Industrial equipment fault intelligent diagnosis method based on multi-sensor fusion AI

Abstract

The invention discloses an industrial equipment fault intelligent diagnosis method based on multi-sensor fusion AI, and belongs to the technical field of industrial equipment state monitoring and fault diagnosis, and the method specifically comprises the following steps: collecting an original measurement signal vector and an external variable vector of target industrial equipment; according to the external variable vector, a physical information neural network is adopted to generate a health signal baseline vector; combining the original measurement signal vector with the health signal baseline vector, and carrying out vector subtraction to generate a residual signal vector; inputting the sequence of the residual signal vectors into a pre-trained fault diagnosis model to generate a health state index; according to the size relation between the health state index and the preset fault alarm threshold value, a final equipment state diagnosis result is generated, accurate and reliable diagnosis of early weak faults of the industrial equipment is achieved, and the false alarm rate and the missing report rate are greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment condition monitoring and fault diagnosis, specifically to an intelligent fault diagnosis method for industrial equipment based on multi-sensor fusion AI. Background Technology

[0002] In high-end manufacturing industries such as mobile phone components, automotive parts, and precision electronic devices, the continuous and stable operation of intelligent equipment such as CNC machine tools and industrial robots is the cornerstone of ensuring product quality and production efficiency. Against this backdrop, many companies are committed to providing intelligent manufacturing solutions, deploying industrial IoT and big data analytics platforms to achieve real-time monitoring of the health status of production equipment and fault early warning—a core requirement of "Industry 4.0" and "Made in China 2025."

[0003] To achieve intelligent equipment diagnostics, the industry widely adopts artificial intelligence (AI) methods based on multi-sensor information fusion. However, existing technologies generally face an inherent technical contradiction in practical applications: on the one hand, to capture subtle fault characteristics such as minute tool wear and early performance degradation of spindle bearings, the sensitivity of the diagnostic system needs to be greatly improved. On the other hand, increased sensitivity often leads to a decrease in the system's robustness to changes in normal operating conditions. For example, processing different batches of parts, changing cutting fluid, fluctuations in workshop ambient temperature, and even slight changes in power grid voltage can all cause normal fluctuations in equipment sensor signals, which may be misjudged as faults by the system, generating a large number of false alarms. Conversely, if sensitivity is deliberately reduced to avoid false alarms, valuable early fault warning windows will be missed, potentially leading to batch defects or even equipment damage.

[0004] The root of this contradiction lies in the fact that fluctuations in equipment data characteristics caused by external variables such as production task switching, material characteristic differences, and changes in the workshop environment are highly similar in signal morphology to fluctuations caused by early, minor faults within the equipment (such as performance degradation and component wear), creating a difficult-to-distinguish "modal aliasing" effect. Traditional, purely data-driven AI models, lacking a deep understanding of the underlying physical mechanisms of the processing, cannot effectively distinguish between these two vastly different signal changes. They simply blindly fit all observed data fluctuations, resulting in poor diagnostic performance under complex and ever-changing real-world production conditions. Therefore, the industry urgently needs a new intelligent diagnostic paradigm that can effectively "decouple" the true fault characteristics from strong background interference (i.e., normal operating conditions and environmental changes). Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent fault diagnosis method for industrial equipment based on multi-sensor fusion AI, so as to solve the problems mentioned in the background art.

[0006] The technical solution of the present invention specifically includes the following steps: Acquire the raw measurement signal vector and external variable vector of the target industrial equipment; Based on the external variable vector, a physical information neural network is used to generate a baseline vector of health signals; The original measurement signal vector and the health signal baseline vector are combined, and a vector subtraction operation is performed to generate the residual signal vector; The sequence of residual signal vectors is input into a pre-trained fault diagnosis model to generate health status indicators. Based on the relationship between health status indicators and preset fault alarm thresholds, the final equipment status diagnosis result is generated; When the health status index is greater than the fault alarm threshold, the equipment status diagnosis result is that the equipment has failed; When the health status index is less than or equal to the fault alarm threshold, the equipment status diagnosis result is that the equipment is in a healthy state.

[0007] Preferably, the step of acquiring the original measurement signal vector specifically includes: Collect fluid dynamic fingerprints characterizing the internal fluid state of the system and construct the first type of feature vector; Acoustic imprints of structural vibrations characterizing the health status of key rotating components are collected and used to construct a second type of feature vector; The first type of feature vector is concatenated with the second type of feature vector to generate the original measurement signal vector.

