on chaotic signals. The real-time threshold calibration and visual maintenance suggestions of the embodiment simplify operation and maintenance response, while existing technologies rely on fixed rules and manual judgment, resulting in a long response delay. Overall, the embodiment performs better in terms of data utilization efficiency and system robustness.

[0028] According to another implementation of the present invention, when the signal processing module performs empirical mode decomposition, it first extracts the input signal. X(t) For extreme points, cubic spline interpolation can be performed using industry-standard algorithm libraries (such as SciPy) to generate the upper envelope. and lower envelope The envelope mean line is calculated using... Formula, difference components The extraction is achieved through a floating-point arithmetic unit. In terms of numerical selection, the sampling frequency can be maintained at 1Hz, and the signal length T can take 10 minutes of data (600 points). The device uses an industrial control computer, and the processor can be configured with an Intel Xeon E series. The material is aluminum alloy heat dissipation shell. The assembly position is in the digital signal processing area of ​​the signal processing board. The module receives the sensor data stream in real time, detects the coordinates of local maxima / minimum points, and the interpolation algorithm updates the envelope curve at a period of 0.5 seconds. It calculates the envelope mean and outputs the difference component. The test uses a standard sine superimposed random noise signal to verify that the envelope fitting error is less than 0.5% and can accurately capture the signal oscillation characteristics. when When the inherent mode function condition is not met, the system initiates iterative processing, and the standard deviation... At that time, 64-bit floating-point precision was used, and the residual signal termination condition was set to amplitude < 0.01 or monotonicity detection passed. The device uses a multi-core processor for parallel computing, with 32GB DDR4 memory. Silicon-based semiconductor chips were selected as the material and assembled in the motherboard slot of the industrial control computer. The envelope was regenerated and new components were calculated in each iteration until SD < 0.25 or the iteration limit was reached. The functional test used a frequency-modulated signal to verify that convergence could be achieved in 3 to 5 iterations. The technical effect is to avoid mode aliasing caused by excessive decomposition. After the decomposition is terminated, the module calculates the residual signal. It then processes the residual signal cyclically, ultimately outputting a signal that satisfies... The number of IMFs, n, is automatically determined. The residual amplitude threshold is set to 0.01. The monotonicity detection uses the first-order difference method. The device transmits data through the PCIe bus. The storage medium is an SSD solid-state drive. The material is FR-4 circuit board substrate. It is assembled in the expansion slot of the industrial control computer. Each IMF component is output after noise reduction by a digital filter. The residual signal is stored separately. The experiment uses the transient data of compressor start-stop to verify that the reconstruction error is <1%. The technical effect is to completely separate the multi-scale components of the signal. Comparative Example 2 The closest existing technology adopts the traditional EMD method: using a fixed number of iterations (usually 20 times), without an adaptive termination criterion, the decomposition process runs on a general server, the residual processing only checks the amplitude threshold of 0.1, there is no monotonicity detection, the signal reconstruction error is about 5-8%, and the computation time is increased by 40%.

[0029] This implementation method improves upon existing technologies through three key techniques: 1) Setting SD=0.25 precisely controls iterative convergence, avoiding under-decomposition / over-decomposition problems in existing technologies; 2) Dual termination conditions (residual amplitude + monotonicity) ensure clear physical meaning, which is more reliable than the single threshold in existing technologies; 3) Hardware integration reduces computational latency and improves processing speed. Tests show that this solution reduces modal aliasing, providing purer components for subsequent chaos analysis.

