A Smart Diagnostic Method for Harmonic Reducers

Abstract

This invention provides an intelligent diagnostic method for harmonic reducers based on improved Markov graph information and graph attention networks, belonging to the field of fault diagnosis technology. The method includes: collecting vibration signals from the harmonic reducer to construct a training set; converting one-dimensional time-series signals into graph-structured data using an improved Markov graph information method, and training a neural network model based on a graph attention mechanism using the graph-structured data to obtain a fault diagnosis model capable of identifying fault categories based on the graph-structured data; collecting vibration signals from the harmonic reducer to be diagnosed to obtain one-dimensional time-series vibration test data, converting the data into test graph-structured data and inputting it into the trained fault diagnosis model, and outputting the fault classification result of the harmonic reducer to be diagnosed. This invention achieves end-to-end intelligent diagnosis, avoids manual feature engineering, has strong robustness to noise and operating condition changes, and has high diagnostic accuracy, making it suitable for predictive maintenance of harmonic reducers in industrial robots.

Description

Technical Field

[0001] This invention relates to the field of industrial robot fault diagnosis, specifically to an intelligent diagnostic method for harmonic reducers based on the collaborative optimization of improved Markov graph information and graph attention network. Background Technology

[0002] Against the backdrop of global manufacturing's intelligent upgrading, industrial robots have become key equipment for improving production efficiency. Harmonic reducers, as their core transmission components, endure high-frequency alternating loads and extreme operating conditions for extended periods; their failure is a major cause of robot system downtime. Therefore, achieving accurate and intelligent diagnostics for harmonic reducers is crucial for ensuring production safety and improving equipment reliability.

[0003] Existing fault diagnosis technologies for harmonic reducers mainly fall into three categories: mechanism modeling, signal processing, and data-driven methods. Mechanism modeling relies on precise physical descriptions, but harmonic reducers have complex structures, making modeling difficult and parameter identification inaccurate. Signal processing methods (such as time-frequency analysis) heavily rely on expert experience for feature engineering and have limited ability to fit nonlinear relationships. Data-driven methods based on deep learning, such as convolutional neural networks (CNNs) and long short-term memory networks (LSTMs), suffer from gradient vanishing / exploding and difficulty in effectively extracting long-range temporal dependencies when directly processing one-dimensional long-sequence vibration signals; while methods that convert time series into two-dimensional time-frequency graphs suffer from information redundancy, low computational efficiency, and reliance on expert selection for transformation methods.

[0004] Graph theory offers a new approach to solving the aforementioned problems. Graph structures can effectively represent the relationships between nodes in non-Euclidean data, providing more concentrated information. Traditional Markov graph information methods can convert time-series signals into graphs, but they directly use the maximum and minimum values ​​of the sequence as boundaries when partitioning states, making them extremely sensitive to noise and outliers in the data. This leads to unstable state partitioning criteria, affecting the robustness of subsequent graph neural network models. Therefore, a more stable and automated intelligent diagnostic method is urgently needed. Summary of the Invention

[0005] In view of the technical defects and drawbacks existing in the prior art, the present invention provides an intelligent diagnostic method for harmonic reducers that overcomes or at least partially solves the above problems. The method can automatically and robustly extract deep fault features from vibration signals, achieve high-precision fault classification, and has strong generalization ability to changes in operating conditions. The specific solution is as follows.

[0006] A method for intelligent diagnosis of harmonic reducers includes a model building stage and a fault diagnosis stage;

[0007] The model building phase includes the following steps:

[0008] S1. Obtain the vibration signals of the harmonic reducer operating under various known fault conditions and construct a one-dimensional time-series vibration training dataset;

[0009] S2. Using the improved Markov graph information method, each time-series sample in the one-dimensional time-series vibration training dataset is converted into corresponding graph structure data, wherein the graph structure data represents the data distribution with state nodes and the state transition probability is represented by the edges between nodes.

[0010] S3. Use the graph structure data to train a neural network model based on graph attention mechanism to obtain a fault diagnosis model that can identify fault categories based on graph structure data;

[0011] The fault diagnosis phase includes the following steps:

[0012] S4. Collect the vibration signal of the harmonic reducer to be diagnosed and obtain one-dimensional time-series vibration test data;

[0013] S5. Using the improved Markov graph information method, the one-dimensional time-series vibration test data is converted into test graph structure data;

[0014] S6. Input the test diagram structure data into the trained fault diagnosis model and output the fault classification result of the harmonic reducer to be diagnosed.

[0015] In some embodiments, the improved Markov graph information method specifically includes the following steps:

[0016] S21. State quantization: quantizing a one-dimensional temporal vibration data sequence of length n. Each data point x in t Quantization is mapped to one of B discrete states to obtain a state sequence. , where q t ∈ {1, 2, ...,B}, where t is the time sequence index;

[0017] S22. Node Attribute Generation: Based on the state sequence Q, count the frequency of each state and calculate the probability of each state, generating a state probability distribution vector P = { with dimension B. ..., p B}, representing the attribute set of all nodes in the graph structure data, where p b This represents the probability of state b occurring;

[0018] S23. Edge attribute generation: Based on the state sequence Q, count the transition frequency between states and calculate the transition probability w from state i to state j. ij Generate a Markov state transition probability matrix W of dimension B×B, and set each non-zero element w in matrix W to...ij As an attribute of the edge connecting node i and node j in the graph structure data;

[0019] S24. Graph Structure Composition: Based on the state probability distribution vector P generated in step S22 and the Markov state transition probability matrix W generated in step S23, construct a graph structure G=(V, E) with B nodes; wherein, each node v in the node set V of the graph structure G... i For a discrete state i, its node attributes are assigned the corresponding probability value p in P. i Each edge e in the edge set E of the graph structure G ij If there is a connection with a non-zero transition probability from state i to state j, the edge attribute is assigned the corresponding transition probability value w in W. ij .

