An industrial equipment fault diagnosis model construction method and related equipment

Abstract

The application provides an industrial equipment fault diagnosis model construction method and related equipment, and relates to the technical field of industrial equipment fault diagnosis. The method comprises the following steps: obtaining a physical mechanism model of the industrial equipment; the physical mechanism model comprises a physical constraint equation; constructing a deep learning network; the deep learning network comprises a composite loss function constructed according to the physical constraint equation; collecting full-time series state data of the industrial equipment during operation; the full-time series state data comprises specified time series state data collected when the industrial equipment is in a healthy state; based on the full-time series state data, the deep learning network is trained for parameter optimization to obtain a final available industrial equipment fault diagnosis model. The method of the application aims to solve the problems of weak early fault identification ability, poor interpretability, lack of physical meaning and strong data dependence in the existing industrial equipment fault diagnosis technology, and effectively improves the intelligent level and operation and maintenance efficiency of the industrial equipment fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment fault diagnosis technology, and more specifically, to a method for constructing an industrial equipment fault diagnosis model and related equipment. Background Technology

[0002] In the current industrial production system, industrial equipment, especially core production equipment containing a large number of rotating mechanical parts, is the foundation for ensuring the stable operation of production activities. If a malfunction in this type of equipment is not detected in time, it can not only lead to production interruptions and huge economic losses, but may also cause serious safety accidents. Therefore, timely and accurate fault diagnosis of industrial equipment, especially the identification of early-stage faults, has become one of the core requirements in the field of industrial operation and maintenance.

[0003] Currently, fault diagnosis technologies for industrial equipment are mainly divided into two categories: The first category is methods based on physical mechanisms: These methods mainly rely on accurate physical models and expert experience for diagnosis, such as finite element analysis, fault tree analysis, and solving physical equations. These methods fully integrate the inherent physical laws governing the equipment's operation, and the resulting diagnostic results have clear physical meaning. However, their inherent drawbacks include the complexity of the modeling process, requiring significant time and specialized personnel to complete the modeling. Furthermore, they struggle to adapt to uncertainties such as performance degradation and changes in operating conditions that may occur during actual equipment operation. In the complex and ever-changing environments of real industrial sites, their application flexibility and adaptability are insufficient.

[0004] The second category is data-driven methods. These methods primarily rely on machine learning and deep learning models for diagnosis, such as support vector machines, random forests, and deep neural networks. These methods do not require complex manual mechanistic modeling; sufficient sample data is all that is needed for model training. While they demonstrate efficiency advantages in certain specific scenarios, they also have significant drawbacks. First, their models are often considered black boxes, with opaque decision-making processes and diagnostic results lacking clear physical evidence, leading to poor interpretability and difficulty in gaining the trust of industrial field maintenance personnel. This also hinders subsequent troubleshooting and repair work. Second, data-driven methods have low accuracy in identifying early, weak fault signals, such as the subtle vibrations caused by early bearing wear. Their sensitivity is insufficient, failing to identify anomalies in the early stages of fault development, missing the optimal repair window and allowing the fault to escalate and cause losses. Third, these methods heavily rely on large amounts of labeled fault data for training. However, in industrial settings, equipment operates normally most of the time, and labeled fault samples are often very scarce, making it difficult to obtain sufficient training data to support the model's ideal diagnostic performance.

[0005] In summary, existing industrial equipment fault diagnosis technologies have the following core shortcomings: First, the ability to identify early faults is weak. Traditional data-driven methods generally have low accuracy in identifying early, weak fault signals and cannot capture the subtle signal anomalies caused by early faults in a timely manner.

[0006] Second, it has poor interpretability. When making diagnostic decisions, the internal logic of deep learning models is not transparent, making it difficult for industrial field maintenance personnel to understand the basis of the diagnostic results. This reduces their trust in the diagnostic system and hinders subsequent fault handling.

[0007] Third, there is a lack of physical meaning. Purely data-driven models often ignore the inherent physical laws and mechanisms of equipment, which may lead to insufficient generalization ability of the model, making it difficult to adapt to different working conditions and operating conditions. When changing the working conditions or replacing the equipment with a similar one, the diagnostic performance may drop significantly.

[0008] Fourth, it is highly dependent on data. Both machine learning and deep learning models require a large amount of labeled fault data for training to achieve good diagnostic performance. However, fault samples from industrial sites are difficult to obtain and are usually scarce, making it difficult to meet the data requirements for model training.

[0009] There is currently no effective technical solution to the above problems. Summary of the Invention

[0010] The purpose of this invention is to provide a method and related equipment for constructing an industrial equipment fault diagnosis model, which aims to solve the problems of weak early fault identification capability, poor interpretability, lack of physical meaning and strong data dependence in existing industrial equipment fault diagnosis technologies, and effectively improve the intelligence level and operation and maintenance efficiency of industrial equipment fault diagnosis.

[0011] In a first aspect, the present invention provides a method for constructing an industrial equipment fault diagnosis model, used to construct an industrial equipment fault diagnosis model for diagnosing faults in mechanical rotating parts of industrial equipment. The method for constructing the industrial equipment fault diagnosis model includes the following steps: S1. Obtain the physical mechanism model of the industrial equipment; the physical mechanism model includes physical constraint equations; S2. Construct a deep learning network; the deep learning network includes a composite loss function constructed based on the physical constraint equations; S3. Collect full time-series status data of the industrial equipment during operation; the full time-series status data includes specified time-series status data collected when the industrial equipment is in a healthy state; S4. Based on the full time-series state data, perform parameter optimization training on the deep learning network to obtain the final usable industrial equipment fault diagnosis model; The parameter optimization training specifically includes the following steps A1-A2: A1. Using the specified time-series state data, a health baseline model is established by pre-training the deep learning network; A2. After establishing the health baseline model, the deep learning network is jointly trained using the full amount of time-series state data.

[0012] The industrial equipment fault diagnosis model construction method provided by this invention combines a physical mechanism model with a deep learning network to construct a fault diagnosis model with physical interpretability. Through a two-stage training strategy, it effectively solves the problem of scarce fault samples in industrial settings and improves the ability to identify early-stage faults. Simultaneously, the model can output a fault severity index and physical interpretability evidence, enhancing the transparency and credibility of the diagnostic results and overcoming the problems of weak early-stage fault identification ability, poor interpretability, lack of physical meaning, and strong data dependence in existing technologies.

[0013] Secondly, the present invention provides an industrial equipment fault diagnosis model construction device for constructing an industrial equipment fault diagnosis model, wherein the industrial equipment fault diagnosis model is used to diagnose faults in mechanical rotating parts of industrial equipment, and the industrial equipment fault diagnosis model construction device comprises: An acquisition module is used to acquire the physical mechanism model of industrial equipment; the physical mechanism model includes physical constraint equations. A building module for constructing a deep learning network; the deep learning network includes a composite loss function constructed based on the physical constraint equations; The data acquisition module is used to acquire full-time status data of the industrial equipment during operation; the full-time status data includes specified time-series status data acquired when the industrial equipment is in a healthy state; The training module is used to perform parameter optimization training on the deep learning network based on the full set of time-series state data to obtain a final usable industrial equipment fault diagnosis model. The parameter optimization training specifically includes the following steps A1-A2: A1. Using the specified time-series state data, a health baseline model is established by pre-training the deep learning network; A2. After establishing the health baseline model, the deep learning network is jointly trained using the full amount of time-series state data.

[0014] Thirdly, the present invention provides an electronic device including a processor and a memory, the memory storing computer-readable instructions, which, when executed by the processor, perform the steps in the industrial equipment fault diagnosis model construction method provided in the first aspect above.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, performs the steps of the industrial equipment fault diagnosis model construction method provided in the first aspect above.

[0016] As can be seen from the above, the industrial equipment fault diagnosis model construction method provided by the present invention effectively solves the core defects of the prior art, such as weak early fault identification capability, poor interpretability, lack of physical meaning, and strong data dependence, by deeply integrating the physical mechanism model of industrial equipment with deep learning network.

[0017] Specifically, this application acquires a physical mechanism model of industrial equipment and integrates its physical constraint equations (including vibration characteristic equations, vibration dynamics equations, and fault characteristic frequency calculation equations) into the composite loss function of a deep learning network to construct a physical mechanism constraint loss term. This combination method enables the model to learn data features while being forced to follow the inherent physical laws of the equipment, thereby significantly improving the model's ability to identify weak early fault signals and overcoming the accuracy bottleneck of traditional data-driven methods in handling early faults.

[0018] Furthermore, this application employs a two-stage training strategy. First, it pre-trains the physically constrained layers of the deep learning network using specified temporal state data under healthy conditions to establish a healthy baseline model. Then, it performs end-to-end joint training based on this baseline using full temporal state data. This strategy effectively alleviates the problem of scarce fault samples in industrial settings, enabling the model to maintain good diagnostic performance and generalization ability even with limited fault data.

[0019] More importantly, the industrial equipment fault diagnosis model of this application can not only extract potential feature vectors with physical mapping relationships from real-time time-series status data and calculate their deviation from the health baseline model to output fault diagnosis results and fault severity index, but also further map the potential feature vectors to different physical dimensions (dynamic equations, frequency domain features, energy conservation) and calculate the reconstruction residuals, thereby identifying the dominant physical factors causing the deviation. This design provides a clear physical interpretability basis for fault diagnosis results, enabling industrial field operation and maintenance personnel to understand the internal logic of diagnostic decisions, enhancing their trust in the diagnostic system, and providing precise guidance for subsequent fault handling work.

[0020] In summary, this application constructs a high-precision, highly interpretable, and low-data-dependency industrial equipment fault diagnosis model by deeply integrating physical mechanisms and data-driven methods, effectively improving the intelligence level and operation and maintenance efficiency of industrial equipment fault diagnosis.

[0021] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings. Attached Figure Description

[0022] Figure 1 This is a flowchart of a method for constructing an industrial equipment fault diagnosis model according to an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of a device for constructing an industrial equipment fault diagnosis model, as provided in an embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0025] Label Explanation: 100. Acquisition module; 200. Construction module; 300. Data acquisition module; 400. Training module; 13. Electronic device; 1301. Processor; 1302. Memory; 1303. Communication bus. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0027] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0028] Traditional industrial equipment fault diagnosis technologies commonly suffer from several problems: weak early fault identification capabilities, specifically insufficient sensitivity to weak early fault signals in rotating mechanical components leading to reduced diagnostic accuracy; poor interpretability stemming from the opacity of the decision-making process within data-driven models, resulting in diagnostic results lacking physical basis and failing to gain the trust of industrial field maintenance personnel; insufficient model generalization ability due to the lack of physical meaning, making it unable to adapt to changes in equipment operating conditions; and strong data dependence, heavily reliant on training with a large number of labeled fault samples, which are difficult and scarce to obtain in industrial settings. In particular, insufficient identification of weak early fault signals directly affects the timeliness of fault warnings; lack of interpretability weakens the credibility of diagnostic results; neglect of physical mechanisms limits the applicability of the model under different operating conditions; and data dependence restricts the efficiency of model deployment in actual industrial environments.

[0029] For example, in the scenario of bearing fault diagnosis for wind turbine generators, the vibration signals generated by the bearings in the early wear stage are weak in amplitude. Existing data-driven models struggle to effectively identify such signals, and the output results lack explanation of the physical process, making it impossible for maintenance personnel to verify the diagnostic basis. Simultaneously, physical mechanism models are difficult to accurately model due to uncertainties in operating conditions such as wind speed fluctuations and load changes, leading to unstable diagnostic results. Furthermore, because fault sample annotation requires on-site operation by professionals and is time-consuming, the available annotation data is extremely limited. The model training process is limited in performance due to insufficient samples, thus affecting the continuous and reliable operation of the equipment. Moreover, the failure to detect weak vibration signals in this scenario may lead to a continuous deterioration of the fault state. The lack of physical evidence makes maintenance decisions lack scientific support, and insufficient adaptability to operating conditions leads to false alarms or missed alarms. The scarcity of data further exacerbates the decline in the model's generalization ability.