[0030] According to another implementation of the present invention, the characteristic analysis module calculates the IMF sequence. Delay time τ i At that time, the autocorrelation function R(τ) Implemented using a standard mathematical library, the sequence mean is calculated using a cumulative averaging algorithm. The sequence length T can take 1024 sampling points. The threshold for the autocorrelation function to drop to zero is set to 0.05. A digital signal processor can be used for computation, installed in the coprocessor slot of the host computer subsystem. Silicon-based integrated circuit chips are selected as the material, and aluminum fins are used for heat dissipation. During operation, the module sequentially calculates the autocorrelation function for different τ values. R(τ) Record when it first becomes less than 0.05 τi The values ​​and parameters were set through historical data analysis. The experiment was conducted using the Lorenz system to verify the accuracy of the calculation. The technical effect is to accurately capture the self-similarity characteristics of the signal. Module calculates embedded dimension m i At that time, m Constructing phase points in 3D phase space Nearest neighbor search uses the kd-tree algorithm, and the initial dimension can be set to... m =3, distance update ratio s(t,m) The calculation uses the Euclidean distance formula, with the mean... E(m) The calculation window can be set to 100 data points, distance change rate Ê The convergence threshold α = 0.01; the device can be configured with GPU-accelerated computing; it is mounted in the PCIe expansion slot of an industrial control computer; the material selected is a copper-based circuit board; dimensions... m Starting from 3 and incrementing, calculate... Ê If the value changes by less than 0.01 twice consecutively, then convergence is determined, and the minimum value that first satisfies the condition is taken. m The experiment uses the Rossler attractor test to verify dimensional convergence, and the technical effect is to optimize the accuracy of phase space reconstruction. When the module performs the dimension increment judgment, m The maximum value can be set to 20. Ê The calculation uses a sliding window average, with a window step size of up to 10 data points. Convergence is determined by two conditions: three consecutive convergences. Ê With a change rate <0.01 and an absolute change <0.001, the device buffers intermediate data through DDR4 memory, is mounted on the motherboard memory slot, and uses an FR-4 substrate. For each additional dimension, the distance between all phase points is recalculated, and the change curve is updated in real time. 15% Gaussian noise is added to the functional test to verify the robustness of the algorithm. The technical effect is to ensure that the phase space is fully expanded.

[0031] Comparative Example 3 Traditional methods use mutual information to calculate the delay time and false nearest neighbor method to calculate the embedding dimension. The first minimum value of mutual information is easily affected by noise. The false nearest neighbor method requires manually setting a percentage threshold (usually 20%), which leads to deviation in dimension estimation. The calculation is performed on a general-purpose CPU and there is no convergence determination mechanism. It takes about 1200ms to process 1024 data points.

[0032] This scheme features three improvements: 1) A zero-crossing threshold of 0.05 for the autocorrelation function enhances noise resistance and provides greater stability compared to the traditional mutual information method; 2) The distance change rate... ÊAutomatic convergence determination (α=0.01) replaces manual proportional threshold, improving the accuracy of dimension estimation; 3) GPU acceleration reduces computation time to within 200ms, and tests show that phase space reconstruction error is reduced, laying the foundation for Lyapunov exponent calculation.

[0033] According to another implementation of the present invention, the characteristic analysis module calculates the delay time for each IMF sequence. and embedding dimension Delay time The first zero-crossing point of the autocorrelation function can be used to determine the value when the autocorrelation function drops to zero. τ Value can be used as Typical values ​​range from 10 milliseconds to 50 milliseconds, embedding dimension The improved spurious nearest neighbor method is used for calculation: m In 3D phase space, the distance between a phase point and its nearest neighbor is defined. As the dimension increases to... m+1 The distance values ​​are updated periodically, and the stability of the mean rate of change of distance is calculated. A preset stability threshold α can be set to 0.05. When the rate of change is below α as the dimension is continuously increased, the dimension is considered to have reached stability. m This refers to the embedding dimension, which typically ranges from 3 to 8 dimensions. This step is performed in an industrial control computer, using a dual-core or higher processor, equipped with 8GB of memory, and connected to the host computer subsystem bus via copper wires. The parameter calculation cycle is controlled to be completed within 200 milliseconds to meet real-time requirements. based on τ i and Reconstructing phase space trajectory Z i (t) Dimension equals The maximum Lyapunov exponent was calculated using a small data set method. λ max : Initial point selection: Z i (t) Two initial phase space points are randomly selected, and the distance between the two points is set to 5% to 15% of the signal standard deviation. Distance iteration: Iteratively track the change in distance between two points along the phase space trajectory, with a step size of [missing information]. Δt Take 20 milliseconds, number of iterations N Set to 1000 times; Exponential calculation: Outputted via the distance logarithmic growth rate formula. λ max The distance values ​​in the formula retain double-precision floating-point precision; Noise reduction: Gaussian white noise superposition test shows that when the signal-to-noise ratio is higher than 30dB, the calculation result deviation is less than 0.005. This process is executed in the numerical coprocessor of the characteristic analysis module, using the IEEE 754 floating-point operation standard. The data is buffered through DDR4 memory. The test uses compressor pressure fluctuation data. The results show that the chaotic sequence identification time is less than 150 milliseconds. according to λ max Classify the sequence based on the values: If λ max A value >0.01 indicates a chaotic sequence; otherwise, it is labeled as a non-chaotic sequence. The classification threshold of 0.01 is validated through Monte Carlo simulation, covering 12 operating conditions including equipment start-up and shutdown, and load surges. The output is a group of labeled IMF sequences. When the chaotic flag, sensor type and original sequence pointer are stored in a structured data format, the sequence group is transmitted to the neural network modeling module through the PCIe bus with a transmission delay of less than 5 milliseconds. Functional tests show that the accuracy of chaotic judgment of high-frequency oscillation signals of flow sensors reaches 98.3% (500 samples). Misjudgments mainly occur in the case of signal-to-noise ratio below 20dB. This implementation achieves rapid and accurate determination of the chaotic characteristics of IMF sequences. The adaptive calculation of phase space reconstruction parameters ensures the reliability of trajectory reconstruction. The Lyapunov exponent calculation process optimizes iteration efficiency. The chaotic classification threshold balances sensitivity and noise resistance. The system can distinguish in real time typical fault modes such as compressor surge (chaotic) and water pump mechanical wear (non-chaotic), providing effective input for subsequent dual-channel modeling.