[0020] In some embodiments, the state quantization step involves quantizing the one-dimensional temporal vibration data sequence in the following manner. Each data point x in t Mapping to state q t ∈ {1, 2, ..., B}:

[0021] First, determine the robust maximum value used for state partitioning. and robust minimum

[0022]

[0023] Where, x nmax and x nmin These are the maximum and minimum values ​​of the fitted normal distribution obtained after fitting the time series data sequence X to a normal distribution;

[0024] Then, based on the robust maximum value... and robust minimum For data point x t Quantify;

[0025]

[0026] Where B is the pre-defined total number of states, k is an integer in the interval [1, B], and q is the quantization result. t That is, the t-th element in the state sequence Q.

[0027] In some embodiments, in the node attribute generation step, the state probability distribution vector P = { ..., p_B} are calculated in the following way:

[0028] Traverse the state sequence Count the number of times each state value b ∈ {1, 2, ..., B} appears in the sequence Q, denoted as N. b ;

[0029] Calculate the probability p of state b occurring using the following formula. b , so as to form the state probability distribution vector P;

[0030]

[0031] Where n is the length of the one-dimensional time-series vibration data sequence X.

[0032] In some embodiments, in the edge attribute generation step, the Markov state transition probability matrix W is calculated in the following manner:

[0033] Traverse the state sequence Count the frequency of transitions from state i to state j, denoted as N. ij , where i, j ∈ {1, 2, ..., B};

[0034]

[0035] Count the total number of times state i appears in the state sequence Q (excluding the last element). ), denoted as N i ;

[0036] The transition probability w from state i to state j is calculated using the following formula. ij :

[0037]

[0038] Here, the element in the i-th row and j-th column of matrix W is w. ij .

[0039] In some embodiments, the neural network model based on the graph attention mechanism is a Graphormer model;

[0040] The Graphormer model is configured to perform attention-based feature transformation on the input graph structure data. It includes a node importance encoding module, a spatial encoding module, an edge encoding module, and a virtual node to implement different encoding functions.

[0041] In some embodiments, the node importance encoding module encodes the input node attributes in the following manner:

[0042] For each node v in the graph structure G=(V, E) i∈ V, obtain its in-degree and out-degree ;

[0043] According to the in-degree From a learnable in-degree embedding vector table Index the corresponding vector ;

[0044] According to the out-degree From a learnable out-degree embedding vector table Index the corresponding vector ;

[0045] The in-degree embedding vector and out-degree embedding vector obtained from the index are compared with the input original node attribute vector x. i Adding them together yields the encoded generalized node feature vector. The calculation is as follows:

[0046]

[0047] in, and These are all parameters that can be learned during model training.

[0048] In some embodiments, the spatial encoding module encodes the structural relationships between nodes in the attention mechanism in the following manner:

[0049] For each pair of nodes (v) in the graph structure G i , v j ), calculate the shortest path distance between them, denoted as . If there is no path between the two nodes, then... Set to the preset disconnection flag value;

[0050] The shortest path distance As an index, the corresponding spatial bias scalar is looked up from a learnable scalar bias table b. ;

[0051] At computing node v i With v j When considering attention scores, the spatial bias scalar is used. The attention score A is introduced as an additive term. ij The calculation method is as follows:

[0052]

[0053] Among them, h i and h jThey are nodes v i and v j Feature vectors (for the first layer of the network, h) i The generalized node features defined in the foregoing embodiments ), W Q and W K Let be the learnable query transformation matrix and key transformation matrix, and d be the dimension of the feature vector. This is the scaling factor.

[0054] In some embodiments, the edge encoding module encodes the edge attribute information of the graph into the attention mechanism in the following manner:

[0055] For each pair of nodes (v) in the graph structure G i , v j Determine the shortest path P between them. ij ;

[0056] Along the shortest path P ij For each edge e on the path n Take its edge attribute features , with a corresponding learnable edge embedding vector Perform the dot product operation;

[0057] Calculate the average of these dot product results to obtain the edge-encoded node pair (v i , v j The contribution value c) ij The calculation formula is as follows:

[0058]

[0059] Where N is the shortest path P ij The number of sides on a vector indicates the dot product operation.

[0060] The contribution value c ij As an additive term, it is added to the attention score that already contains spatial encoding, thus forming the final attention score that integrates node features, structural relationships, and edge attributes. ;

[0061] .

[0062] In some embodiments, the one-dimensional temporal vibration training dataset is constructed in the following manner:

[0063] S11. In an environment including a harmonic reducer fault diagnosis test bench, obtain vibration signals of the harmonic reducer under various preset fault states and various different load-speed combinations.

[0064] S12. The collected vibration signal is segmented to form multiple one-dimensional time-series vibration data samples with fixed lengths;

[0065] S13. Randomly divide all samples into a training set, a validation set, and a test set; wherein the training set is used for training in the model building phase, and the test set is used for performance evaluation in the fault diagnosis phase.