[0008] Preferably, the external variable vector includes ambient temperature and equipment operating command load.

[0009] Preferably, the steps for generating the physical information neural network include: Construct a physical information neural network, taking the external variable vector as input, and outputting a health signal baseline vector with the same dimension and physical units as the original measured signal vector; A composite loss function is used to train the physical information neural network.

[0010] Preferably, the steps for constructing the composite loss function include: Define the data-driven terms, which are used to fit the true signals under historical health conditions; Define the physical regularization term, which is used to force the prediction results of the physical information neural network to follow the preset physical laws; By combining the data-driven term and the physical regularization term, a composite loss function is formed.

[0011] Preferably, the training process of the composite loss function further includes: Before calculating the composite loss function, the data-driven terms are dimensionless using a diagonal matrix, where the diagonal elements of the diagonal matrix are the reciprocals of the feature scales of each signal component. Before calculating the composite loss function, the physical regularization term is dimensionless by using the physical residual scaling matrix.

[0012] Preferably, the fault diagnosis model is a deep learning model based on the Transformer model or the autoencoder model, which is used to capture patterns in time series data.

[0013] Preferably, the steps for setting the fault alarm threshold include: The system collects health status indicators corresponding to the equipment operating in a healthy state and constructs an experience accumulation distribution function. Based on the preset acceptable false alarm rate, the inverse function of the empirical cumulative distribution function is queried to determine the fault alarm threshold.

[0014] This invention provides an improved intelligent fault diagnosis method for industrial equipment based on multi-sensor fusion AI, which has the following improvements and advantages compared with the prior art: 1. By acquiring the original measurement signal vector and external variable vector of the target industrial equipment, and generating a health signal baseline vector based on the external variable vector using a physical information neural network, and then generating a residual signal vector through vector subtraction, this technical approach constructs a physical-causal decoupling paradigm, changing the traditional signal suppression approach. It does not suppress interference, but rather separates the normal signal fluctuations caused by changes in the external environment and workload from the original measurement signal at the causal level. This allows the subsequent fault diagnosis model to be analyzed on the pure residual signal with a very high signal-to-noise ratio, resolving the contradiction between high sensitivity and high robustness, achieving accurate and reliable diagnosis of early weak faults in industrial equipment, and significantly reducing the false alarm rate and missed alarm rate. 2. The original measurement signal vector is specifically defined as a combination of a fluid dynamic fingerprint representing the internal fluid state of the system and a structural vibration acoustic imprint representing the health state of key rotating components. This clear definition of multimodal information sources ensures that the collected signals can comprehensively and deeply reflect the internal health state of the equipment. At the same time, by specifying the external variable vector as ambient temperature and equipment operation command load, the physical information neural network can specifically learn and isolate these two main sources of interference, providing the necessary and sufficient input information for achieving high-precision causal decoupling and improving the accuracy of the entire diagnostic framework. 3. A composite loss function is used to train the physical information neural network. This composite loss function combines a data-driven term for fitting historical health signals with a physical regularization term that forces the network prediction to comply with physical laws. Both terms are dimensionless during training. This design ensures that while the neural network learns the data distribution, its internal logic strictly adheres to known physical laws. As a result, the trained model is no longer a pure black box, but possesses physical interpretability and strong generalization ability. It can accurately predict the health baseline under external operating condition combinations that have not been seen in the training data, ensuring the purity of the subsequent residual signal and the reliability of the diagnostic results. 4. The method uses a deep learning model based on the Transformer model or autoencoder model as the fault diagnosis model. Such advanced models can efficiently capture long-distance dependencies and weak dynamic patterns in time series data. Since the input to the model is a pure residual signal that has been freed from strong environmental interference, the model's learning task is more focused, enabling it to capture early fault features that are completely submerged by noise in the original signal with extremely high sensitivity, thereby achieving earlier and more accurate fault warning. 5. An empirical cumulative distribution function is constructed by collecting indicators under healthy conditions, and its inverse function is queried based on a preset acceptable false alarm rate to determine the fault alarm threshold. This statistically based threshold setting process directly links alarm decisions with controllable false alarm performance, avoiding the subjectivity and instability caused by manually setting thresholds. It ensures that the diagnostic system can provide a statistically guaranteed extremely low false alarm rate while maintaining high sensitivity, significantly improving the credibility and practical value of diagnostic results in actual industrial applications. Attached Figure Description