[0034] According to another implementation of the present invention, the input preprocessing unit receives the IMF sequence group output by the feature analysis module, and the chaos determination threshold. λ max =0.01, determined through Monte Carlo simulation of 12 operating conditions including equipment start-up and shutdown, load surges, etc., non-chaotic sequence channel ( λ max ≤0) Input vector x ( t )Include k A non-chaotic IMF component ( k (usually 3-5) and residual sequences r n (t) Data is transmitted via the PCIe bus, chaotic sequence channel ( λ max >0) Transformed through phase space reconstruction unit Z i ( t ), Reconstruction parameter delay time τi Value range 10~50 milliseconds, embedding dimensionmi The value range is 3 to 8 dimensions. The device can be accelerated by NVIDIA Tesla GPU and is installed in the PCIe expansion slot of the industrial control computer. The material is copper-based circuit board. The functional test uses compressor surge data. The chaotic sequence recognition accuracy reaches 98.3% (500 samples). The fuzzy logic layer uses Gaussian membership functions. ,parameter (Center) and (Width) is generated through training on historical data, and the training data is stored in DDR4 memory. The wavelet transform layer affects the activation intensity. α k For discrete wavelet transform, the wavelet basis functions can be chosen from the Mexican hat wavelet (complex conjugate). ψ *), scale parameters a The value range is 0.1~10, and the translation parameter is... b For the corresponding time series positions, the device can use an Intel Xeon processor, mounted on an industrial control computer motherboard, and made of silicon-based chips. The robustness was verified by adding 15% Gaussian noise in the experiment, and the wavelet coefficient output error was less than 0.5%. Wavelet coefficients input to neural network layers βk Number of neurons in the hidden layer H Set to 50, activation function ϕ (⋅) ReLU is used, weights ω km and bias θ m The output layer is trained using the backpropagation algorithm and then computed. Non-chaotic channel output Chaotic channel output The equipment can be equipped with an industrial control computer with 32GB DDR4 memory, which is installed in a cold and heat source control cabinet. The heat dissipation material is aluminum alloy fins. The functional test simulates water pump mechanical wear (non-chaotic) and valve pressure oscillation (chaotic). The prediction results are transmitted to the result integration module via Modbus protocol.

[0035] This implementation method achieves accurate prediction and anomaly identification of the operating status of the central equipment of the cold and heat source through the dual-channel processing mechanism of the neural network modeling module. The input preprocessing unit is based on chaotic labels ( λ max ≤0 or λ max >0) Intelligent signal separation, integrating the non-chaotic sequence and the residual sequence into an input vector. x ( t Chaotic sequences are then transformed through phase space reconstruction. Z i ( tThis design fully preserves the differences in signal characteristics and solves the problem of insufficient feature extraction of mixed signals.

