A host fault diagnosis method based on digital twin parameter deviation
Patent Information
- Application Number
- CN202611049565.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-15
- Publication Date
- 2026-09-29
AI Technical Summary
例如，船舶主机运行参数呈现出显著的非线性动态特征，高负荷下的正常排气温度可能远高于低负荷下的故障温度，这种强变工况特性使得基于原始数据的状态识别极易产生混淆与误判；另一方面，实船运行数据呈现极端的分布不平衡，海量的正常运行数据与匮乏的故障样本形成了鲜明对比，导致依赖大量标签数据的深度学习算法难以捕捉到关键的故障特征
[0029]与现有技术相比，本发明通过融合物理机理与神经网络补偿构建高保真数字孪生模型，实现了机理通用性与数据高精度的优势互补。在此基础上，创新性地采用“参数偏离度”作为诊断特征，有效实现了数据的动态归一化，彻底消除了主机变工况运行对诊断精度的干扰；同时利用孪生模型进行多维度的故障注入仿真，生成海量虚拟故障样本，克服了实船运行中故障数据匮乏、难以覆盖全故障类型的瓶颈。最终，结合概率神经网络（PNN）优异的非线性映射能力与计算效率，在保证实时性的前提下实现了高达99.43%的故障诊断准确率，显著优于传统的ELM及SVM算法，为船舶主机的智能运维提供了精准、可靠的技术支撑。
Smart Images
Figure CN122838154A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of host fault diagnosis technology, and more specifically to a host fault diagnosis method based on the deviation of digital twin parameters. Background Technology
[0002] Main engines operate in a dynamic and ever-changing environment, and their fault diagnosis faces dual constraints in terms of data quality and quantity. For example, the operating parameters of ship main engines exhibit significant nonlinear dynamic characteristics. The normal exhaust temperature under high load may be much higher than the fault temperature under low load. This highly variable operating condition makes state identification based on raw data prone to confusion and misjudgment. On the other hand, the actual ship operating data shows an extreme imbalance in distribution. The massive amount of normal operating data contrasts sharply with the scarce fault samples, making it difficult for deep learning algorithms that rely on a large amount of labeled data to capture key fault features.
[0003] Faced with the aforementioned technical bottlenecks, existing single diagnostic models are no longer sufficient to meet the needs of precise operation and maintenance. While traditional mechanistic models offer strong physical interpretability, their modeling process is extremely complex and computationally resource-intensive. Furthermore, purely data-driven models often exhibit significantly reduced generalization ability when encountering variable operating conditions not covered in the training set. Therefore, developing a hybrid fault diagnosis method that can embed mechanistic knowledge to eliminate operating condition interference and can be effectively driven by small samples is crucial. Summary of the Invention
[0004] This invention provides a host fault diagnosis method based on digital twin parameter deviation. The purpose is to eliminate operating condition interference and overcome the scarcity of fault samples by using digital twin benchmarks and parameter deviation characteristics, thereby achieving accurate online diagnosis of host faults.
[0005] The above objectives are achieved through the following technical solutions:
[0006] A host fault diagnosis method based on digital twin parameter deviation includes the following steps:
[0007] Step 1: Construct a host digital twin benchmark model; A host digital twin model is established using a combination of mechanism analysis and data-driven approach to obtain the theoretical benchmark values of the host under different operating conditions;
[0008] Step 2: Construct a fault diagnosis knowledge base; Based on the digital twin model established in Step 1, fault simulation is performed by introducing and adjusting performance factors that characterize the health status of each component, generating a training sample set that includes normal state and multiple fault modes.
[0009] Step 3: Calculate parameter deviation features; collect real-time monitoring data from the host, calculate the parameter deviation of the real-time monitoring data relative to the theoretical benchmark value, and construct a deviation feature vector;
[0010] Step 4: Train the probabilistic neural network diagnostic model; Construct a probabilistic neural network by converting the samples into deviation feature vectors using the training sample set generated in Step 2, and then training the network.
[0011] Step 5: Online fault diagnosis; real-time acquisition of the ship's main engine operating condition parameters and monitoring parameters, obtaining the theoretical benchmark value under the current operating condition through a digital twin benchmark model, calculating the real-time deviation feature vector and inputting it into a trained probabilistic neural network, and outputting the fault diagnosis result of the main engine.
[0012] The construction of the host digital twin benchmark model in step 1 specifically includes: constructing a mechanism model based on the host thermodynamic principle, using the host speed and fuel throttle scale as inputs, and outputting the preliminary thermodynamic parameters of each component; using a neural network to construct an error compensation model to correct the output deviation of the mechanism model, and using the corrected result as a theoretical benchmark value that can map the normal operating state of the host.