[0030] If the above problems are not addressed, the fault diagnosis system will be unable to detect early fault states in a timely manner, leading to the accumulation of equipment operational risks and potentially causing unexpected downtime. Insufficient model generalization ability will cause diagnostic results to become invalid when operating conditions change, reducing the overall reliability of the system. Poor interpretability will weaken maintenance personnel's trust in the diagnostic system, affecting fault handling efficiency. Strong data dependency will limit the model's application in real-world industrial environments, increasing the cost of technology application. Therefore, the shortcomings of existing technologies pose significant obstacles to the accuracy, reliability, and practicality of industrial equipment fault diagnosis, urgently requiring a technical solution that can balance physical laws and data characteristics.

[0031] For this, please refer to Figure 1 , Figure 1This is a flowchart of a method for constructing an industrial equipment fault diagnosis model. The method proposed in this invention is used to construct an industrial equipment fault diagnosis model, which is used to diagnose faults in mechanical rotating parts (e.g., bearings) of industrial equipment. The method includes the following steps: S1. Obtain the physical mechanism model of the industrial equipment; the physical mechanism model includes physical constraint equations that can describe the dynamic behavior of the industrial equipment; the physical constraint equations include vibration characteristic equations for decomposing vibration signals of mechanical rotating parts, vibration dynamic equations for describing the vibration motion of mechanical rotating parts, and fault characteristic frequency calculation equations for describing fault characteristics of mechanical rotating parts. S2. Construct a deep learning network with feature extraction and signal reconstruction functions; the deep learning network includes a composite loss function constructed based on physical constraint equations; the composite loss function includes a data-driven loss term used to characterize the deviation between the predicted output and the measured data, a physical mechanism constraint loss term constructed based on physical constraint equations, and a regularization loss term; the physical mechanism constraint loss term includes a dynamic equation constraint term, a frequency domain feature constraint term, and an energy conservation constraint term; S3. Collect full time-series status data of industrial equipment during operation; the full time-series status data includes the specified time-series status data collected when the industrial equipment is in a healthy state; S4. A two-stage training strategy is adopted. Based on the full amount of time-series state data, the parameters of the deep learning network are optimized and trained to obtain the final usable industrial equipment fault diagnosis model. Parameter optimization training specifically includes the following steps A1-A2: A1. Using specified time-series state data, a healthy baseline pattern is established by pre-training the physically constrained layers in a deep learning network; A2. After establishing a health baseline model, the deep learning network is jointly trained end-to-end using full-scale time-series state data; The industrial equipment fault diagnosis model is used to extract potential feature vectors with physical mapping relationships from the input real-time time-series status data, calculate the degree of deviation of the potential feature vectors from the health status feature vectors corresponding to the health benchmark mode, and output fault diagnosis results and fault severity index based on the degree of deviation (according to the numerical range of the fault severity index, the fault severity is divided into healthy state, early fault, moderate fault and severe fault).

[0032] For ease of understanding, the following explains some key terms in this embodiment: Industrial Equipment Fault Diagnosis Model: This model aims to identify and predict potential equipment failures, particularly those affecting rotating mechanical components such as bearings, by analyzing industrial equipment operating data. Its core function is to extract meaningful features from real-time data, assess the degree of deviation of these features from the equipment's health status, and ultimately output diagnostic results and the severity of the fault.

[0033] Physical mechanism model: This model is a mathematical description of the inherent operating mechanism of industrial equipment. It includes a series of physical constraint equations that accurately describe the dynamic behavior of the equipment. By introducing a physical mechanism model, it can be ensured that the fault diagnosis process not only relies on data appearances but also delves into the physical essence of the fault.

[0034] Deep learning networks: These are neural network structures capable of feature extraction and signal reconstruction. They can automatically learn and extract high-level, abstract features from raw time-series data, and simultaneously reconstruct these features back into the original signal. This bidirectional capability allows the network to better understand the intrinsic structure and physical meaning of the data during the learning process.

[0035] Composite loss function: This function is an optimization objective used to guide the training of deep learning networks. It consists of multiple loss terms, including a data-driven loss term, a physical constraint loss term, and a regularization loss term. By comprehensively considering these loss terms, it can be ensured that the network fits the data while also following physical laws, and overfitting is avoided.

[0036] Full-volume time-series status data: This refers to all time-series data collected during the operation of industrial equipment. This data comprehensively records the operating status of the equipment under different operating conditions. It includes specific time-series status data collected when the equipment is in a healthy state; this data is crucial for establishing a healthy baseline model.

[0037] Health baseline model: This model is a reference pattern established by learning and modeling data from equipment operating in a healthy state. It represents the characteristic distribution and behavioral patterns of the equipment during normal operation and serves as the basis for determining whether the equipment has malfunctioned and the severity of the malfunction.

[0038] Latent Feature Vectors: These vectors are abstract feature representations with physical mapping relationships extracted by deep learning networks from real-time time-series state data. These feature vectors can capture key information about the device's operating status, and their dimensionality is usually lower than the original data, which helps simplify subsequent fault analysis.

[0039] Health Status Feature Vector: This vector corresponds to the health baseline pattern and represents the typical characteristics of the device in a healthy state. By comparing the deviation between the real-time extracted latent feature vector and the health status feature vector, the degree of abnormality in the device's operating state can be quantified.

[0040] Fault Severity Index: This index is a quantitative indicator calculated based on the degree of deviation between the potential feature vector and the health status feature vector. It can intuitively reflect the severity of equipment faults and can be used to classify faults into different levels, such as healthy status, early fault, moderate fault, and severe fault.

[0041] This application proposes a method for constructing an industrial equipment fault diagnosis model. This method is used to construct an industrial equipment fault diagnosis model, which is used to diagnose faults in mechanical rotating parts (such as bearings) in industrial equipment.

[0042] When constructing a fault diagnosis model for industrial equipment, the first step is to obtain a physical mechanism model of the equipment. This physical mechanism model includes physical constraint equations that describe the dynamic behavior of the industrial equipment. These physical constraint equations are mathematical expressions of the inherent operating laws of the equipment. For example, dynamic equations derived from Newton's second law or Lagrange's equations can be used to describe the vibration motion of rotating mechanical components. Furthermore, physical equations describing vibration signal decomposition and fault characteristic frequency calculation can be established by analyzing the equipment's structure, material properties, and working principles. For example, for bearing fault diagnosis, calculation equations describing the fault characteristic frequencies of the bearing's inner ring, outer ring, or rolling elements can be established. Obtaining these physical constraint equations provides a solid physical foundation for the subsequent construction and training of deep learning networks, avoiding the potential problem of missing physical meaning in purely data-driven models.

[0043] Furthermore, a deep learning network with feature extraction and signal reconstruction capabilities is constructed. This deep learning network can adopt various architectures, such as an encoder-decoder structure, where the encoder is responsible for extracting latent features from the original time-series state data, and the decoder is responsible for reconstructing the latent features back to the original signal. This deep learning network includes a composite loss function constructed based on physical constraint equations. This composite loss function is the core guiding the network training, integrating multiple optimization objectives. For example, mean squared error (MSE) can be used as a data-driven loss term to minimize the deviation between the predicted output and the measured data. The physical mechanism constraint loss term can be constructed based on the aforementioned physical constraint equations. For example, it can include a kinetic equation constraint term to ensure that the features learned by the network conform to the dynamic laws of the device; a frequency domain feature constraint term to ensure that the network captures fault-related feature frequencies in the frequency domain; and an energy conservation constraint term to ensure that the energy of the network-reconstructed signal is consistent with the theoretical energy. In addition, regularization loss terms, such as L1 or L2 regularization, can be introduced to prevent overfitting and improve the model's generalization ability.

[0044] After constructing the deep learning network, it is necessary to collect full-scale time-series state data of the industrial equipment during operation. This data forms the basis for training and validating the model. Full-scale time-series state data can include various sensor data such as vibration, temperature, and current under different operating conditions and loads. Specifically, full-scale time-series state data includes specified time-series state data collected when the industrial equipment is in a healthy state. This health state data is crucial for establishing a baseline pattern of health, as it represents the behavioral pattern of the equipment during normal operation. For example, operational data can be continuously collected for a period of time during the initial installation of the equipment or after maintenance confirms its health as health state data. This data can be used to pre-train the model, enabling it to learn the characteristic representation of health.

[0045] Subsequently, a two-stage training strategy was adopted. Based on the full range of time-series state data, the deep learning network was trained with optimized parameters to obtain a usable industrial equipment fault diagnosis model. This two-stage training strategy aims to fully utilize health state data and the full range of data to improve the training efficiency and diagnostic performance of the model. In the first stage, a health baseline pattern was established by pre-training the physically constrained layers in the deep learning network using specified time-series state data. For example, the encoder and decoder can be trained using only health state data, enabling them to accurately reconstruct health signals and learn the latent feature representations of health. In the second stage, after establishing the health baseline pattern, the deep learning network was jointly trained end-to-end using the full range of time-series state data. At this stage, the data-driven loss term, the physical mechanism constraint loss term, and the regularization loss term can be optimized simultaneously, allowing the network to maintain physical consistency while further improving feature extraction and fault diagnosis capabilities.

[0046] Ultimately, the industrial equipment fault diagnosis model extracts latent feature vectors with physical mapping relationships from the input real-time time-series state data, calculates the deviation of these latent feature vectors from the health state feature vectors corresponding to the health benchmark model, and outputs fault diagnosis results and a fault severity index based on the deviation. For example, when real-time time-series state data is input into a trained deep learning network, the encoder extracts the corresponding latent feature vectors. Then, these latent feature vectors are compared with the health state feature vectors corresponding to the pre-established health benchmark model, calculating the distance or similarity between them to quantify the degree of deviation. Based on this deviation, fault diagnosis results can be output, such as determining whether the equipment has malfunctioned, and classifying the fault severity into healthy state, early fault, moderate fault, and severe fault according to the numerical range of the deviation. This deviation-based diagnosis method enables the model to have higher sensitivity to early, weak fault signals, improving the ability to identify early faults.

[0047] The following example will provide a more detailed explanation of the above technical solution: Suppose we are on an industrial production line and need to diagnose a critical pump, particularly its internal bearings. Traditional diagnostic methods may suffer from weak early fault identification capabilities, poor interpretability, and heavy reliance on a large number of labeled fault samples.

[0048] To address this issue, this application proposes a method for constructing a fault diagnosis model for industrial equipment. First, for the bearing of the pump, its physical mechanism model is obtained. This includes establishing vibration characteristic equations describing the decomposition of bearing vibration signals. Furthermore, it is necessary to establish fault characteristic frequency calculation equations describing the bearing fault characteristics, such as calculation formulas for the fault frequencies of the bearing's inner ring, outer ring, and rolling elements. The introduction of these physical equations allows the model to follow the actual physical laws of the bearing during subsequent learning, avoiding the "black box" problem that may occur with purely data-driven models.

[0049] Next, a deep learning network with feature extraction and signal reconstruction capabilities is constructed. This network can be an autoencoder structure, where the encoder part is responsible for extracting latent features from the vibration signal of the pump bearing, and the decoder part is responsible for reconstructing these latent features back to the original vibration signal. The training of this network is guided by a composite loss function. This composite loss function includes a data-driven loss term, such as mean squared error, to ensure that the signal reconstructed by the network is as close as possible to the actual measured signal. It also includes a physical mechanism constraint loss term based on the aforementioned physical constraint equations. For example, the dynamic equation constraint term penalizes cases where the network-reconstructed signal does not conform to the bearing dynamic equation; the frequency domain feature constraint term ensures that the network captures the correct amplitude changes at the fault characteristic frequencies; and the energy conservation constraint term ensures that the total energy of the reconstructed signal remains consistent with the theoretical energy. Furthermore, a regularization loss term is added to prevent the network from overfitting.