[0015] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. Example 1

[0017] Please see Figure 1 This invention provides an intelligent fault diagnosis method for industrial equipment based on multi-sensor fusion AI, specifically including the following steps: Acquire the raw measurement signal vector and external variable vector of the target industrial equipment; Based on the external variable vector, a physical information neural network is used to generate a baseline vector of health signals; The original measurement signal vector and the health signal baseline vector are combined, and a vector subtraction operation is performed to generate the residual signal vector; The sequence of residual signal vectors is input into a pre-trained fault diagnosis model to generate health status indicators. Based on the relationship between health status indicators and preset fault alarm thresholds, the final equipment status diagnosis result is generated; When the health status index is greater than the fault alarm threshold, the equipment status diagnosis result is that the equipment has failed; When the health status index is less than or equal to the fault alarm threshold, the equipment status diagnosis result is that the equipment is in a healthy state. The external variable vector can include ambient temperature and equipment operation command load, thus better connecting with the subsequent content; A method for intelligent fault diagnosis of industrial equipment based on multi-sensor fusion AI is provided. The purpose of this method is to accurately distinguish between signal fluctuations caused by external environmental interference and signal fluctuations caused by early internal faults of the equipment through a physical-causal decoupling paradigm, thereby significantly improving the detection sensitivity of weak faults while ensuring the robustness of the diagnostic system. For specific industrial equipment, such as a hydraulic power unit (HPU), the implementation process of this method is initiated. The initial step is to acquire the original measurement signal vector and external variable vector of the target industrial equipment. Here, the original measurement signal vector includes instantaneous measurements and dynamic characteristics calculated through instantaneous measurements. The external variable vector is a data vector containing environmental or operating condition factors that affect the operating status of the equipment but are not generated by the equipment itself. After the acquisition step is completed, the acquired external variable vector is used as the input of the physical information neural network. This network is designed to learn and predict the baseline of normal fluctuations in device signals caused only by changes in external variables under a healthy state. Through this network, a healthy signal baseline vector is generated. Here, the healthy signal baseline vector is a predicted data vector with the same dimension and physical units as the original measured signal vector. It represents the sensor signal readings that a fully healthy device should exhibit under the current external environment and operating load. The real-time acquired raw measurement signal vector is subtracted from the health signal baseline vector generated in the previous steps to generate a residual signal vector. Here, the residual signal vector is a signal vector obtained by stripping the health baseline signal from the raw measurement signal. Theoretically, it only contains abnormal signal components caused by changes in the equipment's internal health status, such as early failures, while the influence of the environment and operating conditions has been eliminated. This calculation process is expressed by the following formula: ; in for The residual signal vector at time step 1. For real-time data collection via sensors The original measurement signal vector at time 1. This is calculated by the physical information neural network in the preceding steps. The baseline vector of the health signal predicted at any given time; Over time, the continuously generated residual signal vectors form a time series. This sequence of residual signal vectors is input into a pre-trained fault diagnosis model, which aims to identify weak dynamic features related to specific fault modes from the pure residual signal time series. This model analyzes and processes the input residual signal sequence to generate health status indicators. Here, the health status indicators are standardized, dimensionless continuous values ​​whose magnitude is positively correlated with the degree to which the equipment deviates from its healthy state. Based on the relationship between the health status index and the preset fault alarm threshold, the final equipment status diagnosis result is generated. Here, the fault alarm threshold is a critical value set based on statistical principles to distinguish between the healthy and faulty states of the equipment. The judgment logic is as follows: when the health status index is greater than the fault alarm threshold, the equipment status diagnosis result is judged as the equipment malfunctioning; conversely, when the health status index is less than or equal to the fault alarm threshold, the equipment status diagnosis result is judged as the equipment being in a healthy state. The technical advantages of this implementation are as follows: by constructing a complete technical closed loop from acquisition, baseline estimation, residual generation to final diagnosis, and by using physical model prediction to remove environmental interference signals instead of traditional signal suppression, the subsequent diagnostic model can work on residual signals with extremely high signal-to-noise ratios. This solves the inherent contradiction between high sensitivity and high robustness, and enables accurate and reliable diagnosis of early-stage minor faults in industrial equipment, significantly reducing false alarm rates and missed alarm rates. Example 2