[0066] The present invention has the following beneficial effects:

[0067] This invention divides the fault diagnosis process of harmonic reducers into two stages: "model building" and "fault diagnosis." It achieves a complete, automated closed-loop diagnostic process from data preparation and feature learning to online application, solving the fundamental problems of existing methods that rely on manual feature extraction and have weak generalization ability. Specifically: First, in the model building stage, a high-precision diagnostic model is built entirely automatically by collecting known fault data, automatically extracting graph structure features using the MFGI method, and then training it using a graph attention network. This avoids the cumbersome process of mechanism modeling, feature design, and time-frequency transformation that heavily relies on expert experience in traditional methods, improving the construction efficiency and intelligence level of the diagnostic system. Second, in the fault diagnosis stage, the trained model is used to perform one-click diagnosis of new data, achieving real-time, rapid, and high-precision fault identification and classification in industrial settings. This makes the solution highly practical for engineering applications and provides a reliable technical tool for predictive maintenance of industrial robots. Attached Figure Description

[0068] Figure 1 A flowchart illustrating an intelligent diagnostic method for harmonic reducers provided in an embodiment of the present invention;

[0069] Figure 2 A schematic diagram of the quantization process provided in an embodiment of the present invention;

[0070] Figure 3 This is a schematic diagram illustrating the process of calculating the state distribution probability vector according to an embodiment of the present invention;

[0071] Figure 4 This is a schematic diagram illustrating the Markov transition matrix calculation process provided in an embodiment of the present invention;

[0072] Figure 5 This is a schematic diagram of the Markov graph information construction process provided in an embodiment of the present invention;

[0073] Figure 6 This is a schematic diagram of the Graphormer network structure provided in an embodiment of the present invention;

[0074] Figure 7 This is a schematic diagram of node importance encoding provided in an embodiment of the present invention;

[0075] Figure 8 This is a schematic diagram of the spatial coding calculation process provided in an embodiment of the present invention;

[0076] Figure 9 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0077] To enable those skilled in the art to better understand the technical solutions of the present invention, exemplary embodiments of the present invention are described below in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0078] Where there is no conflict, the various embodiments of the present invention and the features thereof may be combined with each other.

[0079] As used herein, the term “and / or” includes any and all combinations of one or more related enumerated entries.

[0080] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used herein, the singular forms “a” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that when the terms “comprising” and / or “made of” are used in this specification, the presence of the stated feature, integral, step, operation, element, and / or component is specified, but the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof is not excluded. Terms such as “connected” or “linked” are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect.

[0081] Unless otherwise specified, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art. It will also be understood that terms such as those defined in commonly used dictionaries should be interpreted as having the meaning consistent with their meaning in the context of the relevant art and the invention, and will not be interpreted as having an idealized or overly formal meaning unless expressly so defined herein.

[0082] In the technical solution of this invention, the collection, storage, use, processing, transmission, provision, and disclosure of user personal information all comply with relevant laws and regulations and do not violate public order and good morals. The use of user data in this technical solution follows relevant national laws and regulations (e.g., the "Information Security Technology - Personal Information Security Specification"). For example: appropriate measures are taken for personal information access control; restrictions are imposed on the display of personal information; the purpose of using personal information does not exceed the scope of direct or reasonable association; and explicit identity targeting is eliminated when using personal information to avoid precisely locating a specific individual.

[0083] To address at least one of the technical problems existing in the aforementioned related technologies, the present invention provides an intelligent diagnostic method for harmonic reducers. Figure 1 This is a flowchart illustrating an intelligent diagnostic method for harmonic reducers provided in an embodiment of the present invention. The method includes a model building stage and a fault diagnosis stage.

[0084] The model building phase includes the following steps:

[0085] S1. Obtain the vibration signals of the harmonic reducer operating under various known fault conditions and construct a one-dimensional time-series vibration training dataset;

[0086] S2. Using the improved Markov graph information method, each time-series sample in the one-dimensional time-series vibration training dataset is converted into corresponding graph structure data, wherein the graph structure data represents the data distribution with state nodes and the state transition probability is represented by the edges between nodes.

[0087] S3. Use the graph structure data to train a neural network model based on graph attention mechanism to obtain a fault diagnosis model that can identify fault categories based on graph structure data;

[0088] The fault diagnosis phase includes the following steps:

[0089] S4. Collect the vibration signal of the harmonic reducer to be diagnosed and obtain one-dimensional time-series vibration test data;

[0090] S5. Using the improved Markov graph information method, the one-dimensional time-series vibration test data is converted into test graph structure data;

[0091] S6. Input the test diagram structure data into the trained fault diagnosis model and output the fault classification result of the harmonic reducer to be diagnosed.

[0092] This invention divides the fault diagnosis process of harmonic reducers into two stages: "model building" and "fault diagnosis." It achieves a complete, automated closed-loop diagnostic process from data preparation and feature learning to online application, solving the fundamental problems of existing methods that rely on manual feature extraction and have weak generalization ability. Specifically: First, in the model building stage, a high-precision diagnostic model is built entirely automatically by collecting known fault data, automatically extracting graph structure features using the MFGI method, and then training it using a graph attention network. This avoids the cumbersome process of mechanism modeling, feature design, and time-frequency transformation that heavily relies on expert experience in traditional methods, improving the construction efficiency and intelligence level of the diagnostic system. Second, in the fault diagnosis stage, the trained model is used to perform one-click diagnosis of new data, achieving real-time, rapid, and high-precision fault identification and classification in industrial settings. This makes the solution highly practical for engineering applications and provides a reliable technical tool for predictive maintenance of industrial robots.

[0093] The following are specific examples of data collection and dataset creation:

[0094] In some embodiments, the specific implementation steps of the present invention for obtaining the vibration signal of a harmonic reducer operating under various known fault conditions are as follows:

[0095] (I) Test bench design

[0096] A fault diagnosis test bench for harmonic reducers was constructed. The test bench mainly consists of a transmission unit, a sensor unit, a test object, a data acquisition unit, and a load unit. The transmission unit provides power to drive the harmonic reducer and provides conditions for constant and variable speed operation during the test. The sensor unit, as the core data acquisition module, acquires dynamic performance parameters and operating status data in real time, forming the sole basis for subsequent analysis and evaluation. The sensor unit includes vibration sensors, noise sensors, and temperature sensors, used to measure vibration, noise, and temperature information during operation, respectively. The load unit consists of a magnetohydrodynamic brake, used to apply torque to the harmonic reducer to simulate a load.