[0036] According to another implementation of the present invention, the prediction result receiving unit is connected to the neural network modeling module via an Ethernet interface to acquire the prediction result in real time. and The predicted value is in 32-bit floating-point format, with a transmission frequency of 1Hz and a buffer capacity of 300 sets of historical data. An industrial-grade communication gateway can be used, mounted in a DIN rail slot in the control cabinet, and made of flame-retardant PC plastic. During operation, the unit continuously monitors TCP / IP port 502 (Modbus protocol). When a new predicted value is detected, a data verification mechanism (such as CRC check) is triggered to ensure transmission integrity. Functional testing simulates sensor signal interruption scenarios, verifying a data loss rate of less than 0.1%. Weighted fusion unit execution y ( t ) Calculation, weighting coefficients and The weights are dynamically set through training on historical data: the weight range for non-chaotic models is 0.4~0.7, and the weight range for chaotic models is 0.3~0.6. The weight ω=1 / MAE, and the weights are assigned after normalization. For example, if the MAE of FWNN is 0.1 and the MAE of Chaos-FWNN is 0.15, then... =0.6, =0.4, the computing device can be an ARM Cortex-M7 microcontroller, mounted on an industrial control computer motherboard, with FR-4 circuit boards as the material. During operation, the unit updates the weights every 5 minutes: extract the prediction error (MAE) of each model in the past 24 hours, and allocate the weights inversely proportional to the accuracy (the weight increases by 0.05 for every 10% decrease in error). The functional test uses the compressor running dataset (1000 sets of samples) to verify that the adaptive convergence time of the weights is less than 2 minutes. The threshold comparison unit presets a fault threshold range. θ 1 = 0.5 θ 2=1.0, real-time comparison is achieved through a voltage comparator circuit, and the status marking rules are as follows: y ( t ()∈[0,0.5)→normal state; y ( t )∈(0.5,1.0]→Warning status; y ( t >1.0 → Fault status; The equipment can use comparator chips, mounted in the digital logic area of ​​the signal processing board, made of silicon-based semiconductors. The threshold calibration mechanism supports online adjustment: when seasonal load changes, the system automatically extracts the median of data from the most recent 30 days under similar operating conditions and dynamically updates it. θ 1 and θ 2. Functional testing simulated high-temperature summer conditions to verify that the threshold drift compensation error was less than 5%; The visualization output unit connects to the monitoring room display screen via an HDMI interface, dynamically rendering three curves: real-time running curve (blue), historical trend curve (gray), and status marker points (red / yellow / green). The display refresh rate is set to 25fps, and the resolution supports 1920×1080. The device can be an industrial HMI touch screen, mounted on the control panel, with a tempered glass cover. When the status marker indicates a fault, the system triggers the chaotic marker data from the characteristic analysis module to generate a list of maintenance components (such as the condenser maintenance item corresponding to the temperature chaotic sequence). The functional test triggers the compressor overpressure fault, verifying that the delay from marker output to screen display is less than 500ms.

[0037] This implementation achieves multi-dimensional assessment and real-time visualization of the status of cold and heat source equipment. The prediction result receiving unit ensures the complete transmission of dual model data, the weighted fusion unit optimizes the accuracy of status assessment through dynamic weight allocation, and the adaptive calibration mechanism of the threshold comparison unit effectively suppresses false alarms caused by operating condition drift. Finally, the visualization output unit transforms abstract data into intuitive curves and markers, and generates targeted maintenance suggestions based on chaotic characteristics. The entire module operates stably in the industrial control environment, supporting maintenance personnel to quickly identify equipment anomalies and take intervention measures.

[0038] According to another implementation of the present invention, after receiving the fault marker from the result integration module, the fault status response unit immediately calls the chaotic sequence marker data from the characteristic analysis module. The chaotic sequence related to the temperature sensor is mapped to the condenser or evaporator maintenance item, the pressure-related sequence is mapped to the compressor or valve maintenance item, and the flow-related sequence is mapped to the water pump or pipeline maintenance item. The list is generated using XML data format, and the number of maintenance items is usually 3 to 5. The equipment can be an industrial control computer, installed in the host slot of the control cabinet, and the material is an aluminum alloy shell. During operation, the unit uses SQL queries to associate the chaotic marker with the equipment topology database in real time. For example, when a chaotic sequence of the pressure sensor is detected, it automatically associates the compressor cylinder sealing detection and valve opening calibration items. The functional test simulates the compressor overpressure fault, and verifies that the list generation delay is less than 200 milliseconds. When the early warning response unit is triggered, it generates a checklist with two levels: sensor calibration items (such as zero-point drift correction of temperature sensors) and related equipment detection items (such as vibration detection of water pump bearings). The checklist items are sorted by priority, with high-priority items in the top three. The equipment can be an embedded microprocessor, mounted in the coprocessing area of ​​the signal processing board, and the material is FR-4 circuit board. The working process follows a dynamic update mechanism: historical false alarm records are extracted every 24 hours, and the weight of checklist items is adjusted. For example, when flow false alarms occur frequently, the first item on the checklist is set to clean the ultrasonic probe of the flow meter. The experiment uses vibration datasets from 12 water pumps to verify that the checklist covers 97% of early failure modes. The normal state response unit outputs a performance evaluation report in PDF format, which includes... and The time-series curves were compared, with a data sampling interval of 10 seconds. The visualization interface rendered three curves using the OpenGL library: the real-time running curve (red), the historical trend curve (blue), and the predicted value curve (green). The refresh rate was set to 25fps. The device could be an industrial HMI touchscreen, mounted on the control panel of the console, with a tempered glass surface. During operation, the interface synchronously displayed a summary of maintenance suggestions (such as "Current energy efficiency is in the optimal range"). When the user clicked on a curve node, the system linked to display the raw sensor data at that moment. The functional test ran for 72 hours under full load conditions, verifying that the interface was smooth without any lag.