[0013] The performance factors involved in step 2 include: compressor efficiency factor, turbine efficiency factor, intercooler cooling efficiency factor, indicated thermal efficiency factor, and scavenging coefficient factor; the fault simulation is to simulate compressor blade damage, turbine fouling, intercooler blockage, excessively high cooling water temperature, and abnormal exhaust temperature faults by manually adjusting the values of the above performance factors in a digital twin model, thereby obtaining the corresponding monitoring parameter outputs under different fault modes.
[0014] The formula for calculating the parameter deviation in step 3 is as follows:
[0015]
[0016] Among them, y real For the actual value collected (or the fault simulation value), y twin This is the theoretical baseline value for the digital twin model at the same current speed and power.
[0017] When constructing the deviation feature vector in step 3, the key parameters selected include at least the following 8 items: turbocharger speed A, turbocharger speed B, intercooler outlet pressure, intercooler outlet temperature, exhaust temperature A, exhaust temperature B, turbine A outlet temperature, and turbine B outlet temperature.
[0018] The probabilistic neural network in step 4 includes an input layer, a pattern layer, a summation layer, and an output layer; wherein, the number of nodes in the input layer corresponds to the number of key parameters, and the number of nodes in the output layer corresponds to the number of fault categories; the output layer outputs the probability of the corresponding fault category, and the one with the highest probability is taken as the final diagnosis result.
[0019] The pattern layer neurons calculate the Euclidean distance between the input vector and the sample vector, and are activated by a Gaussian kernel function. The smoothing factor of the Gaussian kernel function is determined through experimental optimization.
[0020] The generation of the training sample set in step 2 includes: setting different operating conditions, adjusting the value of the performance factor under each operating condition, recording the monitoring parameters output by the model, and forming a sample database containing normal states and multiple fault modes.
[0021] The fault diagnosis result in step 5 includes the current health status of the host or the specific fault type. When the diagnosis result is a fault, the corresponding fault category label and fault probability are output synchronously.
[0022] A host fault diagnosis system based on digital twin parameter deviation includes:
[0023] The data acquisition module is used to collect real-time monitoring data and operating parameters of the host computer.
[0024] The digital twin benchmark model module is connected to the data acquisition module and is used to establish a host digital twin model by combining mechanism analysis and data-driven approach, and to obtain the theoretical benchmark values of the host under different operating conditions.
[0025] The fault diagnosis knowledge base module, connected to the digital twin benchmark model module, is used to simulate faults by introducing and adjusting performance factors that characterize the health status of each component, and to generate a training sample set that includes normal state and multiple fault modes.
[0026] The deviation calculation module, connected to the data acquisition module and the digital twin benchmark model module, is used to calculate the parameter deviation of the real-time monitoring data relative to the theoretical benchmark value and construct a deviation feature vector.
[0027] The probabilistic neural network diagnostic module, connected to the deviation calculation module and the fault diagnosis knowledge base module, is used to construct a probabilistic neural network, train the network using the training sample set, and output the fault diagnosis results of the host.
[0028] The beneficial effects of the host fault diagnosis method based on digital twin parameter deviation of the present invention are as follows:
[0029] Compared with existing technologies, this invention constructs a high-fidelity digital twin model by integrating physical mechanisms and neural network compensation, achieving a complementary advantage of mechanism universality and high data accuracy. Based on this, it innovatively adopts "parameter deviation" as a diagnostic feature, effectively realizing dynamic data normalization and completely eliminating the interference of main engine operating conditions on diagnostic accuracy. Simultaneously, it utilizes the twin model for multi-dimensional fault injection simulation, generating massive virtual fault samples, overcoming the bottleneck of scarce fault data and difficulty in covering all fault types in actual ship operation. Finally, combining the excellent nonlinear mapping capability and computational efficiency of probabilistic neural networks (PNN), a fault diagnosis accuracy of up to 99.43% is achieved while ensuring real-time performance, significantly outperforming traditional ELM and SVM algorithms, providing precise and reliable technical support for the intelligent operation and maintenance of ship main engines. Attached Figure Description
[0030] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.