[0097] (II) Data Acquisition Scheme Design

[0098] Based on actual needs, select the fault states of the harmonic reducer to be diagnosed, and simulate these fault states by artificially damaging normal harmonic reducers to create faulty components for experimentation. The same fault type can also be divided into different experimental groups according to the degree of fault. This example involves artificially damaging four normal harmonic reducers to simulate four common fault types: wave generator bearing breakage, input shaft misalignment, flexspline tooth tip pitting, and flexspline tooth root pitting.

[0099] This example designs 20 different combinations under different loads and speeds using the controlled variable method. The motor speeds are divided into four levels: 500rpm, 1000rpm, 1500rpm, and 2000rpm. The loads are divided into five levels: 0Nm (no load), 20Nm, 40Nm, 60Nm, and 80N. The specific operating conditions are shown in Table 1, where S5L0 represents the test with a speed of 500rpm and a load of 0N.

[0100] This experiment tested various operating conditions under different fault states. Taking the fracture of the wave generator bearing as an example, the test steps are explained as follows:

[0101] (1) First, install the harmonic reducer with broken wave generator bearing onto the test bench and strictly ensure that the reducer is tightly fixed in the corresponding position. Then, debug each subsystem of the entire test platform and ensure normal operation before performing overall system debugging to ensure test safety and stable status.

[0102] (2) Adjust the motor speed to 500 rpm and adjust the load to 0 Nm, 20 Nm, 40 Nm, 60 Nm and 80 Nm respectively. Test the working conditions S5L0, S5L2, S5L4, S5L6 and S5L8. After ensuring the motor is stable, start recording data. Repeat the test 5 times for each working condition.

[0103] (3) Adjust the motor speed to 1000rpm, 1500rpm, and 2000rpm respectively, repeat step (2), and record the various operating conditions at different speeds. After completing the wave generator bearing fracture test, organize the test data and prepare for the next test of different operating conditions for the fault type.

[0104] The selection of actual operating conditions can be determined based on the actual working conditions of the harmonic reducer. If necessary, variable operating condition data can also be collected according to the above steps for model training.

[0105] Table 1 is a detailed table of a fixed-condition test provided by an embodiment of the present invention.

[0106]

[0107] Table 1

[0108] (III) Dataset Creation

[0109] In this example, the data acquisition frequency was set to 20480Hz, with each sampling lasting 24.2 seconds, resulting in a time-series vibration data set of 495616 data points. The vibration data obtained from each sampling was divided into 33 smaller samples of 15000 data points each. Each working condition was sampled ten times, resulting in 330 samples for each condition. The dataset contained a total of 1550 samples. The resulting dataset was then randomly shuffled and divided into training, validation, and test sets in a 7:2:1 ratio.

[0110] The following is a specific implementation example of the algorithm:

[0111] In some embodiments, the improved Markov graph information method specifically includes the following steps:

[0112] S21. State quantization: quantizing a one-dimensional temporal vibration data sequence of length n. Each data point x in t Quantization is mapped to one of B discrete states to obtain a state sequence. , where q t ∈ {1, 2, ...,B}, where t is the time sequence index;

[0113] S22. Node Attribute Generation: Based on the state sequence Q, count the frequency of each state and calculate the probability of each state, generating a state probability distribution vector P = { with dimension B. ..., p B}, representing the attribute set of all nodes in the graph structure data, where p b This represents the probability of state b occurring;

[0114] S23. Edge attribute generation: Based on the state sequence Q, count the transition frequency between states and calculate the transition probability w from state i to state j. ij Generate a Markov state transition probability matrix W of dimension B×B, and set each non-zero element w in matrix W to... ij As an attribute of the edge connecting node i and node j in the graph structure data;

[0115] S24. Graph Structure Composition: Based on the state probability distribution vector P generated in step S22 and the Markov state transition probability matrix W generated in step S23, construct a graph structure G=(V, E) with B nodes; wherein, each node v in the node set V of the graph structure G... i For a discrete state i, its node attributes are assigned the corresponding probability value p in P. i Each edge e in the edge set E of the graph structure G ijIf there is a connection with a non-zero transition probability from state i to state j, the edge attribute is assigned the corresponding transition probability value w in W. ij .

[0116] In the above embodiments, through state quantization, node / edge attribute generation, and graph structure synthesis, one-dimensional vibration time-series signals are systematically converted into graph structures that preserve the inherent statistical regularities of the sequence, providing accurate and structured non-Euclidean data input for subsequent graph neural network processing. Specifically, this method creatively encodes the statistical characteristics (state distribution probability) and dynamic evolution laws (state transition probability) of time-series data into node attributes and edge attributes of the graph, respectively. This transforms the complex time-series pattern recognition problem into a relation learning problem on a graph structure, fundamentally overcoming the shortcomings of traditional CNN / LSTM methods in processing long sequences, such as gradient vanishing / exploding and difficulty in capturing long-range dependencies. It also avoids the problems of a large amount of invalid information, low computational efficiency, and reliance on expert selection of transformation methods when converting time series into two-dimensional time-frequency graphs.

[0117] In some embodiments, the state quantization step involves quantizing the one-dimensional temporal vibration data sequence in the following manner. Each data point x in t Mapping to state q t ∈ {1, 2, ..., B}:

[0118] Given a time series Where n is the sequence length, in order to accurately describe the data while improving storage and retrieval efficiency and eliminating some noise disturbances, the entire time series sequence is divided into B states, and each element in the sequence is represented by the following formula. Each one will be uniquely mapped to a state. :

[0119] (1)

[0120] in and Let be the maximum and minimum values ​​in the sequence, respectively. Taking a time series of length 50 as an example to illustrate the quantization process, the maximum and minimum values ​​in the sequence are 0.130 and -0.135, respectively. Choosing B equal to 10, traversing the sequence and substituting into (1), the result is as follows: Figure 2 As shown.