[0050] Before model training, it is necessary to collect full-series time-series status data of the pump during operation. For example, during the initial installation of the pump or after confirming its healthy status, vibration sensor data should be continuously collected. This data serves as the specified time-series status data to establish a healthy baseline pattern. Simultaneously, full-series status data should be continuously collected under various operating conditions, including normal pump operation and potential malfunctions.

[0051] Subsequently, a two-stage training strategy was employed to optimize the parameters of the deep learning network. In the first stage, the layers of the deep learning network related to physical constraints were pre-trained using specified temporal state data under healthy conditions. For example, the autoencoder could be trained using only healthy data, enabling it to accurately reconstruct vibration signals under healthy conditions and learn the latent feature representations of health. The goal of this stage was to establish an accurate health baseline pattern. In the second stage, after establishing the health baseline pattern, the deep learning network was jointly trained end-to-end using all temporal state data. At this stage, data-driven loss, physical mechanism constraint loss, and regularization loss worked together, enabling the network to learn data features while strictly adhering to physical laws and improving its generalization ability to various operating states.

[0052] Ultimately, during actual operation, the pump's real-time status data is input into the trained industrial equipment fault diagnosis model. The model extracts latent feature vectors with physical mapping relationships from this data. For example, these latent feature vectors might correspond to physical quantities such as bearing vibration energy or the amplitude of a specific fault frequency. Then, the model calculates the deviation of these real-time extracted latent feature vectors from the health status feature vectors corresponding to the health baseline model. For example, the deviation is quantified by calculating the Euclidean distance between the two vectors. Based on this deviation, the model outputs a fault diagnosis result, such as "early fault" or "moderate fault," and provides a fault severity index. For example, if the deviation is small, the fault severity index might be 15%, indicating "early fault (planned maintenance)"; if the deviation is large, the index might be 80%, indicating "serious fault (immediate shutdown for maintenance)." This method allows the model to be more sensitive to early, weak fault signals, and because the latent feature vectors have physical mapping relationships, the diagnostic results also have better interpretability.

[0053] The proposed method for constructing industrial equipment fault diagnosis models effectively addresses many problems in existing technologies by deeply integrating the physical mechanism model of industrial equipment with a deep learning network. For example, in the aforementioned pump bearing fault diagnosis example, traditional data-driven methods may struggle to identify the weak vibration signals generated by early bearing wear because these signals lack distinct features and there is a lack of labeled early fault samples for training. However, this application introduces physical constraint equations, guiding the deep learning network to learn features that conform to physical laws during training, enabling the establishment of accurate health baseline patterns even in situations of scarce data. When early wear occurs in the pump bearing, its vibration signals exhibit slight deviations. Although these deviations are not obvious in the original data, under the guidance of physical constraints, the latent feature vectors extracted by the model can more sensitively capture these deviations, thereby achieving early fault identification.

[0054] Furthermore, traditional deep learning models often present diagnostic results as a "black box," making it difficult for maintenance personnel to understand the rationale behind their decisions. This application addresses this by imbuing latent feature vectors with physical mapping relationships. For example, a certain feature dimension might directly correspond to the vibration energy of a bearing or the amplitude of a specific fault frequency, thus giving the diagnostic results physical interpretability. When the model outputs "early fault" and indicates a significant deviation in a certain latent feature dimension, maintenance personnel can infer from the physical meaning of that dimension that a certain physical parameter of the bearing may have malfunctioned, thereby increasing their confidence in the diagnostic results and providing clear guidance for subsequent maintenance decisions.

[0055] Compared to purely data-driven methods that heavily rely on a large number of labeled fault samples, this application employs a two-stage training strategy. Specifically, the first stage utilizes health status data to pre-train the physically constrained related layers, enabling the model to be built with few or even no labeled fault samples. This significantly lowers the barrier to entry for industrial applications, as obtaining a large number of labeled fault samples is typically very difficult in real-world industrial environments. By establishing accurate health baseline patterns, the model can identify faults by detecting deviations from the health status without pre-learning the characteristics of various fault modes, thereby improving the model's generalization ability and enabling it to adapt to the equipment fault diagnosis needs under different operating conditions.

[0056] In some embodiments, the specific expression of the vibration characteristic equation is as follows: ; in, Let be the vibration signal of the rotating mechanical component at time t. The normal vibration signal of the rotating mechanical parts of the industrial equipment at time t when it is in a healthy state. A historical data-based learning approach can be adopted. During the initial installation phase or the health assessment stage, a large amount of vibration data from normal operation can be collected as reference data for normal vibration under healthy conditions. This data is then used to train a baseline Gaussian mixture model to describe the modes and range of healthy vibrations. For new vibration signals, the output signal or reconstructed signal from this baseline Gaussian mixture model can be considered as a response to... (estimates) The abnormal vibration signal caused by the failure of the mechanical rotating component at time t is (for abnormal vibration caused by failure, the sensor can directly measure the deviation from the healthy benchmark in the total vibration signal and the part related to the failure mechanism, and then the signal can be separated by the method of spectrum feature extraction: for example, for periodic failures (such as bearing ball, raceway damage, fatigue damage, etc.). Peaks will occur at specific fault characteristic frequencies (such as inner and outer race frequencies) and their harmonics. This is achieved by analyzing the total signal. By performing Fast Fourier Transform (FFT) spectral analysis and focusing on the energy in these characteristic frequency regions, it is possible to indirectly quantify and separate the energy. The spectral components; or obtained by signal separation using residual analysis to reduce computational complexity: for example, once obtained from the above methods This can be obtained from the measured total signal. Subtracting this from the original signal yields the residual signal. ,However Theoretically included and Therefore, further signal processing (such as filtering and envelope demodulation) is needed to estimate and suppress noise. This allows for a purer extraction. ), Let be the measurement noise at time t.

[0057] Specifically, It is a direct reflection of the equipment's operating status, including all information such as health, faults, and noise (i.e., total vibration signal). It can be obtained directly by acceleration sensors, velocity sensors or displacement sensors installed on mechanical rotating parts (such as bearing housings, motor housings, etc.), or by preprocessing the raw sensor data through preliminary filtering, sampling and other methods. It characterizes the inherent vibration modes of the equipment under fault-free conditions. The data can be obtained using a learning approach based on historical data. For example, during the initial installation phase or the health confirmation phase, a large amount of vibration data from normal operation can be collected as reference data for normal vibration under healthy conditions. This data can then be used to train a baseline Gaussian mixture model (GMM) to describe the modes and ranges of healthy vibrations. For new vibration signals, the output signal or reconstructed signal of this baseline Gaussian mixture model can be considered as a representation of the vibration patterns. The estimation can be performed. Alternatively, an autoencoder model can be trained to learn and reconstruct normal vibration patterns on health data; its reconstructed output can also be used as... The estimate. It is direct evidence of the existence of a fault, usually manifested as a specific frequency component, impulse, or modulation phenomenon that is different from the healthy state. The information can be obtained by using spectral feature extraction methods, for example, for periodic faults (such as bearing ball, raceway damage, fatigue damage, etc.). Peaks will be generated at specific fault characteristic frequencies (such as inner and outer ring frequencies) and their harmonics, which will be observed in the total signal. By performing Fast Fourier Transform (FFT) spectral analysis and focusing on the energy in these characteristic frequency regions, it is possible to indirectly quantify and separate the energy. The spectral components. Alternatively, residual analysis can be used, once the above methods are obtained. The estimate can be obtained from the measured total signal. Subtracting this from the original signal yields the residual signal. The residual signal theoretically contains and Noise can be estimated and suppressed through further signal processing (such as filtering and envelope demodulation). This allows for a purer extraction. . This represents the measurement noise at time t. This noise is typically random and non-periodic, and may originate from the sensor itself, environmental interference, or the data acquisition system. The identification and suppression of noise can be modeled using statistical noise models (e.g., assuming Gaussian white noise) or separated and removed from the signal using adaptive filtering techniques (e.g., Kalman filtering).

[0058] The solution presented in this application provides a clear physical mechanism basis for the construction method of industrial equipment fault diagnosis models through a well-defined vibration characteristic equation. This equation incorporates complex vibration signals. Decomposed into healthy vibration signals Abnormal vibration signals caused by faults and measuring noise Three components with clear physical meaning. This decomposition method allows for a more precise integration of physical constraint equations (including vibration dynamics equations and fault characteristic frequency calculation equations) into the composite loss function when constructing deep learning networks. Specifically, by... The estimation allows the model to establish a healthy baseline pattern, providing a reliable reference for subsequent calculations of the deviation of latent feature vectors from the healthy baseline pattern. Through the estimation of... Effective separation and extraction of fault signals, even weak early-stage fault signals, can be amplified and identified, thus significantly improving the model's ability to identify early-stage faults. Simultaneously, for The clear definition and processing of these physical mechanisms help suppress noise interference during training, ensuring that the model learns effective vibration features rather than noise. This deep integration of physical mechanisms and data-driven approaches makes deep learning networks no longer "black boxes," giving their internal learning processes and diagnostic decisions physical interpretability, thus solving the problems of lack of physical meaning and poor interpretability in traditional data-driven models. Furthermore, It can learn from a large amount of readily available health status data, effectively alleviating the problem of scarce fault samples in industrial fields and reducing the model's dependence on a large amount of labeled fault data.

[0059] The following is a concrete example. Suppose we are diagnosing a fault in the bearing of an industrial fan. First, we collect vibration signals from the fan during operation using an accelerometer mounted on the bearing housing. In order to obtain During the initial installation of the wind turbine or after a comprehensive overhaul confirming its health, it is operated continuously for one week to collect a large amount of vibration data. This health data is used to train a Gaussian mixture model, which can learn and characterize the amplitude distribution, frequency characteristics, and other patterns of bearing vibration under healthy conditions. When new vibration signals appear... During input, the trained Gaussian mixture model is used to... Reconstruction is performed, and the resulting reconstructed signal is used as... The estimate. For Residual analysis can be used. The estimated values ​​can be... From the original signal Subtracting from the middle yields the residual signal. .because It may still contain some environmental noise or non-fault fluctuations during normal equipment operation, which can be addressed. Envelope demodulation is performed to highlight the impact characteristics caused by the fault, and bandpass filtering is combined to remove high-frequency random noise, thereby extracting the signal more purely. . This can be viewed as a random, unstructured component that still exists in the residual signal after the above processing, or it can be characterized by establishing a noise model through statistical analysis of historical noise data. In this way, the original complex vibration signal is clearly decomposed into three physically meaningful components, providing accurate input for the subsequent training of deep learning networks and the application of physical constraints.

[0060] Through the above technical solution, this application provides a clear method for decomposing vibration signals, offering a clear and calculable physical basis for the physical mechanism constraints in the construction of industrial equipment fault diagnosis models, effectively solving the problem of physical constraints being impossible to implement due to the lack of clear decomposition rules. This decomposition expression splits the total vibration signal into three signal components from different sources, conforming to the actual physical process and clearly distinguishing healthy vibration components, faulty vibration components, and interference noise, providing a solid foundation for subsequent extraction of fault features and calculation of physical constraint losses. A Gaussian mixture model is trained using a large amount of readily available unlabeled normal data from equipment installation or the healthy phase to estimate... This approach adapts to the scarcity of fault labeling samples in industrial settings, allowing reliable healthy vibration references to be obtained without relying on a large number of fault labeling samples. It also provides a foundation for subsequent calculations of the deviation of fault signals from the healthy baseline. (For...) The abnormal vibration signal caused by the fault provides two different separation and extraction methods: spectral feature extraction and residual analysis. These methods can adapt to the computational needs of different scenarios. The residual analysis method also reduces computational complexity, making it suitable for real-time processing in industrial settings. The separated, pure fault signal can amplify the characteristics of early, weak faults, significantly improving the ability to identify early faults. To measure noise, noise is separated from the effective signal, which facilitates targeted noise suppression during subsequent processing, prevents noise from masking early weak fault signals, and further improves the model's ability to identify early faults.