[0031] Figure 1 This is a schematic diagram of the steps of a host fault diagnosis method based on digital twin parameter deviation according to the present invention;
[0032] Figure 2 This is a flowchart of a host fault diagnosis method based on digital twin parameter deviation according to the present invention;
[0033] Figure 3 This is a schematic diagram illustrating the principle of hybrid modeling for digital twins according to the present invention;
[0034] Figure 4 This is a flowchart of the present invention for fault injection;
[0035] Figure 5 This is a comparison diagram of the principle of parameter deviation feature extraction of the present invention;
[0036] Figure 6 This is a diagram of the PNN neural network topology for fault diagnosis according to the present invention. Detailed Implementation
[0037] like Figures 1 to 6 As shown, in order to achieve the technical effect of "real-time monitoring and intelligent control of the gas turbine through real-time evaluation of the various components and overall status of the gas turbine, ensuring that the gas turbine can be restored and maintained in its optimal operating state under the premise of economic efficiency, ease of operation and environmental compliance", the following describes in detail a main engine fault diagnosis method based on digital twin parameter deviation, including the following steps:
[0038] Step 1: In order to obtain the standard baseline values (i.e. "health values") of the host operating parameters under different operating conditions, a digital twin baseline model driven by a combination of mechanism analysis and data-driven approach is constructed to obtain the theoretical baseline values of the host under different operating conditions.
[0039] First, based on the thermodynamic principles of the main engine, an average value modeling method is used to establish the main engine's mechanistic model. Average value models are constructed for components including the air filter, compressor, intercooler, cylinder block, and exhaust turbine. The main engine speed and fuel throttle position are used as inputs, and the theoretical thermodynamic parameters of each component are output.
[0040] Specifically, the input parameters include the main engine speed n. d Main unit power P w (or fuel throttle position and other operating condition parameters), its key components are modeled as follows:
[0041] The compressor model is based on the compressor characteristic curve and the first law of thermodynamics to calculate the compressor outlet temperature T. out_c and power consumption W c Introducing compressor efficiency As a performance parameter.
[0042] The compressor model is based on the compressor characteristic curve and the first law of thermodynamics to calculate the compressor outlet temperature T. out_c and power consumption W c Introducing compressor efficiency As a performance parameter.
[0043] The intercooler model is based on the compressor outlet parameters and cooling water temperature to calculate the intercooler outlet temperature T. out_ac Introducing cooling efficiency As a performance parameter.
[0044] The turbine model is based on the exhaust temperature T. p and exhaust gas flow As input, calculate the turbine outlet temperature. and turbine output power.
[0045] The above component models are coupled with each other, and the power output of the turbine is balanced with the power consumed by the compressor. The turbocharger speed is calculated by the turbocharger rotor dynamics equation.
[0046] Due to inherent errors in the simplified mechanistic model, this embodiment further utilizes a multilayer perceptron (MLP) neural network to construct an error compensation model. Operating parameters (such as speed and power) are used as input, and the calculated deviation between the mechanistic model output and the actual monitored values is used as the output for training. In online application, the operating parameters are input into the trained compensation model to correct the output deviation of the mechanistic model, ultimately outputting a high-precision theoretical reference value that accurately maps the normal operating state of the host machine. ).
[0047] Step 2: Since fault samples are extremely scarce in actual ship operation data, the digital twin model established in Step 1 is used to perform fault injection simulation, generate a training sample set containing normal state and multiple fault modes, and build a fault diagnosis knowledge base.
[0048] First, performance factors characterizing the health status of each component are introduced as adjustment parameters for fault injection. Under normal conditions, the value of each factor is 1. Specifically, these include: compressor efficiency factors (K1, K2), which characterize the efficiency of compressor A and compressor B, respectively; turbine efficiency factors (K3, K4), which characterize the efficiency of turbine A and turbine B, respectively; intercooler cooling efficiency factor (K5), which characterizes the intercooler cooling efficiency; indicated thermal efficiency factor (K6), which characterizes the indicated thermal efficiency and is associated with cylinder liner water temperature; and scavenging coefficient factor (K7), which characterizes the scavenging coefficient and is associated with exhaust temperature anomalies.
[0049] Then, the values of the aforementioned performance factors are manually adjusted in the digital twin model to simulate typical faults of different degrees: when simulating compressor A / B blade damage, the corresponding compressor efficiency factor is gradually reduced from the normal value of 1.0 (e.g., reduced to 0.9); when simulating turbine A / B fouling, the corresponding turbine efficiency factor is gradually reduced from the normal value; when simulating intercooler pipe blockage, the intercooler cooling efficiency factor is gradually reduced from the normal value; when simulating excessively high coolant temperature, the indicated thermal efficiency factor is gradually reduced from the normal value; when simulating abnormal exhaust temperature (abnormal scavenging), the scavenging coefficient factor is gradually reduced from the normal value. Through the above fault injection, the corresponding monitoring parameter outputs under different fault modes are obtained.