[0039] This implementation achieves intelligent maintenance decision support for equipment status. The fault status response unit transforms abstract signals into targeted maintenance items through precise mapping of chaotic sequences to physical components, shortening fault location time. The dynamic list generation mechanism of the early warning status response unit optimizes inspection priorities based on historical data, improving the efficiency of early hazard identification. The visualized reports and curve comparisons of the normal status response unit provide intuitive evidence for performance evaluation. The entire mechanism operates stably in industrial control environments, supporting maintenance personnel to quickly execute preventative maintenance strategies. According to another implementation of the present invention, the sensor group includes a temperature sensor (range -50℃ to 150℃), a pressure sensor (range 0 to 10MPa), a flow sensor (range 0 to 1000L / min), and a current / voltage sensor (range 0 to 100A / 0 to 500V). The temperature sensor can be installed on the inlet and outlet pipes of the condenser, the pressure sensor is assembled at the compressor flange interface, and the flow sensor is placed in the straight section of the water pump pipeline. Data is collected at a frequency of 1Hz and forms a feature vector. XSubsequently, data is transmitted via RS485 bus. The device can be an industrial IoT gateway, installed on the terminal block of the control cabinet. The material is flame-retardant engineering plastic. During operation, the gateway performs data verification (CRC16 algorithm), and abnormal data is automatically rejected. Functional testing was conducted at 85℃ for 24 hours, verifying a data loss rate of less than 0.1%. Empirical mode decomposition is run on an industrial control computer, with standard deviation as the threshold. ε =0.25, the residual signal termination amplitude is set to 0.01, and the decomposed output IMF sequence is... c i ( t Phase space reconstruction analysis: Delay time τ i The calculation employs the zero-crossing method of the autocorrelation function (threshold 0.05), with an embedding dimension of [missing information]. m i Convergence determination based on the rate of change of distance (threshold) α =0.01), Lyapunov index λ max Calculate step size Δ t =20 ms The device can use a multi-core processor, is installed in the PCIe slot of an industrial control computer, and is made of copper-based heat sink. The working process is fully automated. Each IMF sequence analysis takes about 150ms. The experiment uses the Lorenz chaotic system to inject 15% noise to verify that the accuracy of chaotic identification is better than that of traditional methods. Non-chaotic sequence input vector x ( t It contains 3-5 IMF components and residual sequences; the chaotic sequence is reconstructed as follows: Zi ( t Then input the chaotic modeling channel, the fuzzy logic layer membership function center. c kl and width δ kl Wavelet transform scaling parameters are generated through training with historical data. a The range is set to 0.1~10, the number of neurons in the hidden layer of the neural network is fixed at 50, the activation function is ReLU, the device can be an edge computing box, assembled in the ventilated area of ​​the control cabinet, the material is aluminum alloy shell, and the state label threshold is set. θ 1 = 0.5 θ Version 2=1.0 supports seasonal adaptive calibration. The visualization interface is output to a 21.5-inch industrial monitor via HDMI, dynamically rendering three curves: real-time value (red), predicted value (blue), and threshold line (black). The functional test triggers a condenser blockage fault, verifying that the end-to-end latency from data acquisition to screen display is less than 3 seconds.

[0040] The method of this implementation realizes a closed-loop process for fault diagnosis of cold and heat source equipment. After adaptive decomposition of sensor data, the chaotic characteristic analysis module accurately distinguishes the dynamic characteristics of the signal (such as pressure oscillation chaos vs. current drift non-chaotic). The dual-path modeling strategy processes the two types of signals simultaneously. The time-frequency characteristics of the signal are preserved through fuzzy wavelet neural network. The dynamic threshold mechanism of the result integration module suppresses the interference of operating condition fluctuations. The visualization output transforms complex data into actionable decision-making basis for operation and maintenance personnel (such as temperature chaotic sequence directly related to evaporator maintenance items). The entire system operates stably and continuously in the industrial environment, supporting the implementation of preventive maintenance strategies.