[0061] In some embodiments, the specific expression of the vibration dynamics equation is as follows: ; in, For equivalent quality, for The second derivative, The damping coefficient is... for The first derivative, This is the stiffness coefficient. For the normal excitation force at time t, The additional excitation force caused by the fault at time t (which can be obtained by installing a force sensor or attaching a stress strain gauge, and is used to measure the magnitude of the unbalanced force of the additional excitation source caused by the fault in the rotational motion of the mechanical rotating parts).

[0062] Specifically, equivalent mass refers to the equivalent value of the mass of all moving parts of a complex mechanical system when simplified to a single-degree-of-freedom vibration system, converted to the vibration direction. This parameter can be obtained through theoretical calculations, finite element analysis, or experimental modal analysis. For example, for a bearing-rotor system, the equivalent mass can be approximated as the mass of the rotor plus a portion of the bearing housing's mass. The damping coefficient characterizes the system's ability to dissipate energy during vibration, such as friction, air resistance, and material internal friction. The damping coefficient can be determined through free decay vibration experiments, the half-power bandwidth method, or numerical simulation. The stiffness coefficient reflects the system's ability to resist deformation, i.e., the force required per unit deformation. The stiffness coefficient can be determined through calculations using mechanics of materials formulas, finite element analysis, or static loading experiments. For example, for a bearing system, its stiffness mainly comes from the contact stiffness between the rolling elements and the inner and outer rings of the bearing, as well as the structural stiffness of the bearing housing. Vibration signals are typically measured in real time using accelerometers, velocity sensors, or displacement sensors. Normal excitation force refers to the periodic or random excitation caused by normal operating loads, unbalanced forces, gear meshing forces, etc., under healthy operating conditions. This force can be measured through theoretical calculations, historical data analysis, or by installing a force sensor under healthy conditions. The additional excitation force caused by a fault refers to the extra, abnormal excitation force generated when a rotating mechanical component malfunctions, such as bearing wear, cracks, or misalignment. This additional excitation force can be measured by installing a force sensor or attaching a strain gauge, for example, by installing a force sensor or attaching a strain gauge to the pump bearing housing or housing frame, to capture the magnitude of the unbalanced force generated by the additional excitation source due to the fault during the rotational motion of the mechanical device.

[0063] This application's solution provides a solid physical foundation for the aforementioned industrial equipment fault diagnosis model construction method by introducing a clear vibration dynamics equation. This equation, based on the classic single-degree-of-freedom forced vibration model, accurately describes the actual vibration characteristics of rotating mechanical components while avoiding overly complex calculations, thus adapting to the training and inference needs of deep learning models. By incorporating the three inherent structural parameters of the mechanical vibration system—equivalent mass, damping coefficient, and stiffness coefficient—into the equation, the model can closely match the actual structural properties of the target component, ensuring the accuracy of physical constraints and thus adapting to rotating mechanical components of different specifications and types. Furthermore, in the equation… and its first derivative and second derivative This approach comprehensively covers the dynamic changes of vibration in the time domain, enabling a complete description of the vibrational motion patterns of rotating mechanical components. By dividing the excitation into normal excitation forces and additional excitation forces caused by faults, it clearly reflects the actual impact of faults on the vibration system, facilitating the extraction of fault-related features by the subsequent model. This deep integration of physical mechanisms and data-driven approaches allows the deep learning network to learn features consistent with the actual vibration patterns of rotating mechanical components during training, effectively improving the model's generalization ability and interpretability, and enabling earlier identification of weak early fault signals. Combined with the aforementioned vibration characteristic equations, the vibration dynamics equations provided by this scheme offer more comprehensive and accurate physical constraints for the deep learning network, allowing the model to not only learn patterns from data but also understand the physical mechanisms behind these patterns, thus maintaining robustness even when data is scarce or operating conditions change.

[0064] As a specific implementation, the vibration dynamics equations can be applied to bearing fault diagnosis in an industrial pump. In this scenario, the equivalent mass can be defined as the equivalent mass of the pump rotor and the inner ring of the bearing, the value of which can be calculated from the pump's design parameters and material density. The damping coefficient can be estimated by conducting a free vibration experiment on the pump under healthy conditions and measuring its vibration decay curve. The stiffness coefficient can be calculated based on the bearing type, material properties, and structural stiffness of the bearing housing, or obtained through finite element analysis. Vibration signals and their derivatives can be acquired in real time by a triaxial accelerometer mounted on the bearing housing and obtained through digital differentiation. The normal excitation force can be obtained by measuring the radial and axial forces on the bearing using a force sensor under normal pump operation. The additional excitation force caused by the fault can be calculated by attaching stress-strain gauges to the bearing housing or pump casing to monitor abnormal stress changes generated when a fault occurs. The acquisition of these parameters and signals provides specific inputs for the deep learning network to construct the physical mechanism constraint loss term, enabling the model to learn and optimize based on real physical laws.

[0065] The above technical solution provides a clear vibration dynamics equation that conforms to the classical laws of mechanical vibration for deep learning networks, thus providing an accurate and practical physical basis for deep learning fault diagnosis models constrained by physical mechanisms. This effectively achieves a deep integration of physical mechanism knowledge and data-driven methods, and can specifically address the problems of poor model interpretability, weak early fault identification ability, insufficient generalization ability, and reliance on a large number of labeled fault samples in existing fault diagnosis technologies. The equation form can accurately describe the generation and variation of vibration without introducing overly complex calculations, adapting to the training and inference needs of deep learning models. At the same time, by dividing the excitation into normal excitation force and additional excitation force caused by fault, the actual impact of fault on the vibration system is clearly reflected, which facilitates the model to extract fault-related features, thereby helping the model learn physically meaningful features, effectively improving the model's generalization ability and interpretability, and enabling the model to identify weak early fault signals earlier.

[0066] In some embodiments, when the mechanical rotating component is a bearing, the fault characteristic frequency calculation equation includes the calculation equation for the bearing inner ring fault frequency, the calculation equation for the bearing outer ring fault frequency, and the calculation equation for the bearing rolling element fault frequency. The specific equation for calculating the failure frequency of the bearing inner ring is as follows: ; in, This refers to the failure frequency of the bearing inner ring. This refers to the number of rolling elements in the bearing. The diameter of the rolling element, For bearing pitch diameter, Contact angle, The rotational frequency of the axis; The specific equation for calculating the failure frequency of the bearing outer ring is as follows: ; in, This refers to the failure frequency of the bearing outer ring. The specific equation for calculating the failure frequency of bearing rolling elements is as follows: ; in, This represents the failure frequency of the bearing rolling elements.

[0067] When bearings are the rotating mechanical components, they are widely used support and transmission components in industrial equipment, and their operating status directly affects the overall performance and reliability of the equipment. Bearing fault diagnosis is a crucial aspect of industrial equipment maintenance and has significant practical application value. The calculation equations for fault characteristic frequencies include those for the inner ring fault frequency, the outer ring fault frequency, and the rolling element fault frequency. Bearing fault modes are diverse, mainly concentrated in key components such as the inner ring, outer ring, and rolling elements. Faults in different locations will exhibit specific frequency components in the vibration signal. Refining the calculation of fault characteristic frequencies into independent equations for these main fault locations allows for more accurate capture and quantification of the physical characteristics of bearing faults, providing a refined physical basis for subsequent frequency domain characteristic constraints, thus enabling the model to effectively distinguish between different types of bearing faults. The calculation equation for the inner ring fault frequency is used to quantify the characteristic vibration frequency generated when the inner ring of the bearing fails. Its calculation depends on the bearing's geometric parameters (e.g., number of rolling elements, rolling element diameter, bearing pitch diameter, contact angle) and the equipment's operating parameters (e.g., shaft rotation frequency). These parameters can typically be obtained by consulting the bearing manufacturer's technical manuals, equipment design drawings, or through on-site measurements (such as using calipers or tachometers), thus providing an accurate physical frequency reference for diagnosing inner ring faults. The calculation equation for the bearing outer ring fault frequency is similar to that for the inner ring fault frequency; this equation aims to determine the characteristic frequency corresponding to the bearing outer ring fault. Its input parameters also include the bearing's geometric parameters and the shaft's rotational frequency. The acquisition of these parameters is similar to that for the inner ring fault frequency calculation, ensuring the accuracy and operability of the calculation results and providing precise location for the frequency domain analysis of outer ring faults. The calculation equation for the bearing rolling element fault frequency is used to calculate the characteristic frequency generated when the bearing rolling element fails. Its calculation principle is based on the kinematic characteristics of the rolling elements, combined with the bearing's geometric parameters and the shaft's rotational frequency. Through this equation, specific frequency information of rolling element faults can be obtained, providing a clear physical basis for the model to identify rolling element faults in the frequency domain.

[0068] This application, in constructing an industrial equipment fault diagnosis model, first obtains the physical mechanism model of the industrial equipment, which includes physical constraint equations that describe the dynamic behavior of the industrial equipment. These physical constraint equations are further refined into vibration characteristic equations for decomposing vibration signals of rotating mechanical components, vibration dynamic equations for describing the vibration motion of rotating mechanical components, and fault characteristic frequency calculation equations for describing the fault characteristics of rotating mechanical components. Based on this, for bearings, a common rotating mechanical component in the industrial field, this application further clarifies the specific implementation method of the fault characteristic frequency calculation equations. By providing calculation equations for the fault frequencies of the bearing inner ring, outer ring, and rolling elements, this application concretizes the abstract "fault characteristic frequency calculation equations" into a quantifiable physical model. These specific calculation equations, utilizing the inherent structural parameters of the bearing (such as the number of rolling elements, rolling element diameter, bearing pitch diameter, and contact angle) and operating parameters (such as shaft rotation frequency), can accurately calculate the characteristic frequencies corresponding to faults in different parts of the bearing. These accurately calculated fault characteristic frequencies will serve as the quantitative basis for the frequency domain characteristic constraint term in the physical mechanism constraint loss term, an important component of the composite loss function in the deep learning network. Specifically, in the frequency domain feature constraint term, the deep learning network is guided to learn and identify frequency domain features related to bearing faults by comparing the difference between the amplitude of the Fourier transform function of the reconstructed signal at these specific fault frequencies and the preset physical constraints. This method, which deeply integrates the physical mechanism of bearing faults into the training process of the deep learning model, enables the model to not only learn patterns from data but also follow the inherent physical laws of the equipment. This helps the model to identify fault features more sensitively and accurately when faced with early, weak fault signals, overcoming the shortcomings of purely data-driven models in early fault identification. Simultaneously, because the model is constrained by physical laws during training, the features and diagnostic results it learns are more physically meaningful, thereby improving the model's generalization ability and interpretability, enabling industrial field maintenance personnel to better understand the basis of diagnostic results.

[0069] The following is a concrete example to illustrate this. Assume that the rotating mechanical component in a certain industrial equipment is a deep groove ball bearing, model 6205, with known structural parameters: 9 rolling elements, rolling element diameter of 7.94 mm, bearing pitch diameter of 39.5 mm, and contact angle of 0 degrees (for deep groove ball bearings, this is usually simplified to 0 degrees). The shaft rotation frequency of this bearing during operation is 1500 rpm, or 25 Hz. Based on the calculation equations provided in this application, the characteristic frequencies of the bearing when failure occurs in different parts can be calculated.

[0070] Calculation of bearing inner ring failure frequency: =9 / 2*(1+7.94 / 39.5*cos(0))*25≈4.5*(1+0.201)*25≈135.11Hz.

[0071] Calculation of bearing outer ring failure frequency: =9 / 2*(1-7.94 / 39.5*cos(0))*25≈4.5*(1-0.201)*25≈89.89Hz.

[0072] Calculation of bearing rolling element failure frequency: =39.5 / (2*7.94)*[1-(7.94 / 39.5*cos(0))^2]*25≈2.487*[1-(0.201)^2]*25≈59.66Hz.