[0018] The specific steps for acquiring the raw measurement signal vector include: Collect fluid dynamic fingerprints characterizing the internal fluid state of the system and construct the first type of feature vector; Acoustic imprints of structural vibrations characterizing the health status of key rotating components are collected and used to construct a second type of feature vector; The first type of feature vector is concatenated with the second type of feature vector to generate the original measurement signal vector; The external variable vector includes ambient temperature and equipment operating command load; Based on the method in Example 1, the steps for data acquisition have been specified in order to capture multimodal information that can comprehensively and three-dimensionally reflect the status of the equipment. The steps for acquiring the original measurement signal vector are further refined; a fluid dynamic fingerprint characterizing the internal fluid state of the system is acquired and constitutes a first-class feature vector; here, the first-class feature vector... It is a set of parameters used to describe the core performance of the equipment's hydraulic system. In this embodiment, the vector includes pressure build-up time, flow response delay, pressure decay rate during the pressure holding stage, and hydraulic oil temperature. These parameters together characterize the operating efficiency and sealing performance of the fluid power system. Simultaneously, structural vibration acoustic imprints characterizing the health status of key rotating components are collected and used to construct a second type of feature vector; here, the second type of feature vector... It is a set of parameters used to describe the mechanical health status of rotating parts such as motors and pumps in equipment. In this embodiment, the vector includes the root mean square value of the vibration signal, the amplitude of the characteristic frequency in the spectrum, the signal kurtosis, and the acoustic energy of a specific frequency band. These parameters can effectively reflect the wear, imbalance, or lubrication status of the parts. The first type of feature vector is concatenated with the second type of feature vector to generate the original measurement signal vector. This splicing operation mathematically forms a higher-dimensional unified state vector, ensuring that information from different physical domains can be comprehensively analyzed within the same framework. In addition, external variable vector Specifically defined as including ambient temperature With equipment operation command load Here, the ambient temperature The temperature of the operating environment of the equipment affects key physical parameters such as the viscosity of the hydraulic oil; the load of the equipment's operating commands... The intensity or instructions of the task currently being performed by the equipment directly determine the equipment's operating power and stress level; In this embodiment, to simplify the model, the external variables mainly consider ambient temperature and workload. However, in practical applications, other important external variables, such as depth and salinity, can be added according to the specific scenario. The advantages of this implementation method are as follows: by clearly defining the multimodal sensor information sources, namely fluid dynamics and vibroacoustics, it ensures that the acquired raw measurement signals can more comprehensively and deeply reflect the internal health status of the equipment; at the same time, by accurately defining the external variables as ambient temperature and operating load, the subsequent physical information network can specifically learn and isolate these two main and strongest sources of interference, providing the necessary and sufficient input information for achieving high-precision causal decoupling, thereby improving the accuracy of the entire diagnostic framework. Example 3