[0121] However, the extreme values ​​of vibration signals are unstable and easily affected by external interference. Using them as the standard for quantization timing will change the essential characteristics of the timing sequence and increase the feature differences between samples of the same fault type. To solve these problems, we use the normal distribution extreme value quantization sequence to ensure that the encoding is not affected by extreme values. At the same time, since the vibration signals of rotating machinery are generally symmetrical, directly dividing the data segments by the maximum value will lead to an eccentric state distribution. Therefore, we compare the absolute values ​​of the maximum and minimum values ​​of the normal distribution and take the smaller value to describe the state of the vibration signal. Specifically, we use (2) and (3) instead of (1) to quantize the timing signal.

[0122] (2)

[0123] (3)

[0124] in and These are the maximum and minimum values ​​of the normal distribution of time series data.

[0125] In the above embodiments, robust extrema are obtained by using a method that fits the data to a symmetric normal distribution. As a boundary for state partitioning, replacing the original sequence extrema which are susceptible to noise and outliers, this significantly improves the stability and anti-interference capability of the state quantization process, thereby enhancing the consistency of graph structural features under the same fault type. Specifically, since instantaneous impacts and sampling errors in vibration signals can generate abnormal maxima or minima, directly using these as quantization boundaries would lead to drastically different state partitioning results for samples under the same fault state due to accidental noise, resulting in huge differences in the generated graph structural features and severely impairing the recognition accuracy and robustness of subsequent models. This scheme obtains a robust boundary that reflects the distribution characteristics of the main data subject through normal distribution fitting, effectively filtering out the interference of outliers and ensuring that the core vibration mode is stably and consistently encoded into discrete states. This improvement makes the graph structures converted from similar fault samples have higher similarity, fundamentally improving the robustness of the entire diagnostic method to noise and its generalization ability to different samples.

[0126] In some embodiments, in the node attribute generation step, the state probability distribution vector P = { ..., p_B} are calculated in the following way:

[0127] Traverse the state sequence Count the number of times each state value b ∈ {1, 2, ..., B} appears in the sequence Q, denoted as N. b ;

[0128] Calculate the probability p of state b occurring using the following formula.b , so as to form the state probability distribution vector P;

[0129] (4)

[0130] Where, p b This represents the probability that state b occurs in the state sequence Q. The number of times state b appears in the state sequence Q is represented by n, where n represents the length of the one-dimensional temporal vibration data sequence X. Figure 3 The example illustrates how to obtain the state distribution probability vector, where state 4 appears the most frequently in the state sequence, appearing a total of 9 times, and its corresponding state distribution probability component is 0.18.

[0131] In the above embodiments, the state probability distribution vector P is accurately calculated by statistically analyzing the frequency of each state in the state sequence. This provides each node in the graph with a core attribute reflecting the overall statistical characteristics of the data distribution. This attribute is a key input for subsequent node-level feature learning and fusion in the graph attention network. Specifically, the probability p of the state occurrence is... b As a node attribute, it provides graph neural networks with a quantitative measure of the "importance" or "activity" of each state (node) within the entire observation window. This statistically based probabilistic information has stronger representational power and stability than the original vibration amplitude itself. In networks such as Graphormer, modules such as node importance encoding can more effectively learn the relative weights of different state nodes in fault representation based on this probabilistic attribute, thereby guiding the network's attention to focus more on key state patterns related to the fault, improving the targeting of feature extraction and the accuracy of diagnosis.

[0132] In some embodiments, in the edge attribute generation step, the Markov state transition probability matrix W is calculated in the following manner:

[0133] The state is defined by time t. up to time t+1 For a state To transition to state j, traverse the entire state sequence sequentially to obtain n-1 state transitions, and count all transitions from state j. to state The number of transfers is denoted as The number of times the state appears, as counted in (4) This allows us to calculate the state. to state transition probability :

[0134] (5)

[0135] in, All Combining these elements yields the Markov transition matrix W:

[0136] (6)

[0137] The Markov transition matrix represents the transition probabilities between states, indicating the likelihood that the current state will become a different state in the next time step. Figure 2 The example illustrates the calculation process. State 2 appears three times in the state sequence, forming three transitions: from state 2 to state 6, from state 2 to state 4, and from state 2 to state 7. Substituting these into (5), we can calculate... Repeating the above steps will yield the Markov transition matrix. The specific calculation process is as follows: Figure 4 As shown.

[0138] In the above embodiments, by accurately counting the transition frequencies between states and calculating the transition probabilities, a Markov state transition probability matrix W is constructed. This encodes the dynamic evolution characteristics and temporal dependencies of the vibration signal as connecting edges between nodes in the graph and their weights, thus fully characterizing the dynamic behavior of system state changes. Specifically, the edge attribute w... ij The probability of transitioning from state i to state j quantitatively describes the transition patterns between different states during system operation. These patterns are often closely related to specific fault modes (such as periodic shocks and modulation phenomena). By using this transition probability matrix as an edge attribute of the graph, the complex temporal causal relationships and nonlinear dynamic characteristics implicit in one-dimensional time-series signals are successfully transformed into visible and computable relational information in a graph structure. This provides graph attention networks (especially their edge encoding modules) with the possibility of directly learning fault-related state transition patterns, enabling the model to distinguish different faults from the perspective of "how states evolve" rather than just "how states are distributed," greatly enriching the information dimension of fault features. This is crucial for identifying fault types that are similar in steady-state distribution but significantly different in dynamic processes.