[0073] The specific frequency values ​​obtained from these calculations (such as 135.11Hz, 89.89Hz, and 59.66Hz) will be used to construct the frequency domain feature constraint term in the composite loss function of the deep learning network. For example, when the vibration signal reconstructed by the model shows a significant amplitude near 135.11Hz, combined with other physical constraints, the model may be more inclined to diagnose it as a bearing inner race fault. In this way, the model can utilize this precise physical prior knowledge to optimize parameters during training, thereby improving the accuracy and reliability of identifying faults in different parts of the bearing.

[0074] Through the above technical solutions, this application clarifies the specific calculation methods for the characteristic frequencies corresponding to faults in different parts of bearings, which are the most common and prone to failure in industrial rotating mechanical components. This provides a feasible quantitative basis for the physical mechanism constraints in the entire fault diagnosis model construction process, enabling deep learning fault diagnosis based on physical constraints to accurately adapt to the diagnostic needs of bearing faults and significantly improve the physical consistency and diagnostic reliability of the model. Dividing the bearing fault characteristic frequency calculation equations into three different categories—inner ring faults, outer ring faults, and rolling element faults—can match the physical characteristics of faults in different parts, providing physical support for subsequently distinguishing different types of bearing faults. Specifically, by providing the calculation equation for the bearing inner ring fault frequency, which uses the directly obtainable structural and operating parameters of the bearing itself as calculation inputs, it conforms to the vibration generation mechanism of the bearing inner ring fault. This provides an accurate target frequency position for the amplitude verification of the inner ring fault in the subsequent frequency domain feature constraints, helping the model accurately capture the frequency domain characteristics corresponding to the inner ring fault. Similarly, the calculation equations for the bearing outer ring fault frequency and the bearing rolling element fault frequency are also based on readily available bearing parameters, conforming to the vibration characteristics of their respective faults. These equations accurately output the corresponding characteristic frequencies, providing precise locational basis for frequency domain physical constraints and helping the model identify outer ring and rolling element faults. These three types of calculation equations cover the three main fault locations in bearings, meeting common needs for bearing fault diagnosis in industrial scenarios. They provide a clear characteristic frequency range for the frequency domain characteristic constraint term in the composite loss function, enabling physical mechanism constraints to be truly quantified. This solves the previous problem of only specifying the purpose of the fault characteristic frequency calculation equations without concrete implementation methods. Ultimately, this allows the entire fault diagnosis model integrating physical mechanisms to be truly applied to bearing fault diagnosis, while ensuring that the physical meaning of the model conforms to the actual equipment characteristics, improving the model's interpretability and generalization ability, especially providing more accurate and reliable diagnostic results when facing early, weak fault signals.

[0075] In some embodiments, the specific expression of the composite loss function is as follows: ; in, This is the composite loss function (i.e., the total loss). For data-driven loss terms, and All are the first preset weights. For the physical mechanism constraint loss term, This is the regularization loss term; The specific expression for the data-driven loss term is: ; in, The total number of the first training samples, To preset reconstruction coefficients, for The mean square error (used to enable deep learning networks to learn normal vibration patterns in a healthy state by minimizing the mean square error between the predicted output and the measured data). For the i-th first training sample, the actual vibration signal is... The vibration signal reconstructed from the i-th first training sample. The weighted normalization coefficients for classification loss, For about The cross-entropy loss function (used to improve the ability of deep learning networks to distinguish different fault types by minimizing the cross-entropy between the predicted fault probability and the actual fault label). The true fault label for the i-th first training sample ( Represents the fault category, for example (like (Indicates a Type 1 fault) (Total number of fault categories (finite state machine, including healthy states)) The predicted fault probability of the i-th first training sample (denoted as the i-th first training sample belonging to...) (Probability of class of failures).

[0076] The specific expression for the physical mechanism constraint loss term is as follows: ; in, This represents the total number of the second training samples. , and All are the second preset weights. For the dynamic equation constraint term of the j-th second training sample, For the frequency domain feature constraint term of the j-th second training sample, For the j-th second training sample, there is an energy conservation constraint term. The specific expressions for the constraint terms in the dynamic equations are as follows: ; in, Indicated as to Find the L2 norm. for The second derivative, for The first derivative, Let be the reconstructed signal (estimate) of the mechanical rotating component of the j-th second training sample at time t. Let be the estimated excitation force at time t; The specific expression for the frequency domain characteristic constraint term is: ; in, The set of feature frequencies for a specified fault class (Fault class), whose symbol corresponds to C: the state classification result of the total number of fault categories. To indicate as to Modulus, For about The Fourier transform function of the reconstructed signal of the j-th second training sample (the reconstructed signal of the vibration signal in the vibration dynamics equation is in the time domain, and needs to be FFT (Fast Fourier Transform) to the frequency domain for further signal separation and residual calculation; in fault diagnosis, the vibration signal is usually analyzed in the frequency domain, however, the reconstructed signal (i.e., the decoder output): Because the characteristics are in the time domain, a Fourier transform is needed to convert the time-domain signal to the frequency domain, obtaining a complex spectrum. Furthermore, faults often produce peaks at specific frequencies. By performing a Fourier transform on the reconstructed signal based on the vibration dynamics equations, its reconstructed spectral characteristics are obtained, i.e. It is a complex function whose amplitude represents the energy of the signal at different frequencies. for The p-th characteristic frequency (fault frequency) in the equation. for The physical constraint corresponding to the amplitude of the p-th characteristic frequency in the model (can be obtained through experimental measurements and threshold settings of vibration frequency characteristics and amplitude characteristics, such as those of common bearing failures and fault classifications; for unknown state samples or without a large number of fault samples, a healthy state constraint is usually used). =0), making the model tend to reconstruct a healthy spectrum. If the sample is determined to be a certain type of fault, the degree of amplitude deviation of the reconstructed signal at the fault frequency can be analyzed (i.e., ... This deviation reference value provides a quantitative indicator of the relationship between fault type and severity. The specific expression for the energy conservation constraint term is: ; in, The time window length for integrating the signal energy (usually taken as one vibration period or a fixed analysis window length, and must be consistent with...) The corresponding integration intervals are consistent. Let be the expected vibration energy of the j-th second training sample. (In a mechanical vibration system, vibration energy usually refers to the total energy of the vibration signal in the time or frequency domain, reflecting the overall intensity of the system vibration. The expected vibration energy is not directly measured, but is a theoretical value calculated based on the vibration dynamics equations. For example, the total system energy consists of kinetic and potential energy.) .in, (represented as the energy of the reconstructed signal). The specific expression for the regularization loss term is: ; in, The first preset regularization coefficient is used. Indicated as to Find the L2 norm (weight decay term). It is represented as the set of all trainable weights. This is the second preset regularization coefficient. Indicated as to Find the L1 norm (sparse constraint term). For the first The output activation vector of the layer, The number of network layers to impose sparsity constraints. It's important to note that when training data is limited or noisy, neural networks tend to memorize noise and specific patterns in the training samples rather than learning general rules. Regularization, by penalizing model complexity, forces the network to learn smoother, more general feature representations. Ensuring that the trained model maintains good performance on unseen data (different operating conditions, different equipment, different loads) is crucial for industrial applications.

[0077] The composite loss function is designed as a weighted sum of data-driven loss, physical mechanism constraint loss, and regularization loss. and Weights are used to balance the importance of different loss terms during model training. For example, these weights can be determined through empirical setting, grid search, or Bayesian optimization to adapt to different device types, failure modes, or data characteristics, thereby achieving an optimal balance between data fitting, adherence to physical laws, and model generalization. The data-driven loss term measures the deviation between the deep learning network's predicted output and the actual observed data. This loss term combines reconstruction error and classification error; for example, the mean squared error can be used to quantify the reconstructed signal. Compared with actual vibration signals The differences between them prompt the model to learn normal vibration patterns under healthy conditions; at the same time, the cross-entropy loss is used to evaluate the predicted failure probability. Compared to actual fault labels The difference between them is addressed to improve the model's ability to distinguish different fault types. The physical mechanism constraint loss term aims to incorporate the inherent physical laws of rotating mechanical parts of industrial equipment into the training process of the deep learning network. Here, M represents the total number of the second training samples. , and These weights are used to adjust the relative contributions of dynamic equation constraint terms, frequency domain characteristic constraint terms, and energy conservation constraint terms to the physical mechanism constraints. These weights can be adjusted based on the importance of the physical mechanism or sensitivity to specific failure modes; for example, they can be appropriately increased when focusing on the dynamic response of the device. When emphasizing the characteristic frequency of faults, it can improve The constraints in the dynamic equations are used to ensure that the deep learning network reconstructs the signal. It conforms to the vibration dynamics equations of a rotating mechanical component. This term is calculated by determining the second derivative of the reconstructed signal. and first derivative The constraint term, combined with equivalent mass, damping coefficient, stiffness coefficient, and estimated excitation force, measures the degree to which the reconstructed signal satisfies the physical dynamic equations. For example, this constraint term can use the L2 norm to quantify the deviation between the model's predicted dynamic behavior and the theoretical physical model, thereby guiding the model to learn vibration modes that conform to physical laws. The frequency domain characteristic constraint term is used to impose physical constraints on the reconstructed signal in the frequency domain. This term compares the Fourier transform function of the reconstructed signal... Modulus value and preset physical constraint amplitude This ensures that the model's performance at critical failure frequencies matches physical expectations. For example, for a healthy state, It can be set to 0 to induce the model to reconstruct a stable spectrum; for specific faults, This can be set to the characteristic frequency amplitude corresponding to the fault, guiding the model to identify and quantify fault characteristics. The energy conservation constraint term ensures that the energy of the reconstructed signal conforms to the energy conservation principle of the physical system. This term compares the reconstructed signal... The difference between the energy integral within the time window T and the theoretically expected vibrational energy constrains the model's output. For example, the expected vibrational energy can be calculated based on the vibrational dynamics equations, ensuring that the model also follows physical laws at the energy level. Regularization loss terms aim to improve the generalization ability of deep learning models and prevent overfitting. For example, the L2 norm smooths the model by penalizing large weight values, reducing the model's sensitivity to noise in the training data; the L1 norm, on the other hand, encourages the model to learn sparse feature representations, further improving the model's robustness and generalization ability.

[0078] This composite loss function optimizes the training of deep learning network parameters by integrating data fitting ability, physical mechanism constraints, and model generalization constraints into a single training objective. Specifically, the data-driven loss term, combining mean squared error (MSE) and cross-entropy loss, enables the deep learning network to learn the normal vibration patterns of industrial equipment under healthy conditions from full-series time-series state data and improves its ability to distinguish different fault types. The MSE term ensures the model accurately reconstructs the input signal, capturing the inherent patterns in the data; the CrossEntropy term directly optimizes the model's classification performance, enabling accurate fault identification. Building upon this, the introduction of the physical mechanism constraint loss term, through a combination of dynamic equation constraints, frequency domain feature constraints, and energy conservation constraints, deeply integrates the inherent physical laws of the rotating mechanical components of industrial equipment into the training process of the deep learning network. The dynamic equation constraints ensure that the reconstructed vibration signal conforms to the fundamental physical laws of mechanical vibration in the time domain, i.e., satisfies… This dynamic equilibrium relationship. The frequency domain feature constraint term constrains the spectral characteristics of the reconstructed signal in the frequency domain. By comparing the Fourier transform magnitude of the reconstructed signal with the preset physical constraint amplitude, the model can accurately reflect physical phenomena at key fault characteristic frequencies. For example, it can suppress the amplitude of fault characteristic frequencies in a healthy state or highlight their characteristics when a fault occurs. The energy conservation constraint term further constrains the model from an energy perspective, ensuring that the energy of the reconstructed signal is consistent with the theoretically expected vibration energy, thereby enhancing the physical rationality of the model. Furthermore, the regularization loss term, through the weight decay term and the sparsity constraint term, effectively suppresses the overfitting phenomenon that may occur in deep learning networks during training. The weight decay term penalizes excessively large model weights, prompting the model to learn smoother and more generalized feature representations; the sparsity constraint term encourages the network to learn sparse activation patterns, further improving the model's robustness and generalization ability to unseen data. Through the first preset weight... and and the second preset weight , and This allows for flexible adjustment of the contributions of data-driven approaches, physical mechanism constraints, and regularization to the total loss, thereby balancing the optimization objectives of data fitting, adherence to physical laws, and generalization ability. This refined loss function design enables deep learning networks, under a two-stage training strategy, to more effectively utilize full-series temporal state data, establish a health baseline pattern with physical mapping relationships, accurately calculate the deviation of latent feature vectors from the health baseline pattern, and ultimately output reliable fault diagnosis results and fault severity indices.