[0019] The steps for generating a physical information neural network include: Construct a physical information neural network, taking the external variable vector as input, and outputting a health signal baseline vector with the same dimension and physical units as the original measured signal vector; The PINN network directly predicts physical quantities, such as viscosity or flow rate, through its output layer, and uses these predictions to compare with physical equations in the loss function, thereby embedding physical laws into the training of the model. The network architecture can be a multilayer perceptron, containing an input layer, multiple hidden layers, and an output layer. For example, the specific number of layers, the number of neurons per layer, and the choice of activation functions can be adjusted according to the actual application scenario and data characteristics. The output layer is connected to the original measurement signal vector. The dimensions are consistent to ensure that subsequent vector subtraction operations have physical meaning; A composite loss function is used to train the physical information neural network; The steps for constructing a composite loss function include: Define the data-driven terms, which are used to fit the true signals under historical health conditions; Define the physical regularization term, which is used to force the prediction results of the physical information neural network to follow the preset physical laws; In this embodiment, for the hydraulic power unit, the physical regularization term... It can include physical equations describing the relationship between hydraulic oil viscosity and temperature and pressure, such as the Arrhenius equation or the Walther equation. For example, a residual term can be introduced to ensure that the viscosity predicted by the network is accurate. With temperature The relationship conforms to the laws of physics: ; in These are constants in the Arrhenius equation; Physical regularization term; : Network-predicted viscosity; Ambient temperature; Combining data-driven terms and physical regularization terms, a composite loss function is constructed; The training process for the composite loss function also includes: Before calculating the composite loss function, a diagonal matrix is ​​used. The data-driven terms are dimensionless, where The diagonal elements are the reciprocals of the characteristic scales of each signal component; Before calculating the composite loss function, the physical regularization term is dimensionless by using the physical residual scaling matrix; Based on the method in Example 1, the process of generating the health signal baseline vector is described in detail; the core of this process is to construct and train a physical information neural network. Its purpose is to serve as a high-precision virtual health model, capable of predicting in real time the signal performance of health devices under any given external conditions. The output vector is designed to be similar to the original measured signal vector. Completely consistent in dimension and physical units, this design is the foundation for ensuring that subsequent vector subtraction operations have clear physical meaning; To make the network It possesses the characteristics of accurate prediction and strong generalization ability, and is trained using a composite loss function; the construction steps of this composite loss function include two core parts; Identify data-driven terms; data-driven terms aim to utilize historical data to fit true signals from historical health states. Determine the physical regularization term; the physical regularization term aims to encode the known physical operating laws of the equipment into the training process, so as to force the prediction results of the physical information neural network to follow the preset physical laws; By combining data-driven terms and physical regularization terms, a composite loss function is constructed. Since different signal components and physical equations have different physical dimensions, direct addition will cause the training process to be dominated by a term with a larger dimension. To solve this problem, the training process of the composite loss function also includes a dimensionless step. Before calculating the final loss value, a data scaling matrix is ​​used to dimensionlessly process the data-driven terms, and a physical residual scaling matrix is ​​used to dimensionlessly process the physical regularization terms; the data scaling matrix It is a diagonal matrix, and the diagonal elements are the reciprocals of the characteristic scales of each signal component, which are derived from a large amount of health data. The matrix is ​​constructed as follows: Collect a large number of samples of each signal component from historical health datasets, such as pressure, temperature, and vibration amplitude. For each signal component, calculate its standard deviation or interquartile range as its feature scale. Construct a diagonal matrix using the reciprocals of these feature scales as diagonal elements. This method ensures that the composite loss function... In the calculation, signal components with different physical dimensions are normalized to a similar numerical range, thus avoiding the training process being dominated by a component with a larger dimension. Similarly, the physical residual scaling matrix and Used to normalize the dimensions of residuals from different physical equations; Modified composite loss function It is expressed as: ; in : The corrected composite loss function; These are the trainable parameters of the network. For the corresponding external variable vector, Sample index, Data scaling matrix The total number of samples, For neural network models, and The physical residual function is constructed based on prior knowledge. and For the weight hyperparameters determined by methods such as grid search or Bayesian optimization on independent validation sets, during grid search, and The value range can be set to . The step size is 0.1; or Bayesian optimization can be used to automatically find the optimal value. and This represents the number of placement points selected within the solution domain; : Configuration point index; : Physical residual scaling matrix used to normalize the dimensions of residuals from different physical equations; L2 norm; The physical residual function describes the relationship between structural vibration and the rotational speed and load of rotating components; Among them, the physical residual function and Used to encode physical laws into mathematical equations; for example... It can contain equations describing the relationship between hydraulic oil viscosity and temperature and pressure; when the physical quantities predicted by the network do not satisfy this relationship, the value of the function is not zero, thus resulting in a loss; The gain effect of this implementation method is as follows: by designing a composite loss function that includes data-driven terms and physical regularization terms, and introducing strict dimensionless processing, it is ensured that while the neural network learns the data distribution, its internal logic strictly follows known physical laws. This makes the trained model no longer a pure black box, but has physical interpretability and strong generalization ability. It can accurately predict the health baseline under external operating condition combinations that have not been seen in the training data, and ensure the purity of the subsequent residual signal and the reliability of the diagnostic results. Example 4