[0041] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. An intelligent fault diagnosis system for cold and heat source centers, characterized in that, include: The data acquisition subsystem uses a sensor array including temperature, pressure, flow, current, and voltage sensors to collect real-time operating data from the central cold and heat source equipment. The collected data includes real-time temperature parameters of the equipment. T i Pressure parameters P i Flow parameters F i Current parameters I i Voltage parameters V i Composition of running feature vectors X =[ T i , P i , F i , I i , V i ]; The status signal database is connected to the data acquisition subsystem and is used to store the operational data collected by the data acquisition subsystem. The host computer subsystem includes the following components connected in sequence: The signal processing module connects to the status signal database and extracts the operational feature vector of the central cold / heat source equipment at time step t. X(t) The empirical mode decomposition method is used to process the data, and n IMF sequences are output. and residual sequences r n (t) ; The feature analysis module analyzes IMF sequences. Analyze the signal characteristics and output the chaotic / non-chaotic determination result; The neural network modeling module selects a modeling path based on the chaos determination results output by the characteristic analysis module and generates corresponding predicted values. ; The results integration module integrates the prediction results of different modeling paths generated by the neural network modeling module and outputs the fault judgment conclusion. The results display module receives the fault judgment conclusions output by the results integration module and visualizes these conclusions.

2. The system as described in claim 1, characterized in that, The specific steps of empirical mode decomposition in the signal processing module include: 1) For the input signal X(t) Extract extreme points and generate the upper envelope using cubic spline interpolation. and lower envelope ; 2) Calculate the envelope mean line ; 3) Extract the difference component ; 4) If If both conditions of the IMF are met, then it is determined to be... The dual conditions are that the number of extreme points is equal to or differs by at most one from the number of zero-crossing points, and the mean of the upper and lower envelopes at any point is approximately zero. 5) If the dual conditions are not met, then As new input, iteratively execute steps 1) to 3) until the... k Sub-iteration component Satisfying the standard deviation criterion ,in, The length of the original signal is used for output. = ; 6) Calculate the residual signal ; 7) with Repeat steps 1) to 6) as new inputs until the termination condition is met: the residual signal amplitude is lower than a preset threshold or becomes a monotonic function, and the final output is obtained. n IMF sequences { ,…, } and residual sequences Simultaneously satisfying: .

3. The system as described in claim 2, characterized in that, The characteristic analysis module analyzes signal characteristics through the following steps: For each IMF sequence output by the signal processing module For each IMF sequence output by the signal processing module ,implement: Delay time τ i Calculation: via autocorrelation function Confirmed, among which , is the sequence mean. T Let be the sequence length, and take . The τ corresponding to the first drop to zero is taken as ; Embedding dimension calculate: exist m Calculate the phase point in 3D phase space. Nearest Neighbor The distance is ; b. When the dimension increases to m When +1, the distance is updated to ; c. Definition about The mean is Define the rate of change of distance Where N is the number of effective phase points; d. Judgment criteria: Gradually increase the number of dimensions. m Calculate the corresponding ,when Follow m When the dimension increases and tends to stabilize, that is, when the dimension is continuously increased... The rate of change is less than the preset threshold α The minimum dimension corresponding to time. m That is, to determine the embedding dimension of the sequence. .