[0079] As a specific implementation method, suppose that when building a fault diagnosis model for an industrial fan bearing, parameter optimization training of a deep learning network is required. During training, firstly, for the calculation of the data-driven loss term, N first training samples are selected from the collected full-series time-series state data. For each sample i, the model outputs a reconstructed vibration signal. and a predicted failure probability At this point, the system will calculate... Compared with actual vibration signals The mean square error between them, and Compared to actual fault labels The cross-entropy loss between them. For example, if the actual vibration signal The model reconstructs the original time-domain signal measured by the sensor. If it is the signal output by the decoder, then the MSE term will prompt... as close as possible If the actual fault label This indicates an "inner loop failure," while the model predicts... The lower probability of "inner circle faults" leads to a larger loss in the CrossEntropy term, prompting the model to adjust its parameters to improve the accuracy of identifying "inner circle faults." The preset reconstruction coefficient and classification loss weighted normalization coefficient can be set to 0.6 and 0.4, respectively, to balance reconstruction accuracy and classification performance. Secondly, for the calculation of the physical mechanism constraint loss term, M second training samples are selected from the full time-series state data. For each sample j, the model outputs a reconstruction signal. The system will calculate the constraint terms of the dynamic equations based on preset equivalent mass, damping coefficient, stiffness coefficient, and estimated excitation force. For example, if the physical parameters of the wind turbine bearing are known, and the signal for model reconstruction is... After substituting into the dynamic equation, there is a significant deviation from the estimated excitation force. This will be relatively large, prompting the model to adjust to conform to the laws of physics. Simultaneously, it will reconstruct the signal. Perform Fourier transform to obtain And calculate the frequency domain characteristic constraint terms. For example, if, under healthy conditions, the amplitude at the fault characteristic frequency of the bearing should be close to zero, but the model reconstruction... If a significant peak occurs at this frequency, then This will result in losses. In addition, energy conservation constraints will be calculated, where... This is the theoretically expected vibration energy calculated based on the physical model of the wind turbine bearing. For example, if the energy of the reconstructed signal from the model does not match the theoretical energy, then... This will result in a loss. Second preset weight. , and These can be set to 0.3, 0.4, and 0.3 respectively to balance the importance of different physical constraints. Finally, the regularization loss term is calculated based on the set of all trainable weights of the deep learning network. and the output activation vectors of each layer Perform. For example, through calculation. L2 norm and activation vector The L1 norm, multiplied by a first preset regularization coefficient and a second preset regularization coefficient (e.g., =0.001, =0.0005), to penalize model complexity and encourage sparsity, thereby improving the model's generalization ability and enabling it to maintain good performance when faced with different operating conditions or data without faults. The total loss is obtained by weighted summation of the above losses. The backpropagation algorithm is then used to optimize the parameters of the deep learning network until the model converges.

[0080] Through the above technical solution, this application provides a specific quantitative construction method for a composite loss function, integrating data fitting ability, physical mechanism constraints, and model generalization constraints into the same training objective. This solves the problem that physical constraints are difficult to effectively integrate into the calculation of training loss in deep learning in traditional methods, significantly improving model performance. Specifically, the introduction of the data-driven loss term enables the model to fully utilize data features and improve the accuracy of fault mode identification; the detailed construction of the physical mechanism constraint loss term ensures that the model follows the inherent physical laws of the rotating mechanical parts of industrial equipment during the learning process, thereby enhancing the physical rationality and interpretability of the model, reducing the dependence on a large amount of labeled fault data, and showing higher accuracy and sensitivity, especially in the identification of early weak fault signals. The addition of the regularization loss term effectively prevents model overfitting and improves the model's generalization ability under different working conditions, different equipment, and different load conditions. By flexibly adjusting the weights of various losses, the optimization objectives between data-driven approaches, physical constraints, and model generalization can be balanced according to the needs of actual application scenarios. This enables the constructed industrial equipment fault diagnosis model to not only accurately diagnose faults but also provide physically meaningful explanations, thereby gaining the trust of industrial field maintenance personnel. This effectively solves the core defects of traditional technologies, such as weak early fault identification capabilities, poor interpretability, lack of physical meaning, and strong data dependence.

[0081] In some embodiments, after outputting fault diagnosis results and fault severity index based on the degree of deviation, the industrial equipment fault diagnosis model further includes the following steps: S51. Map the latent eigenvectors to the physical dimensions corresponding to the constraints of the dynamic equation, the frequency domain eigenvalues, and the energy conservation constraints, and calculate the reconstruction residuals under each physical dimension. S52. Based on the magnitude distribution of each reconstructed residual, identify the dominant physical factors that cause the potential feature vector to deviate from the healthy state feature vector; S53. Output the dominant physical factors and use them as the basis for the physical interpretability of the fault diagnosis results.

[0082] The process involves mapping latent feature vectors to the physical dimensions corresponding to constraints in the dynamic equations, frequency domain features, and energy conservation constraints. This aims to transform the abstract latent feature vectors extracted by the deep learning network into quantities with explicit physical meaning. Latent feature vectors are compact, high-dimensional representations learned and extracted by the deep learning network from raw time-series state data, containing rich information about the device's operating state. To achieve physical interpretability, these abstract features need to be linked to the inherent physical laws of the device. One approach is to set a specific output layer in the decoder part of the deep learning network, such as a linear transform layer or a small fully connected network, to convert the latent feature vectors into physical quantities or their surrogate values ​​related to the dynamic equations, frequency domain features, or energy conservation. Another approach is to use a part of the model decoder to reconstruct the latent feature vectors into signals or parameters related to each physical dimension, such as reconstructed vibration signals or their spectra, and then calculate the deviation from theoretical physical laws based on these reconstructed signals / parameters.

[0083] After completing the above mapping, the reconstruction residuals for each physical dimension need to be calculated separately. The reconstruction residual is a quantitative indicator that measures the difference between the model's output or reconstruction result and the expected physical law (or health baseline) in a specific physical dimension. These residuals can intuitively reflect the degree of anomaly in the corresponding physical dimension. Specifically, for the constraints of the dynamic equations, the residual can be the unbalanced force or its norm calculated after substituting the reconstructed vibration signal into the vibration dynamic equations. For the frequency domain characteristic constraints, the residual can be the difference between the amplitude of the Fourier transform of the reconstructed signal at a specific fault characteristic frequency and the amplitude of the health baseline (e.g., the amplitude at this frequency should be zero in a healthy state). For the energy conservation constraints, the residual can be the difference between the integral energy of the reconstructed signal within a specific time window and the theoretically expected energy calculated based on the physical mechanism.

[0084] Subsequently, based on the magnitude distribution of each reconstructed residual, the dominant physical factors causing the potential feature vector to deviate from the healthy state feature vector are identified. By comparing the magnitudes of the residuals in different physical dimensions, it can be determined which physical dimension exhibits the most significant anomaly, thereby identifying the primary physical cause of equipment failure. One identification method is to set a preset threshold for the residuals of each physical dimension; when the residual of a certain dimension exceeds this threshold, that dimension is considered to have an anomaly. Another identification method is to normalize the residuals of each physical dimension and then compare the magnitudes of the normalized residuals; the physical dimension with the largest residual value is identified as the dominant physical factor.

[0085] Finally, the dominant physical factors are output as the basis for the physical interpretability of the fault diagnosis results. This means presenting the identified dominant physical factors to maintenance personnel in human-understandable language or form as a clear explanation of the fault diagnosis results. For example, text descriptions such as "dynamic imbalance," "periodic wear," or "abnormal system energy dissipation" can be output directly. Alternatively, a visualization interface can be used to display the residuals and their magnitude distribution for each physical dimension in chart form, highlighting the dominant factors for a more intuitive explanation.

[0086] This application's scheme maps latent feature vectors to the physical dimensions corresponding to constraints in the dynamic equation, frequency domain features, and energy conservation, respectively. It then calculates the reconstruction residuals for each physical dimension and identifies the dominant physical factors based on the distribution of these residuals. Finally, it outputs these dominant physical factors as the physical interpretability basis for the fault diagnosis results. This series of steps enables industrial equipment fault diagnosis models to transform abstract feature deviations into concrete physical causes, thus solving the problem of insufficient interpretability of deep learning models as "black boxes." This scheme relies on the multi-dimensional physical mechanism constraints already incorporated into the model during training, providing a clear physical interpretation for the fault diagnosis results. The dynamic equation constraints, frequency domain features, and energy conservation constraints are chosen as the analysis dimensions because these dimensions are themselves used during model training to constrain the model's learning of the physical laws of the equipment; they correspond to the inherent physical characteristics of different levels of equipment faults. This design avoids constructing additional analysis dimensions, reducing the additional computational complexity of interpretability analysis and ensuring the consistency between the interpretability analysis process and the model's learning logic, avoiding the problem of additional analysis being disconnected from the model logic. By quantifying the reconstruction residuals of each physical dimension, the degree of anomaly in the corresponding physical dimension can be clearly reflected, providing a reliable quantitative basis for subsequent judgment of the source of anomalies. By comparing the magnitudes of the residuals in different physical dimensions, it is possible to accurately pinpoint which aspect of physical anomaly is the main cause of the overall deviation from a healthy state, completing the transformation from abstract feature deviation to specific physical causes, and allowing the abstract model output to be grounded in an understandable physical level.

[0087] For example, when monitoring the bearing of an industrial pump in real time, the industrial equipment fault diagnosis model first extracts a latent feature vector based on the input real-time time-series state data. Then, this latent feature vector is mapped to the physical dimensions corresponding to the constraints of the dynamic equation, frequency domain feature constraints, and energy conservation constraints. For instance, the latent feature vector can be input into three independent, pre-trained small neural networks, each responsible for converting it into a reconstructed signal or parameter associated with a physical dimension. For example, one network might reconstruct the bearing's vibration signal, another might reconstruct the signal's spectrum, and a third might reconstruct the signal's energy. Subsequently, the system calculates the residuals between these reconstructed results and the corresponding physical laws under a health baseline model. For example, if substituting the reconstructed vibration signal into the vibration dynamic equation results in a significantly increased unbalanced force residual, while the frequency domain feature residual and energy conservation residual are relatively small, it indicates that the reconstructed residuals corresponding to the constraints of the dynamic equation have a maximum value distribution. In this case, the system will identify dynamic imbalance as the dominant physical factor and output the explanation that "the monitored equipment has dynamic imbalance or misalignment." This allows maintenance personnel to directly obtain the physical causes of the fault, rather than just an abstract fault conclusion without any basis. This increases the maintenance personnel's trust in the diagnostic results and provides clear guidance for subsequent targeted maintenance.

[0088] Through the above technical solution, this application provides clear physical interpretability for industrial equipment fault diagnosis models, solving the "black box" problem of traditional deep learning models and enabling maintenance personnel to understand the underlying physical causes of fault diagnosis results. This not only improves the reliability and credibility of diagnostic results but also provides strong support for accurate fault location and targeted maintenance, thereby improving the operating efficiency and safety of industrial equipment.