[0020] The fault diagnosis model is a deep learning model based on the Transformer model or the autoencoder model, which is used to capture patterns in time series data. When using the Transformer model, its architecture can be adjusted according to the characteristics of the input sequence. For example, multi-layer encoders and multi-head attention mechanisms can be used. During training, optimizers such as AdamW can be used, combined with learning rate scheduling strategies to improve training efficiency and model performance. The model can be trained on a dataset containing 10,000 samples for 50 training epochs. For the Transformer model, the residual signal vector sequence is taken as input, and the dependencies within the sequence are learned through its self-attention mechanism. Its output layer can be a fully connected layer, which maps the features extracted by the attention mechanism to a single health status indicator. When using an autoencoder model, the encoder part can consist of three convolutional layers to extract time-series features; the decoder consists of three deconvolutional layers. The model is trained by minimizing the mean square error between the input sequence and its reconstructed sequence to learn the normal pattern of the residual signal in a healthy state. During inference, the reconstruction error between the input residual sequence and the reconstructed sequence is calculated as an indicator of the health state. ; The autoencoder model is trained to minimize the difference between the input residual sequence and its reconstructed sequence. In a healthy state, the reconstruction error is small; when a fault occurs, the residual signal deviates from the healthy mode, causing the reconstruction error to increase, and this error value serves as an indicator of the health status. Based on the method in Example 1, the fault diagnosis model used to identify fault modes from the residual signal sequence is specifically defined; the pre-trained fault diagnosis model Designed as a high-efficiency pattern recognizer, it is specifically used to capture the weak dynamic patterns contained in residual signal time series data that indicate the initiation and development of specific faults. In this embodiment, the fault diagnosis model is specifically designated as a deep learning model based on either a Transformer model or an autoencoder model. The Transformer model is chosen because its self-attention mechanism can effectively capture long-range dependencies in time-series data, which is crucial for identifying early faults with long evolution cycles and inconspicuous features. The autoencoder model is chosen because of its powerful feature extraction and anomaly detection capabilities; it can learn the normal distribution pattern of healthy residual signals and give high scores to any minor anomalies deviating from this pattern. This model is designed to capture patterns in time-series data, receiving a fixed-length... residual signal time series window As input, and output a single health status indicator ; The model The training process is performed on a labeled dataset containing two types of data: one is a health residual sequence extracted during the healthy operation of the equipment, and the other is a fault residual sequence containing various known fault types. Through supervised learning, the model parameters... The model is optimized to output health status indicators with significant discriminative power for different categories of input sequences; The advantages of this implementation are as follows: by employing advanced deep learning models such as Transformer or autoencoder, the system's ability to automatically learn and identify complex fault modes from high-dimensional time series is greatly enhanced; since the input to the model is a pure residual signal that has been freed from strong environmental interference, the model's learning task becomes more focused, enabling it to capture early fault features in the original signal that are completely submerged by noise with extremely high sensitivity, thereby achieving earlier and more accurate fault warnings.

[0021] The steps for setting fault alarm thresholds include: The system collects health status indicators corresponding to the equipment operating in a healthy state and constructs an experience accumulation distribution function. Based on the preset acceptable false alarm rate, the inverse function of the empirical cumulative distribution function is queried to determine the fault alarm threshold; In health data statistics for a specific device and operating condition, when the acceptable false alarm rate is set to... At that time, the fault alarm threshold pass The calculation shows that its value could be 0.85; this would make the value verifiable. Based on the method in Example 1, the method for setting the fault alarm threshold has been specified and standardized. The purpose is to abandon the manual setting that relies on experience and establish an automated alarm decision-making mechanism with clear statistical significance and controllable performance. The steps for setting fault alarm thresholds are as follows: Collect health status indicators corresponding to the equipment operating under a large volume of healthy conditions over a long period. This forms a representation of the health status of the equipment. Value sample set; In practical applications, it is recommended to collect health data from the equipment under typical operating conditions for at least 200 hours to ensure that the sample set comprehensively reflects the statistical distribution of the equipment's health status. This process should be repeated under multiple combinations of different loads and ambient temperatures to improve the representativeness of the sample set. By performing statistical analysis on this large sample set, a more robust empirical cumulative distribution function can be constructed. This ensures that the final set fault alarm threshold is met. It is reliable; Based on this sample set, an empirical cumulative distribution function for health status indicators was constructed. ; Empirical cumulative distribution function This describes a health indicator value that is less than or equal to a certain value under healthy conditions. The probability of; Based on the preset acceptable false alarm rate Acceptable false alarm rate This is the tolerable probability of the system falsely reporting a fault when the device is healthy, set by the system designer based on application requirements. For example, it can be set to... ; Through calculation, under a typical operating condition, when the acceptable false alarm rate is set to At that time, the fault alarm threshold It could be 0.85". This would make the instructions more specific and easier to understand; By querying the inverse function of the aforementioned empirical cumulative distribution function To determine the final fault alarm threshold. The calculation is performed using the following formula: ; in The fault alarm threshold, It is the inverse function of the empirical cumulative distribution function. This is a preset acceptable false alarm rate; the underlying logic of this formula lies in finding a critical value. This allows for the calculation of health indicators when the equipment is in good condition. The probability of exceeding this value is exactly equal to the preset false alarm rate. ; The advantages of this implementation method are as follows: by establishing a statistically based threshold setting process, alarm decisions are directly linked to the acceptable false alarm rate, making the false alarm performance of the diagnostic system precise and controllable. This method avoids the subjectivity and instability caused by manually setting thresholds, ensuring that the diagnostic system can provide a statistically guaranteed, extremely low false alarm rate while maintaining high sensitivity, greatly improving the credibility and practical value of the diagnostic results in actual industrial applications.