4. The system as described in claim 3, characterized in that, The signal characteristics analyzed by the characteristic analysis module further include: S1, Receive And calculate the delay time for each sequence. and embedding dimension Based on τ i and m i Reconstructing phase space trajectory Z i (t) ,in: In the reconstruction Z i (t) The maximum Lyapunov exponent is calculated using a small data set method. λ max : S2, in Z i (t) Select two initial points in the phase space. and ; S3, Calculate the initial distance ; S4. Iterate along the phase space trajectory N times to obtain... ; S5. Calculated using the distance logarithmic growth rate formula λ max : ,in, In time step distance, It is the time step; S6, according to λ max Determining chaotic properties: If λ max >0, mark If the sequence is chaotic, it is marked as chaotic; otherwise, it is marked as non-chaotic. Output the IMF sequence set with chaotic labeling. } 5. The system as described in claim 4, characterized in that, The neural network modeling module is configured as follows: The input preprocessing unit, connected to the output of the feature analysis module, is used to receive the IMF sequence set {c} with chaotic labels. i (t)}, and divide the sequence into two types of data channels according to the chaotic label: First channel reception λ max Non-chaotic sequences ≤0 and residual sequences r n (t) , constitute the input vector , ,..., ,r(t)] ,in, k is the number of non-chaotic IMF sequences, and is the time step; Second channel reception λ max A chaotic sequence with a value greater than 0 is transformed into a phase space reconstruction unit. ; The fuzzy logic layer uses Gaussian membership functions. Process the input vector, where, For the input corresponding to the first k The membership degree of a fuzzy set. and These are the center and width parameters of the membership function, respectively, obtained from historical data training; Wavelet transform layer affects activation intensity Perform discrete wavelet transform and output ,in, For the first k Rule on scale a and displacement b The wavelet coefficients below, For the complex conjugate of wavelet basis functions, a For scale parameters, b These are translation parameters; neural network layer input It outputs through the hidden layer. ,in, For the hidden layer m The output of each neuron For the input layer k The node is connected to the hidden layer. m The weight of each node, For the hidden layer m Bias of each neuron As the activation function, then through Output, where, For the model at time step t The output, To get from the hidden layer m The weights from each node to the output layer For output layer bias, H This represents the number of neurons in the hidden layer. When the output vector is At that time, the output of the neural network layer is ; When the output vector is At that time, the output of the neural network layer is .

6. The system as described in claim 5, characterized in that, The result integration module is configured as follows: Receive prediction results and prediction results ; The prediction results from the same modeling path are fused using a weighted formula: Among them, the weighting coefficient and The prediction accuracy of each model is set based on historical data; Will Compared with the preset fault threshold range [ θ 1, θ 2] Comparison: when ∈[0, θ 1 Output a normal status flag; when ∈( θ 1, θ 2] Output warning status flag; when > θ 2. Output fault status flags; The results display module dynamically displays the device status change curve and real-time markers.

7. The system as described in claim 6, characterized in that, The results display module is configured as follows: The device status flags output by the result integration module; When a fault condition is identified, a list of components to be repaired is generated based on the chaotic markers of the IMF sequence output by the characteristic analysis module. The list maps physical components according to the sensor type of the chaotic sequence source. The chaotic sequence related to the temperature sensor corresponds to the maintenance items for the condenser or evaporator. The chaotic sequence related to the pressure sensor corresponds to the compressor or valve maintenance items; The chaotic sequence related to the flow sensor corresponds to the maintenance items for the water pump or pipeline. When marked as alarm status, a system checklist is generated that includes sensor group calibration items and associated device test items; When marked as normal, a performance evaluation report is output, which includes the output of the neural network modeling module. and / or ; The device's real-time operating curves, historical status change trends, and current maintenance recommendations are displayed synchronously through a visual interface.