[0089] In some embodiments, the specific steps in step S53 include: S53a. When the reconstruction residual corresponding to the constraint term of the dynamic equation is the maximum value distribution, output the explanation information that the monitored equipment has dynamic imbalance or misalignment; S53b. When the reconstructed residual corresponding to the frequency domain feature constraint term is the maximum value distribution, output the explanation information of the periodic wear fault of the monitored equipment; S53c. When the reconstruction residual corresponding to the energy conservation constraint term is the maximum value distribution, output the explanation information of the abnormal system energy dissipation of the monitored equipment.

[0090] Specifically, when the reconstructed residual corresponding to the constraint term of the dynamic equation has a maximum value distribution, it indicates that the signal reconstructed by the model deviates significantly from the expected healthy pattern in terms of dynamic behavior. This deviation usually originates from abnormal forces or torques experienced by rotating mechanical components during operation, causing their motion trajectory, velocity, or acceleration, and other dynamic parameters to deviate from the normal range. The system can preset an interpretation information database containing fault descriptions related to different physical factors. When the maximum residual of the constraint term of the dynamic equation is detected, the system retrieves and outputs the text information "The monitored equipment has a dynamic imbalance or misalignment" from the database. Alternatively, the system can dynamically generate or combine interpretation information based on the magnitude and trend of the residual. For example, if the residual continues to increase, descriptions such as "and there is a deteriorating trend" may be added to the output information to provide more detailed diagnostic suggestions.

[0091] When the reconstruction residual corresponding to the frequency domain feature constraint term has a maximum value distribution, it means that the energy or amplitude of the reconstructed signal at a specific fault characteristic frequency differs significantly from the expected value under healthy conditions. This frequency domain anomaly is usually due to localized damage or wear of mechanical components (e.g., bearings), which periodically generate impacts or vibrations during rotation, thus manifesting as an enhancement of specific frequency components in the spectrum. The system can maintain a mapping table between fault frequencies and typical fault types (e.g., periodic wear). When the frequency domain residual is at its maximum, the system outputs an explanation of "the monitored equipment has a periodic wear fault" based on this mapping table. Alternatively, the system can further analyze which specific fault characteristic frequency (e.g., the bearing inner ring fault frequency, bearing outer ring fault frequency, or bearing rolling element fault frequency) has the largest residual, thereby outputting more specific wear location information, such as "the monitored equipment has a periodic wear fault in the bearing inner ring".

[0092] When the reconstruction residual corresponding to the energy conservation constraint term has a maximum value distribution, it indicates that the signal reconstructed by the model deviates significantly from the theoretically expected healthy energy in terms of total energy. This overall energy anomaly may reflect problems such as increased internal friction, altered damping characteristics, or decreased energy transfer efficiency, leading to abnormal energy dissipation or accumulation in the system. The system can set a threshold; when the residual of the energy conservation constraint term exceeds this threshold and reaches its maximum, it outputs the explanation information "The monitored device exhibits abnormal system energy dissipation." Alternatively, the system can combine other sensor data (such as temperature and power consumption) to further infer possible specific causes based on the abnormal energy dissipation, such as "The abnormal system energy dissipation may be related to excessive friction leading to temperature rise."

[0093] This application's solution maps the potential feature vectors calculated by the industrial equipment fault diagnosis model to the physical dimensions corresponding to the constraints of the dynamic equation, frequency domain feature constraints, and energy conservation constraints, respectively. It then calculates the reconstruction residuals under each physical dimension to identify the dominant physical factors causing the potential feature vectors to deviate from the healthy state feature vectors. Furthermore, this application combines the physical meanings of different physical constraint dimensions to directly map the abstract dominant physical factors to specific fault interpretation conclusions that industrial field maintenance personnel can directly understand. For example, when the reconstruction residual of the dynamic equation constraint is the largest, the explanation information of "dynamic imbalance or misalignment" is directly output. This relies on the characteristic of the dynamic equation constraint itself describing the vibration motion law of mechanical rotating parts; the largest residual indicates the highest degree of deviation of the reconstructed signal from the healthy state dynamic law, and dynamic imbalance and misalignment are the most common faults causing such deviations in dynamic law. When the reconstruction residual of the frequency domain characteristic constraint term is at its maximum, the explanation information for "periodic wear fault" is output. This relies on the physical mechanism of the frequency domain characteristic constraint term itself based on the fault characteristic frequency. The maximum residual indicates that the amplitude of the reconstructed signal at the fault characteristic frequency deviates from the expected state to the greatest extent, and periodic wear fault is a typical fault that will cause abnormal amplitude at the characteristic frequency. When the reconstruction residual of the energy conservation constraint term is at its maximum, the explanation information for "abnormal system energy dissipation" is output. This relies on the characteristics of the energy law of the vibration system itself based on the energy conservation constraint term itself. The maximum residual indicates that the total energy of the reconstructed signal deviates from the theoretically expected energy to the greatest extent, and abnormal energy dissipation corresponds precisely to this type of fault manifestation of overall energy deviation. This direct and physically based mapping transforms physical interpretability from abstract dominant factors into fault cause conclusions that can be directly used by operation and maintenance, giving full play to the advantages of physical mechanism embedding, and conforming to the actual occurrence law of industrial rotating component faults, thus ensuring the accuracy of the explanation.

[0094] The following is a concrete example to illustrate this. Consider an industrial pump whose rotating mechanical components (e.g., bearings) are in operation. Real-time time-series status data is collected and input into a pre-trained industrial equipment fault diagnosis model. The model extracts latent feature vectors from this data and calculates their deviation from the health state feature vector corresponding to the health baseline model. Subsequently, these latent feature vectors are mapped to the physical dimensions corresponding to the kinetic equation constraint terms, frequency domain feature constraint terms, and energy conservation constraint terms, respectively, and the reconstruction residuals under each physical dimension are calculated. Suppose that at a certain moment, the system detects that the magnitude distribution of the reconstruction residuals corresponding to the kinetic equation constraint terms is significantly greater than the residuals of the frequency domain feature constraint terms and the energy conservation constraint terms. At this time, the system will output the explanatory information: "The monitored equipment has a kinetic imbalance or misalignment." Upon receiving this information, maintenance personnel can immediately check the pump's dynamic balance or shaft alignment, for example, by using a laser alignment instrument to check the coupling alignment, or by conducting on-site dynamic balancing tests, thereby quickly locating and resolving the fault.

[0095] Through the above technical solution, this application significantly improves the physical interpretability of fault diagnosis results. By directly mapping abstract physical factors (such as the maximum residual of constraint terms in dynamic equations) to specific fault types (such as dynamic imbalance or misalignment) that industrial field maintenance personnel can understand, the dilemma of requiring further analysis to clarify the cause of the fault is avoided. This enables maintenance personnel to quickly and accurately understand the physical nature of the fault, thereby significantly shortening the time for fault diagnosis and repair, and improving the operating efficiency and safety of industrial equipment.

[0096] In some embodiments, the specific steps of the industrial equipment fault diagnosis model to calculate the deviation of the potential feature vector from the health state feature vector corresponding to the health baseline mode, and to output the fault diagnosis result and fault severity index based on the deviation include: The fault severity index is calculated using the following formula (based on the degree of deviation, the L2 norm of the fault feature space is obtained, i.e., the Euclidean distance in the vector space): ; in, This is a severity index of the fault. Indicated as to Find the L2 norm (i.e., the first Euclidean distance in the vector space). Indicated as to Find the L2 norm (i.e., the second Euclidean distance in the vector space). For potential feature vectors, The center of the health state feature vector (i.e., the mean of all specified time-series state data used as training samples for health states). This is a preset feature vector of severe faults; The fault diagnosis result can include the classification of the fault severity: FSI = 0%: Completely healthy state; 0% < FSI < 30%: Early fault (planned maintenance can be carried out); 30% ≤ FSI < 70%: Moderate fault (it is recommended to repair as soon as possible); FSI ≥ 70%: Severe fault (stop the machine immediately for repair); FSI > 100%: The fault is more serious than the "severe fault" defined.

[0097] This calculation method aims to quantify the deviation degree between the current device state and the healthy state and standardize it. By calculating the Euclidean distance in the fault feature space, the difference between the potential feature vector and the center of the healthy state feature vector can be intuitively characterized. This quantification method based on the Euclidean distance can effectively capture the overall deviation trend in the multi-dimensional feature space and provide a reliable numerical basis for the subsequent evaluation of the fault severity. Among them, the above formula defines the calculation method of the fault severity index FSI. The numerator represents the Euclidean distance between the current potential feature vector and the center of the healthy state feature vector, reflecting the actual deviation degree between the current device state and the healthy benchmark. The denominator The Euclidean distance between the centers of the preset severe fault feature vector and the healthy state feature vector is used to normalize the deviation in the numerator, making the FSI a relative value between 0 and 1 (or higher), thus achieving comparability of fault severity under different equipment and operating conditions. The latent feature vector (Z) is a feature representation with a physical mapping relationship extracted from the input real-time time-series state data by the industrial equipment fault diagnosis model. This vector is generated by the encoder part of a deep learning network, effectively compressing the original high-dimensional time-series data while retaining key information related to equipment operating status and fault modes. The dimension and specific meaning of the latent feature vector Z can be designed according to the structure and training objectives of the deep learning network. For example, its components can correspond to physical quantities such as total vibration energy, main fault frequency energy, high-frequency impact energy, or non-stationarity measures. The center of the healthy state feature vector represents the typical feature pattern of industrial equipment in a fully healthy state. It is usually obtained by processing a large amount of specified time-series state data collected under healthy equipment conditions and calculating the mean of the latent feature vectors corresponding to these healthy samples. Using the mean as the health benchmark can effectively eliminate the randomness or measurement noise of individual health samples, making the health benchmark more stable and accurate, thereby improving the robustness of the fault severity index calculation. The pre-defined severe fault feature vector represents the typical characteristic patterns of industrial equipment under severe fault conditions. This vector can be determined in various ways; for example, it can be based on expert experience, with domain experts annotating it according to historical fault data and equipment characteristics; or, it can be extracted from actual severe fault data and the mean or representative value of its potential feature vector can be calculated. As a crucial reference point in FSI calculation, it provides an upper limit benchmark for quantifying fault severity. The classification of fault severity aims to map the continuous Fault Severity Index (FSI) to discrete fault levels with clear operational guidance. These classification intervals and their corresponding descriptions (such as "Complete Health Status," "Early Fault (Planned Repair)," "Moderate Fault (Recommended Repair as Soon as Possible)," "Severe Fault (Immediate Shutdown for Repair)," and "Fault More Severe Than Defined as 'Severe Fault'") provide operations personnel with intuitive and easily understandable fault status assessments. By setting different FSI thresholds, refined management of equipment health can be achieved, supporting preventative maintenance and emergency shutdown decisions.

[0098] This solution cleverly transforms the abstract deviation of feature vectors into an intuitive and quantifiable fault severity index through a calculation method based on Euclidean distance ratio. This quantification method not only accurately captures subtle changes in equipment status but also, due to its standardized nature, ensures horizontal comparability of diagnostic results across different equipment and operating conditions. Building upon this, the solution further divides the FSI value into multiple clearly defined intervals, each corresponding to a specific fault severity level, along with corresponding maintenance recommendations. For example, an FSI of 0% indicates a perfectly healthy state; 0% to 30% indicates an early-stage fault requiring planned maintenance; 30% to 70% indicates a moderate fault requiring prompt maintenance; and 70% and above indicates a severe fault requiring immediate shutdown for maintenance. This grading mechanism significantly lowers the barrier to interpreting fault diagnosis results, enabling maintenance personnel to quickly and accurately assess equipment health based on the FSI value and its corresponding level, and make timely and appropriate maintenance decisions. This solution, combined with the steps of the industrial equipment fault diagnosis model—extracting potential feature vectors with physical mapping relationships from the input real-time time-series status data and calculating the deviation of these potential feature vectors from the health status feature vectors corresponding to the health baseline model—constitutes a complete fault diagnosis chain from data acquisition, feature extraction, status assessment to decision support. The physical mapping relationship of the potential feature vectors ensures the physical meaning of FSI, while the standardization and classification of FSI transforms complex physical states into operable operation and maintenance instructions, thereby significantly improving the practicality and interpretability of fault diagnosis.