[0139] In some embodiments, in the graph structure G , This represents the attribute of the i-th node in the graph. Representing the edge attributes between connecting nodes, MFGI transforms temporal information into a state distribution probability vector P of dimension B and a state distribution probability vector P of dimension B. The state transition probability matrix W, where each element in P... As an attribute of node i, each non-zero element in W As an attribute of the edge connecting nodes i and j, a graph consisting of B nodes can be constructed. The specific steps are as follows: Figure 5 As shown.

[0140] In some embodiments, the neural network model based on the graph attention mechanism is a Graphormer model;

[0141] GraphMer is a graph neural network based on the Transformer architecture, and its specific structure is as follows: Figure 5 As shown, this method encodes node importance and distribution features and directly adds them to node attributes to form new generalized node features, which are then input into the self-attention module to calculate the similarity matrix. Spatial encoding and edge encoding are directly added to the similarity matrix to form a new generalized similarity matrix, used to update the node features of the next layer. The Graphormer network mainly includes a node importance encoding module, a spatial encoding module, an edge encoding module, and virtual nodes, such as... Figure 6 As shown, this is the Graphormer network structure of the present invention.

[0142] The above embodiments explicitly designate the neural network model based on the graph attention mechanism as the Graphormer model. Through its unique architectural design, this model can achieve multi-level, global deep feature extraction and fusion of graph structure data generated by MFGI, thereby significantly improving the representation ability of complex graph structure features and the accuracy of fault classification. Specifically, compared with ordinary graph neural networks (GNNs), the four core components of the Graphormer model (node ​​importance encoding, spatial encoding, edge encoding, and virtual nodes) form a synergistic advantage. It not only aggregates neighbor information locally but also overcomes the inherent defects of traditional GNNs in terms of limited receptive field, difficulty in modeling long-range node dependencies, and effective utilization of edge information by introducing global attention and structured encoding.

[0143] In some embodiments, node attributes can be directly substituted to achieve node information fusion based on feature similarity between nodes. However, this method only utilizes node attributes and node connections, essentially wasting information. In fact, the importance of different nodes in a graph structure is not consistent; for example, celebrities have a much higher weight in social networks than ordinary people. In Graphormer, node centrality is used as a criterion for measuring node importance. Two embedding vectors are assigned to each node based on its in-degree and out-degree, and these are added as inputs to the node attributes.

[0144] Specifically:

[0145] For each node v in the graph structure G=(V, E) i ∈ V, obtain its in-degree and out-degree ;

[0146] According to the in-degree From a learnable in-degree embedding vector table Index the corresponding vector ;

[0147] According to the out-degree From a learnable out-degree embedding vector table Index the corresponding vector ;

[0148] The in-degree embedding vector and out-degree embedding vector obtained from the index are compared with the input original node attribute vector x. i Adding them together yields the encoded generalized node feature vector. The calculation is as follows:

[0149] (7)

[0150] in, and These are all parameters that can be learned during model training.

[0151] See Figure 7 The diagram shown is a schematic diagram of node importance encoding provided in an embodiment of the present invention.

[0152] The above embodiments enhance the encoding of node attributes by using the in-degree and out-degree information of nodes, assigning differentiated weights to each node in the graph that reflect its importance in the topological structure. This allows the model to automatically focus on more critical and information-rich state nodes in the network during subsequent attention calculations. Specifically, in the graph generated by MFGI, the "importance" of different state nodes is not the same. This scheme maps the in-degree and out-degree of a node (i.e., the frequency to which and out of the state) into learnable embedding vectors and adds them to the original node attributes. This directly injects the structural centrality information of the graph into the node features, giving higher initial weights to nodes that are in pivotal positions in the state transition network (involved in high-frequency transitions). This prior weight guidance helps the Graphormer attention mechanism learn the core state patterns most relevant to the fault dynamics more quickly and accurately, avoiding slow convergence or getting stuck in suboptimal solutions that might result from treating all nodes equally in the early stages of training, thereby improving training efficiency and the targeting of feature learning.

[0153] In some embodiments, the Transformer architecture exhibits significant advantages in processing sequence data due to its global awareness field—tokens at each level can dynamically capture association information at any position in the entire sequence. However, this global interaction mechanism also introduces inherent challenges to position-dependent modeling: the model needs to explicitly encode the spatial relationships between elements to compensate for the lack of sequence order. In traditional sequence modeling scenarios, researchers have addressed this issue by using absolute position embedding (assigning an independent vector to each position) or relative position encoding (compiling the association function of the distance between elements). However, when processing graph-structured data, the multidimensional spatial distribution of nodes and the non-sequence topology characteristics render the above methods ineffective, necessitating the development of new spatial encoding paradigms.

[0154] Specifically, consider a function that measures the spatial relationship between arbitrary nodes i and j. The function can be defined by the connectivity between nodes in the graph. When two nodes are connected, the distance (SPD) of the shortest path between them is chosen as the threshold. If not, then Let a special value be set: -1. Each pair of nodes will be assigned a learnable scalar, which will be added as a bias term in the self-attention module to the corresponding element of the similarity matrix:

[0155] (8)

[0156] Among them, h i and h j They are nodes v i and v j Feature vectors (for the first layer of the network, h) i The generalized node features defined in the foregoing embodiments ), W Q and W K Let be the learnable query transformation matrix and key transformation matrix, and d be the dimension of the feature vector. This is the scaling factor.

[0157] Where b is a group consisting of... The index is a learnable scalar, shared across all layers, and its computation process is as follows: Figure 8 As shown.