8. A condition assessment method for an intelligent fault diagnosis system for a cold / heat source center, based on the system described in claim 7, characterized in that, Includes the following steps: Data acquisition steps: Real-time acquisition of operating data of the central equipment of the cold and heat source through sensor group. The sensor group includes temperature sensor, pressure sensor, flow sensor, current sensor and voltage sensor. The data types acquired include temperature parameter T, pressure parameter P, flow parameter F, current parameter I and voltage parameter V, which form an operating feature vector X=[T,P,F,I,V], where T, P, F, I and V correspond to the real-time measurement values ​​of the equipment. Data storage steps: The collected operational data is stored in the status signal database, which provides a data source for subsequent processing; Signal processing steps: Extract running feature vectors from the state signal database X(t) t is the time step, and the empirical mode decomposition method is used for processing. X(t) Output n IMF sequences and residual sequences r n (t) Empirical mode decomposition includes the following sub-steps: 1) For the input signal X(t) Extract extreme points and generate the upper envelope using cubic spline interpolation. and lower envelope ; 2) Calculate the envelope mean line ; 3) Extract the difference component ; 4) If If both conditions of the IMF are met, then it is determined to be... The dual conditions are that the number of extreme points is equal to or differs by at most one from the number of zero-crossing points, and the mean of the upper and lower envelopes at any point is approximately zero. 5) If the dual conditions are not met, then As new input, iteratively execute steps 1) to 3) until the... k Sub-iteration component Satisfying the standard deviation criterion ,in, Given the sequence length, the output will be... = ; 6) Calculate the residual signal ; 7) with Repeat steps 1) to 6) as new inputs until the termination condition is met: the residual signal amplitude is lower than a preset threshold or becomes a monotonic function, and the final output is obtained. n IMF sequences { ,…, } and residual sequences Simultaneously satisfying: ; Characteristic analysis step: For each IMF sequence output from the signal processing step Analyzing signal characteristics includes the following sub-steps: Calculate the delay time τ i Through autocorrelation function Confirmed, among which , is the sequence mean. T Let be the sequence length, and take . The τ corresponding to the first drop to zero is taken as ; Calculate the embedding dimension ,include: a. in m Calculate the phase point in 3D phase space. Nearest Neighbor The distance is ; b. When the dimension increases to m When +1, the distance is updated to ; c. Definition about The mean is Define the rate of change of distance Where N is the number of effective phase points; d. Judgment criteria: Gradually increase the number of dimensions. m Calculate the corresponding ,when Follow m When the dimension increases and tends to stabilize, that is, when the dimension is continuously increased... The rate of change is less than the preset threshold α The minimum dimension corresponding to time. m That is, to determine the embedding dimension of the sequence. ; Based on τ i and m i Reconstructing phase space trajectory ； In reconstruction Z i (t) The maximum Lyapunov exponent is calculated using a small data set method. λ max : exist Z i (t) Select two initial points in the phase space. and ; Calculate the initial distance ; Iterate N times along the phase space trajectory to obtain ; Calculated using the distance logarithmic growth rate formula λ max : ,in, In time step distance, It is the time step; according to λ max Determining chaotic properties: If λ max >0, mark If the sequence is chaotic, it is marked as chaotic; otherwise, it is marked as non-chaotic. Output the IMF sequence set with chaotic labeling. }; Neural network modeling steps: Based on the chaos markers output from the feature analysis step, a modeling path is selected, including the following sub-steps: Input preprocessing categorizes IMF sequences into two types: non-chaotic and chaotic sequences. Non-chaotic sequences include... λ max,i ≤0 and residual sequences r n (t) , constitute the input vector , ,..., ,r(t)] , k The number of non-chaotic IMF sequences; chaotic sequences include... λ max >0 Transformed through phase space reconstruction ; The fuzzy logic layer processes the input vector. Process the input vector, where, For the input corresponding to the first k The membership degree of a fuzzy set. and These are the center and width parameters of the membership function, respectively, obtained from historical data training; Wavelet transform layer affects activation intensity Perform discrete wavelet transform and output ,in, For the first k Rule on scale a and displacement b The wavelet coefficients below, For the complex conjugate of wavelet basis functions, a For scale parameters, b These are translation parameters; neural network layer input It outputs through the hidden layer. ,in, For the hidden layer m The output of each neuron For the input layer k The node is connected to the hidden layer. m The weight of each node, For the hidden layer m Bias of each neuron As the activation function, then through Output, where, For the model at time step t The output, To get from the hidden layer m The weights from each node to the output layer For output layer bias, H This represents the number of neurons in the hidden layer. When the output vector is At that time, the output of the neural network layer is ; When the output vector is At that time, the output of the neural network layer is ; Result integration step: Receive the output of the neural network modeling step. and Through weighted formula Fusion prediction results ， Weighting coefficient and The prediction accuracy of each model is set based on historical data; Will Compared with the preset fault threshold range [ θ 1, θ 2] Comparison: when ∈[0, θ 1 Output a normal status flag; when ∈( θ 1, θ 2] Output warning status flag; when > θ 2. Output fault status flags; The results display module steps are as follows: Receive the device status flags output by the results integration module; When a fault condition is identified, a list of components to be repaired is generated based on the chaotic markers of the IMF sequence output by the characteristic analysis module. The list maps physical components according to the sensor type of the chaotic sequence source. The chaotic sequence related to the temperature sensor corresponds to the maintenance items for the condenser or evaporator. The chaotic sequence related to the pressure sensor corresponds to the compressor or valve maintenance items; The chaotic sequence related to the flow sensor corresponds to the maintenance items for the water pump or pipeline. When marked as alarm status, a system checklist is generated that includes sensor group calibration items and associated device test items; When marked as normal, a performance evaluation report is output, which includes the output of the neural network modeling module. and / or ; The device's real-time operating curves, historical status change trends, and current maintenance recommendations are displayed synchronously through a visual interface.

Images