[0050] Fault injection formula: Different degrees of faults can be simulated by modifying the above factors:
[0051] Simulated compressor A / B failure: Updated efficiency, where .
[0052]
[0053] Simulated turbine A / B failure: updated efficiency, where .
[0054]
[0055] Simulated intercooler pipe blockage: Updated cooling efficiency, where .
[0056]
[0057] Simulated overheating of cooling water: Updated indicated thermal efficiency, where .
[0058]
[0059] Simulated exhaust temperature fault (scavenging abnormality): Updated scavenging coefficient, where .
[0060]
[0061] Finally, different operating points (e.g., 25%, 50%, 75%, 100% load) are set, and the above K value is adjusted under each operating condition to cover faults of different severity. The monitoring parameters output by the model are recorded, and the fault state data and normal state data generated by the simulation are collected to construct a sample database containing fault labels as a training sample set for subsequent diagnostic models.
[0062] Step 3: In order to eliminate the influence of changes in operating conditions on the absolute values of parameters and make the diagnostic features independent of operating conditions, the parameter deviation is extracted as the feature vector for fault diagnosis.
[0063] Parameter deviation is defined as the degree of relative deviation between real-time monitoring data and the theoretical baseline value of the digital twin model under the same current operating conditions. The calculation formula is as follows:
[0064]
[0065] Among them, y real For the actual value collected (or the fault simulation value), y twin This is the theoretical baseline value for the digital twin model at the same current speed and power.
[0066] When constructing the deviation feature vector, eight key parameters that are most sensitive to faults are selected. Specifically, these include: turbocharger speed deviation A, turbocharger speed deviation B, intercooler outlet pressure deviation, intercooler outlet temperature deviation, exhaust temperature deviation A, exhaust temperature deviation B, turbine outlet temperature deviation A, and turbine outlet temperature deviation B. These eight deviation parameters together constitute a deviation feature vector, which serves as the input feature for the subsequent probabilistic neural network diagnostic model.
[0067] Step 4: Use a probabilistic neural network (PNN) for fault classification. Using the training sample set generated in Step 2, convert the samples into deviation feature vectors as described in Step 3 and then train the network.
[0068] The PNN network consists of four layers: the input layer has 8 nodes, corresponding to the 8 deviation features mentioned above; the pattern layer (hidden layer) has the same number of nodes as the total number of training samples, and each neuron calculates the Euclidean distance between the input vector and the sample vector, and is activated by a Gaussian kernel function. The activation function is:
[0069]
[0070] in, The smoothing factor is determined experimentally in this embodiment; the summation layer has 8 nodes, corresponding to 8 health status categories. This layer performs a weighted summation of the outputs of pattern layer neurons belonging to the same category; the output layer uses a competitive transfer function, and the category with the largest value in the output summation layer is taken as the final diagnosis result.
[0071] The fault category labels for the output layer are set as follows: Label0 represents normal; Label1 represents turbine A fault; Label2 represents turbine B fault; Label3 represents compressor A fault; Label4 represents compressor B fault; Label5 represents intercooler fault; Label6 represents cooling system abnormality; Label7 represents main engine exhaust temperature abnormality.
[0072] Step 5: In the actual ship application phase, the system executes the online fault diagnosis process. Real-time data is collected from the ship's main engine operating parameters (such as speed and power) and sensor data from the aforementioned eight key monitoring points. The real-time operating parameters are input into the digital twin benchmark model constructed in step S1 to obtain the theoretical benchmark values for the eight monitoring points under the current operating conditions. The deviation of the real-time monitoring parameters from the theoretical benchmark values is calculated according to the deviation calculation formula described in step S3. Construct a real-time deviation feature vector; input the real-time deviation feature vector into the probabilistic neural network trained in step S4, and output the probability of various faults corresponding to the current state from the output layer. Take the one with the highest probability as the final fault diagnosis result and output the current health status or specific fault type of the host.
[0073] For example, if the input features show that the intercooler outlet temperature deviation is significantly positive (e.g., +7.4%) and the temperature of subsequent components shows a positive deviation, the PNN model will output Label5 (intercooler fault) to achieve real-time intelligent fault diagnosis of the host.
[0074] It should be noted that the steps shown in the above process or in the flowchart of the accompanying figures can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0075] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.