[0158] The above embodiments explicitly inject global topological information of the graph into the attention mechanism by encoding the shortest path distance (SPD) between node pairs in the graph as a learnable bias term in the attention score. This enables the model to perceive and utilize the structured spatial relationships between nodes, enhancing the model's understanding of the overall graph structure. Specifically, this scheme calculates the SPD and uses it as a distance-related bias. Adding to the attention score creatively endows the attention mechanism with "structure awareness," which allows the model to consider the distance between nodes in the topological space when calculating the correlation between nodes. For example, it can learn to generate stronger attention between nodes that are close to each other in the topology, or adjust the intensity of information transmission according to distance. This ability is crucial for understanding graphs composed of state transitions, as it helps the model distinguish between locally closely related clusters of states and globally distant state dependencies, thereby enabling more accurate modeling of the propagation path and scope of impact of faults.

[0159] In some embodiments, traditional methods for effectively encoding edge features in graph-structured data suffer from inherent limitations in information transmission: existing schemes either rely on node feature overlay (directly injecting edge features into the feature space of associated nodes) or feature coupling during neighborhood aggregation (synchronously integrating edge and node features into local information). Both paradigms are limited to indirectly transmitting edge information to adjacent nodes, making it difficult to fully realize the contribution of edge features to the global graph representation. Therefore, it is necessary to develop an algorithm that can encode edge features into global information. Specifically, for each pair of nodes (v... i , v j The average value of the dot product of the edge features and their corresponding learnable embedding vectors along the shortest path is calculated and then directly added to the similarity matrix. The specific calculation is as follows:

[0160] For each pair of nodes (v) in the graph structure G i , v j Determine the shortest path P between them. ij ;

[0161] Along the shortest path P ij For each edge e on the path n Take its edge attribute features , with a corresponding learnable edge embedding vector Perform the dot product operation;

[0162] Calculate the average of these dot product results to obtain the edge-encoded node pair (v i , v j The contribution value c) ij The calculation formula is as follows:

[0163]

[0164] Where N is the shortest path P ij The number of sides on a vector indicates the dot product operation.

[0165] The contribution value c ijAs an additive term, it is added to the attention score that already contains spatial encoding, thus forming the final attention score that integrates node features, structural relationships, and edge attributes. ;

[0166] .

[0167] The above embodiments aggregate edge attribute (state transition probability) information along the shortest path between nodes and incorporate it as a contribution value into the attention score. This achieves the direct and effective integration of key edge features (reflecting fault dynamics) into global graph representation learning and node relationship modeling, overcoming the difficulty of efficiently utilizing edge information in traditional methods. Specifically, in the graph generated by MFGI, edge attribute w ij Transition probabilities are one of the core features for fault diagnosis. Traditional GNNs often utilize edge information locally through message passing, making global sharing difficult. This scheme innovatively uses the "shortest path" as a medium to encode the transition probability features of all edges on the path connecting a pair of nodes, and assigns this encoded value c. ij Adding this as a bias term to the attention mechanism means that the attention strength between a pair of nodes depends not only on their own features and topological distance, but also directly on the state transition probability patterns on the paths connecting them. This greatly enriches the information dimensions considered by the attention mechanism, enabling the model to keenly capture the anomalous state transition sequence patterns corresponding to specific faults. Consequently, it has a stronger discriminative power when distinguishing fault types, especially between faults with similar dynamic behaviors but different transition probability distributions.

[0168] In some embodiments, the virtual node is specifically:

[0169] In a network, a virtual node is set up, assuming it is directly connected to all other nodes. However, this connection is completely virtual and independent, and it does not act as an intermediary for normal node connections. Since its distance to all nodes is 1, a special design is needed to distinguish it from real physical nodes. The function implements special spatial encoding by using a special Readout method, which uses the output of the virtual node in the last layer of the network as the final output of the entire network.

[0170] The algorithm implemented in the above embodiments can first transform the one-dimensional vibration dataset of the harmonic reducer obtained in the dataset production into a graph structure dataset, and input it into the Graphormer network. The network will automatically extract features and output one-hot vectors. The hyperparameters and network parameters in the model training process are shown in Table 2 and Table 3, respectively.

[0171]

[0172] Table 2 Training Parameters

[0173]

[0174] Table 3 Network Parameters

[0175] Based on the same inventive concept, embodiments of the present invention also provide an electronic device. Figure 9 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Figure 9 As shown, an embodiment of the present invention provides an electronic device including: one or more processors 101, a memory 102, and one or more I / O interfaces 103. The memory 102 stores one or more programs, which, when executed by the one or more processors, enable the one or more processors to implement any of the harmonic reducer intelligent diagnostic methods described in the above embodiments; the one or more I / O interfaces 103 are connected between the processor and the memory, configured to enable information interaction between the processor and the memory.

[0176] The processor 101 is a device with data processing capabilities, including but not limited to a central processing unit (CPU); the memory 102 is a device with data storage capabilities, including but not limited to random access memory (RAM, more specifically SDRAM, DDR, etc.), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), and flash memory (FLASH); the I / O interface (read / write interface) 103 is connected between the processor 101 and the memory 102, and can realize information interaction between the processor 101 and the memory 102, including but not limited to a data bus (Bus).

[0177] In some embodiments, the processor 101, memory 102, and I / O interface 103 are interconnected via bus 104, and thus connected to other components of the computing device.

[0178] In some embodiments, the one or more processors 101 include a field-programmable gate array.

[0179] This invention also provides a computer-readable medium. The computer-readable medium stores a computer program, which, when executed by a processor, implements the steps of any of the harmonic reducer intelligent diagnostic methods described in the above embodiments. The computer-readable storage medium can be volatile or non-volatile.

[0180] This invention also provides a computer program product, including computer-readable code, or a non-volatile computer-readable storage medium carrying computer-readable code. When the computer-readable code is run in the processor of an electronic device, the processor in the electronic device executes any of the above-described intelligent diagnostic methods for harmonic reducers.

[0181] Those skilled in the art will understand that all or some of the steps, systems, and apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software can be distributed on a computer-readable storage medium, which may include computer storage media (or non-transitory media) and communication media (or transient media).