[0099] As a specific implementation method, it is assumed that in the industrial equipment fault diagnosis model, the latent feature vector is a two-dimensional vector used to characterize a certain operating state of the equipment. By processing a large amount of time-series state data under specified healthy states, the center of the healthy state feature vector is calculated. The value is [0.1, 0.2]. Simultaneously, based on historical severe fault data or expert experience, a pre-defined severe fault feature vector is established. The value is determined to be [1.5, 2.0]. When the potential feature vector of the device is monitored in real time as [0.8, 1.0], the Fault Severity Index (FSI) can be calculated by following these steps: First, calculate the Euclidean distance (i.e., the first Euclidean distance in the vector space) between the center of the current potential feature vector and the center of the health status feature vector. Secondly, calculate the Euclidean distance (i.e., the second Euclidean distance in the vector space) between the centers of the preset severe fault feature vector and the healthy state feature vector: Finally, the Fault Severity Index (FSI) is calculated: FSI = 1.063 / 2.280 ≈ 0.4662, or 46.62%. Based on the preset fault severity classification criteria, since the calculated FSI of 46.62% falls within the range of 30% ≤ FSI < 70%, the industrial equipment can be diagnosed as having a moderate fault, and prompt repair is recommended. This example clearly demonstrates how standardized FSI calculation and classification can provide maintenance personnel with clear fault diagnosis results and action recommendations.

[0100] This application's solution addresses the lack of intuitive quantitative references for fault severity in traditional methods by establishing a standardized fault severity index quantification method, coupled with clear fault severity grading rules. By calculating the Fault Severity Index (FSI) based on the deviation of the latent feature vector from the health baseline and utilizing the ratio of Euclidean distances in the fault feature space, the deviation of the tested sample's features from the health baseline can be accurately quantified. This method uses health status features and preset severe fault features as reference benchmarks, making the calculated FSI a standardized relative value. This ensures the comparability of diagnostic results across different operating conditions and equipment, avoiding incomparability issues caused by differences in feature space numerical ranges. Furthermore, using the mean of all healthy training samples as the center of the health status feature vector effectively eliminates the random interference of individual healthy samples, making the health baseline more stable and accurate, thereby improving the accuracy of the fault severity index calculation. By dividing the Fault Severity Index (FSI) into different numerical ranges and assigning clear fault severity levels and maintenance recommendations to each range, maintenance personnel can quickly determine the fault situation based on the output FSI and make appropriate maintenance decisions. This significantly reduces the threshold for interpreting fault diagnosis results and greatly improves the practicality and decision support capabilities of fault diagnosis methods.

[0101] Please refer to Figure 2 , Figure 2 This invention relates to an industrial equipment fault diagnosis model building device (which employs the industrial equipment fault diagnosis model building method described in the above embodiments, with specific steps as described above) in some embodiments. The device is used to build an industrial equipment fault diagnosis model for diagnosing faults in rotating mechanical parts of industrial equipment. The device is integrated into a back-end control device in the form of a computer program. The industrial equipment fault diagnosis model building device includes: The acquisition module 100 is used to acquire the physical mechanism model of the industrial equipment; the physical mechanism model includes physical constraint equations. Module 200 is used to build deep learning networks; the deep learning networks include composite loss functions constructed based on physical constraint equations. The acquisition module 300 is used to acquire full-time status data of industrial equipment during operation; the full-time status data includes the specified time-series status data acquired when the industrial equipment is in a healthy state; Training module 400 is used to optimize the parameters of the deep learning network based on the full amount of time-series state data, so as to obtain the final usable industrial equipment fault diagnosis model. Parameter optimization training specifically includes the following steps A1-A2: A1. Using specified time-series state data, establish a healthy baseline model by pre-training a deep learning network; A2. After establishing a health baseline model, the deep learning network is jointly trained using full-scale time-series state data.

[0102] Please refer to Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The present invention provides an electronic device 13, including: a processor 1301 and a memory 1302. The processor 1301 and the memory 1302 are interconnected and communicate with each other through a communication bus 1303 and / or other forms of connection mechanism (not shown). The memory 1302 stores computer-readable instructions executable by the processor 1301. When the electronic device is running, the processor 1301 executes the computer-readable instructions to execute the industrial equipment fault diagnosis model construction method in any optional implementation of the above embodiments, so as to achieve the following functions: obtaining the physical mechanism model of the industrial equipment; the physical mechanism model includes physical constraint equations; constructing a deep learning network; the deep learning network includes a composite loss function constructed according to the physical constraint equations; collecting full time-series state data of the industrial equipment during operation; the full time-series state data includes specified time-series state data collected when the industrial equipment is in a healthy state; and performing parameter optimization training on the deep learning network based on the full time-series state data to obtain a final usable industrial equipment fault diagnosis model.

[0103] This invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it performs the industrial equipment fault diagnosis model construction method in any optional implementation of the above embodiments to achieve the following functions: obtaining a physical mechanism model of the industrial equipment; the physical mechanism model includes physical constraint equations; constructing a deep learning network; the deep learning network includes a composite loss function constructed based on the physical constraint equations; collecting full-series time-series state data of the industrial equipment during operation; the full-series time-series state data includes specified time-series state data collected when the industrial equipment is in a healthy state; and performing parameter optimization training on the deep learning network based on the full-series time-series state data to obtain a final usable industrial equipment fault diagnosis model.

[0104] The computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0105] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0106] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0107] Furthermore, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0108] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0109] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An industrial equipment failure diagnosis model construction method for constructing an industrial equipment failure diagnosis model for diagnosing a failure of a mechanical rotating component in an industrial equipment, characterized by, The method for constructing the industrial equipment fault diagnosis model includes the following steps: S1. Obtain the physical mechanism model of the industrial equipment; the physical mechanism model includes physical constraint equations; the physical constraint equations include vibration characteristic equations, vibration dynamics equations, and fault characteristic frequency calculation equations; S2. Construct a deep learning network; the deep learning network includes a composite loss function constructed based on the physical constraint equations; the composite loss function includes a data-driven loss term, a physical mechanism constraint loss term constructed based on the physical constraint equations, and a regularization loss term; the physical mechanism constraint loss term includes a dynamic equation constraint term, a frequency domain feature constraint term, and an energy conservation constraint term. S3. Collect full time-series status data of the industrial equipment during operation; the full time-series status data includes specified time-series status data collected when the industrial equipment is in a healthy state; S4. Based on the full time-series state data, perform parameter optimization training on the deep learning network to obtain the final usable industrial equipment fault diagnosis model; The parameter optimization training specifically includes the following steps A1-A2: A1. Using the specified time-series state data, a health baseline model is established by pre-training the deep learning network; A2. After establishing the health baseline model, the deep learning network is jointly trained using the full amount of time-series state data; When the mechanical rotating component is a bearing, the fault characteristic frequency calculation equation includes the calculation equation for the bearing inner ring fault frequency, the calculation equation for the bearing outer ring fault frequency, and the calculation equation for the bearing rolling element fault frequency. The specific equation for calculating the failure frequency of the bearing inner ring is as follows: ； in, This refers to the failure frequency of the bearing inner ring. This refers to the number of rolling elements in the bearing. The diameter of the rolling element, For bearing pitch diameter, Contact angle, The rotational frequency of the axis; The specific equation for calculating the failure frequency of the bearing outer ring is as follows: ； in, This refers to the failure frequency of the bearing outer ring. The specific equation for calculating the failure frequency of bearing rolling elements is as follows: ； in, This represents the failure frequency of the bearing rolling elements.

2. The method for constructing an industrial equipment fault diagnosis model according to claim 1, characterized in that, The specific expression for the vibration characteristic equation is as follows: ； in, The vibration signal of the mechanical rotating component at time t. The signal represents the normal vibration of the rotating mechanical component of the industrial equipment at time t when it is in a healthy state. The abnormal vibration signal caused by the failure of the mechanical rotating component at time t. Let be the measurement noise at time t.

3. The method for constructing an industrial equipment fault diagnosis model according to claim 1, characterized in that, The specific expression for the vibration dynamics equation is as follows: ； in, For equivalent quality, for The second derivative, The damping coefficient is... for The first derivative, This is the stiffness coefficient. For the normal excitation force at time t, The additional excitation force caused by the fault at time t.

4. The method for constructing an industrial equipment fault diagnosis model according to claim 1, characterized in that, The specific expression for the composite loss function is as follows: ； in, The composite loss function, For the data-driven loss term, and All are the first preset weights. The physical mechanism constraint loss term, The regularization loss term is mentioned above. The specific expression for the data-driven loss term is: ； in, The total number of the first training samples, To preset reconstruction coefficients, for The mean square error, For the i-th first training sample, the actual vibration signal is... The vibration signal reconstructed from the i-th first training sample. The weighted normalization coefficients for classification loss, For about The cross-entropy loss function, For the i-th of the first training samples, Let be the predicted failure probability of the i-th first training sample; The specific expression for the physical mechanism constraint loss term is as follows: ； in, This represents the total number of the second training samples. , and All are the second preset weights. For the j-th second training sample, the dynamic equation constraint term is... For the j-th second training sample, the frequency domain feature constraint term is... For the j-th second training sample, this is the energy conservation constraint term; The specific expression for the constraint terms of the dynamic equation is as follows: ； in, Indicated as to Find the L2 norm. for The second derivative, for The first derivative, The reconstructed signal of the mechanical rotating component in the j-th second training sample at time t. Let be the estimated excitation force at time t; The specific expression for the frequency domain feature constraint term is as follows: ； in, Specify the set of characteristic frequencies for the fault class. To indicate as to Modulus, For about The Fourier transform function of the reconstructed signal of the j-th second training sample, for The physical constraint corresponding to the amplitude of the p-th characteristic frequency in the equation; The specific expression for the energy conservation constraint term is as follows: ； in, The length of the time window for integrating the signal energy. Let be the expected vibrational energy of the j-th second training sample; The specific expression for the regularization loss term is as follows: ； in, The first preset regularization coefficient is used. Indicated as to Find the L2 norm. It is represented as the set of all trainable weights. This is the second preset regularization coefficient. Indicated as to Find the L1 norm. For the first The output activation vector of the layer, The number of network layers to which sparse constraints are applied.

5. An industrial equipment fault diagnosis model construction apparatus employing the industrial equipment fault diagnosis model construction method as described in any one of claims 1-4, used to construct an industrial equipment fault diagnosis model, wherein the industrial equipment fault diagnosis model is used to diagnose faults in mechanical rotating parts of industrial equipment, characterized in that, The industrial equipment fault diagnosis model construction device includes: An acquisition module is used to acquire the physical mechanism model of industrial equipment; the physical mechanism model includes physical constraint equations. A building module for constructing a deep learning network; the deep learning network includes a composite loss function constructed based on the physical constraint equations; The data acquisition module is used to acquire full-time status data of the industrial equipment during operation; the full-time status data includes specified time-series status data acquired when the industrial equipment is in a healthy state; The training module is used to perform parameter optimization training on the deep learning network based on the full set of time-series state data to obtain a final usable industrial equipment fault diagnosis model. The parameter optimization training specifically includes the following steps A1-A2: A1. Using the specified time-series state data, a health baseline model is established by pre-training the deep learning network; A2. After establishing the health baseline model, the deep learning network is jointly trained using the full amount of time-series state data.

6. An electronic device, characterized in that, It includes a processor and a memory, the memory storing computer-readable instructions, which, when executed by the processor, perform the steps in the industrial equipment fault diagnosis model construction method as described in any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it performs the steps in the industrial equipment fault diagnosis model construction method as described in any one of claims 1-4.

Images