[0022] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent diagnosis of industrial equipment failure based on multi-sensor fusion AI, characterized in that, Specifically comprising the following steps: Collecting an original measurement signal vector and an external variable vector of a target industrial equipment; Generating a health signal baseline vector by using a physical information neural network according to the external variable vector; Performing vector subtraction operation on the original measurement signal vector and the health signal baseline vector to generate a residual signal vector; Inputting a sequence of the residual signal vector into a pre-trained fault diagnosis model to generate a health state indicator; Generating a final equipment state diagnosis result according to the size relationship between the health state indicator and a preset fault alarm threshold; When the health state indicator is greater than the fault alarm threshold, the equipment state diagnosis result is that the equipment has failed; When the health state indicator is less than or equal to the fault alarm threshold, the equipment state diagnosis result is that the equipment is in a healthy state. 2.The multi-sensor fusion AI-based industrial equipment fault intelligent diagnosis method according to claim 1, characterized in that, The step of collecting the original measurement signal vector specifically comprises: Collecting a fluid dynamics fingerprint representing the internal fluid state of the system to form a first type of feature vector; Collecting a structural vibration acoustic imprint representing the health state of the key rotating components to form a second type of feature vector; Splicing the first type of feature vector and the second type of feature vector to generate the original measurement signal vector. 3.The multi-sensor fusion AI-based industrial equipment fault intelligent diagnosis method according to claim 1, characterized in that, The external variable vector includes the environmental temperature and the equipment operation instruction load. 4.The multi-sensor fusion AI-based industrial equipment fault intelligent diagnosis method according to claim 1, characterized in that, The generation steps of the physical information neural network include: Building a physical information neural network, taking the external variable vector as input, and outputting a health signal baseline vector consistent with the dimension and physical quantity of the original measurement signal vector; Training the physical information neural network by using a composite loss function. 5.The multi-sensor fusion AI-based industrial equipment fault intelligent diagnosis method according to claim 4, characterized in that, The construction steps of the composite loss function include: Determining a data-driven term, which is used to fit the real signal under the historical health state; Determining a physical regularization term, which is used to force the prediction result of the physical information neural network to comply with the preset physical law; Combining the data-driven term and the physical regularization term to form the composite loss function. 6.The multi-sensor fusion AI-based industrial equipment fault intelligent diagnosis method according to claim 5, characterized in that, The training process of the composite loss function further includes: Before calculating the composite loss function, using a diagonal matrix to perform dimensionless processing on the data-driven term, wherein the diagonal elements of the diagonal matrix are the inverses of the characteristic scales of the signal components; Before calculating the composite loss function, using a physical residual scaling matrix to perform dimensionless processing on the physical regularization term. 7.The multi-sensor fusion AI-based industrial equipment fault intelligent diagnosis method according to claim 1, characterized in that, The fault diagnosis model is a deep learning model based on a Transformer model or an autoencoder model, which is used to capture patterns in time series data. 8.The multi-sensor fusion AI-based industrial equipment fault intelligent diagnosis method according to claim 1, characterized in that, The setting steps of the fault alarm threshold include: Collecting the health state indicator corresponding to the healthy state of the equipment, and constructing an empirical cumulative distribution function; According to the preset acceptable false alarm rate, querying the inverse function of the empirical cumulative distribution function to determine the fault alarm threshold.

Images