Claims
1. A host fault diagnosis method based on digital twin parameter deviation, characterized in that, Includes the following steps: Step 1: Construct a host digital twin benchmark model; A host digital twin model is established using a combination of mechanism analysis and data-driven approach to obtain the theoretical benchmark values of the host under different operating conditions; Step 2: Build a fault diagnosis knowledge base; Based on the digital twin model established in step 1, fault simulation is performed by introducing and adjusting performance factors that characterize the health status of each component, generating a training sample set that includes normal state and multiple fault modes. Step 3: Calculate the parameter deviation characteristics; Collect real-time monitoring data from the host, calculate the parameter deviation of the real-time monitoring data relative to the theoretical benchmark value, and construct a deviation feature vector; Step 4: Train the probabilistic neural network diagnostic model; Construct a probabilistic neural network by converting the samples into deviation feature vectors using the training sample set generated in Step 2, and then training the network. Step 5: Online fault diagnosis; real-time acquisition of the ship's main engine operating condition parameters and monitoring parameters, obtaining the theoretical benchmark value under the current operating condition through a digital twin benchmark model, calculating the real-time deviation feature vector and inputting it into a trained probabilistic neural network, and outputting the fault diagnosis result of the main engine.
2. The method according to claim 1, characterized in that, The construction of the host digital twin benchmark model in step 1 specifically includes: constructing a mechanism model based on the host thermodynamic principle, using the host speed and fuel throttle scale as inputs, and outputting the preliminary thermodynamic parameters of each component; using a neural network to construct an error compensation model to correct the output deviation of the mechanism model, and using the corrected result as a theoretical benchmark value that can map the normal operating state of the host.
3. The method according to claim 1, characterized in that, The performance factors involved in step 2 include: compressor efficiency factor, turbine efficiency factor, intercooler cooling efficiency factor, indicated thermal efficiency factor, and scavenging coefficient factor; the fault simulation is to simulate compressor blade damage, turbine fouling, intercooler blockage, excessively high cooling water temperature, and abnormal exhaust temperature faults by manually adjusting the values of the above performance factors in a digital twin model, thereby obtaining the corresponding monitoring parameter outputs under different fault modes.
4. The method according to claim 1, characterized in that, The formula for calculating the parameter deviation in step 3 is as follows: Among them, y real For the actual value collected (or the fault simulation value), y twin This is the theoretical baseline value for the digital twin model at the same current speed and power.
5. The method according to claim 1, characterized in that, When constructing the deviation feature vector in step 3, the key parameters selected include at least the following 8 items: turbocharger speed A, turbocharger speed B, intercooler outlet pressure, intercooler outlet temperature, exhaust temperature A, exhaust temperature B, turbine A outlet temperature, and turbine B outlet temperature.
6. The method according to claim 1, characterized in that, The probabilistic neural network in step 4 includes an input layer, a pattern layer, a summation layer, and an output layer; wherein, the number of nodes in the input layer corresponds to the number of key parameters, and the number of nodes in the output layer corresponds to the number of fault categories; the output layer outputs the probability of the corresponding fault category, and the one with the highest probability is taken as the final diagnosis result.
7. The method according to claim 6, characterized in that, The pattern layer neurons calculate the Euclidean distance between the input vector and the sample vector, and are activated by a Gaussian kernel function. The smoothing factor of the Gaussian kernel function is determined through experimental optimization.
8. The method according to claim 1, characterized in that, The generation of the training sample set in step 2 includes: setting different operating conditions, adjusting the value of the performance factor under each operating condition, recording the monitoring parameters output by the model, and forming a sample database containing normal states and multiple fault modes.
9. The method according to claim 1, characterized in that, The fault diagnosis result in step 5 includes the current health status of the host or the specific fault type. When the diagnosis result is a fault, the corresponding fault category label and fault probability are output synchronously.
10. A host fault diagnosis system based on digital twin parameter deviation, characterized in that, include: The data acquisition module is used to collect real-time monitoring data and operating parameters of the host computer. The digital twin benchmark model module is connected to the data acquisition module and is used to establish a host digital twin model by combining mechanism analysis and data-driven approach, and to obtain the theoretical benchmark values of the host under different operating conditions. The fault diagnosis knowledge base module, connected to the digital twin benchmark model module, is used to simulate faults by introducing and adjusting performance factors that characterize the health status of each component, and to generate a training sample set that includes normal state and multiple fault modes. The deviation calculation module, connected to the data acquisition module and the digital twin benchmark model module, is used to calculate the parameter deviation of the real-time monitoring data relative to the theoretical benchmark value and construct a deviation feature vector. The probabilistic neural network diagnostic module, connected to the deviation calculation module and the fault diagnosis knowledge base module, is used to construct a probabilistic neural network, train the network using the training sample set, and output the fault diagnosis results of the host.