[0182] As is known to those skilled in the art, computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information such as computer-readable program instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), static random access memory (SRAM), flash memory or other memory technologies, portable compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, it is known to those skilled in the art that communication media typically contain computer-readable program instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0183] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0184] The computer program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the invention.

[0185] The computer program product described herein can be implemented specifically through hardware, software, or a combination thereof. In one alternative embodiment, the computer program product is specifically embodied in a computer storage medium; in another alternative embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0186] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0187] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0188] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0189] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction, which contains one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0190] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A method for intelligent diagnosis of harmonic reducers, characterized in that, This includes the model building phase and the fault diagnosis phase; The model building phase includes the following steps: S1. Obtain the vibration signals of the harmonic reducer operating under various known fault conditions and construct a one-dimensional time-series vibration training dataset; S2. Using the improved Markov graph information method, each time-series sample in the one-dimensional time-series vibration training dataset is converted into corresponding graph structure data, wherein the graph structure data represents the data distribution with state nodes and the state transition probability is represented by the edges between nodes. S3. Use the graph structure data to train a neural network model based on graph attention mechanism to obtain a fault diagnosis model that can identify fault categories based on graph structure data; The fault diagnosis phase includes the following steps: S4. Collect the vibration signal of the harmonic reducer to be diagnosed and obtain one-dimensional time-series vibration test data; S5. Using the improved Markov graph information method, the one-dimensional time-series vibration test data is converted into test graph structure data; S6. Input the test diagram structure data into the trained fault diagnosis model and output the fault classification result of the harmonic reducer to be diagnosed.

2. The intelligent diagnostic method for harmonic reducers according to claim 1, characterized in that, The improved Markov graph information method specifically includes the following steps: S21. State quantization: Quantizing a one-dimensional temporal vibration data sequence X of length n... Each data point x in t Quantization is mapped to one of B discrete states to obtain a state sequence. , where q t ∈ {1, 2, ..., B}, where t is the time sequence index; S22. Node Attribute Generation: Based on the state sequence Q, count the frequency of each state and calculate the probability of each state, generating a state probability distribution vector P = { with dimension B. ..., p B }, representing the attribute set of all nodes in the graph structure data, where p b This represents the probability of state b occurring; S23. Edge attribute generation: Based on the state sequence Q, count the transition frequency between states and calculate the transition probability w from state i to state j. ij Generate a Markov state transition probability matrix W of dimension B×B, and set each non-zero element w in matrix W to... ij As an attribute of the edge connecting node i and node j in the graph structure data; S24. Graph Structure Composition: Based on the state probability distribution vector P generated in step S22 and the Markov state transition probability matrix W generated in step S23, construct a graph structure G=(V, E) with B nodes; wherein, each node v in the node set V of the graph structure G... i For a discrete state i, its node attributes are assigned the corresponding probability value p in P. i Each edge e in the edge set E of the graph structure G ij If there is a connection with a non-zero transition probability from state i to state j, the edge attribute is assigned the corresponding transition probability value w in W. ij .

3. The intelligent diagnostic method for harmonic reducers according to claim 2, characterized in that, In the state quantization step, the one-dimensional time-series vibration data sequence X = is quantized in the following manner. Each data point x in t Mapping to state q t ∈ {1, 2, ..., B}: First, determine the robust maximum value used for state partitioning. and robust minimum ; ； Where, x nmax and x nmin These are the maximum and minimum values ​​of the fitted normal distribution obtained after fitting the time series data sequence X to a normal distribution; Then, based on the robust maximum value... and robust minimum For data point x t Quantify; ； Where B is the pre-defined total number of states, k is an integer in the interval [1, B], and q is the quantization result. t That is, the t-th element in the state sequence Q.

4. The intelligent diagnostic method for harmonic reducers according to claim 2, characterized in that, In the node attribute generation step, the state probability distribution vector It is calculated in the following way: Traverse the state sequence Count the number of times each state value b ∈ {1, 2, ..., B} appears in the sequence Q, denoted as N. b ; Calculate the probability p of state b occurring using the following formula. b , so as to form the state probability distribution vector P; ； Where n is the length of the one-dimensional time-series vibration data sequence X.

5. The intelligent diagnostic method for harmonic reducers according to claim 2, characterized in that, In the edge attribute generation step, the Markov state transition probability matrix W is calculated in the following way: Traverse the state sequence Count the frequency of transitions from state i to state j, denoted as N. ij , where i, j∈ {1, 2, ..., B}; ； Let N be the total number of times state i appears in the state sequence Q. i ; The transition probability w from state i to state j is calculated using the following formula. ij : ； Here, the element in the i-th row and j-th column of matrix W is w. ij .

6. The intelligent diagnostic method for harmonic reducers according to claim 1, characterized in that, The neural network model based on the graph attention mechanism is the Graphormer model; The Graphormer model is configured to perform attention-based feature transformation on the input graph structure data. It includes a node importance encoding module, a spatial encoding module, an edge encoding module, and a virtual node to implement different encoding functions.

7. The intelligent diagnostic method for harmonic reducers according to claim 6, characterized in that, The node importance encoding module encodes the input node attributes in the following way: For each node v in the graph structure G=(V, E) i ∈ V, obtain its in-degree and out-degree ; According to the in-degree The corresponding vector is indexed from a learnable in-degree embedding vector table z⁻. ; According to the out-degree The corresponding vector is indexed from a learnable out-degree embedding vector table z⁺. ; The in-degree embedding vector and out-degree embedding vector obtained from the index are compared with the input original node attribute vector x. i Adding them together yields the encoded generalized node feature vector. The calculation is as follows: ； in, and These are all parameters that can be learned during model training.

8. The intelligent diagnostic method for harmonic reducers according to claim 6 or 7, characterized in that, The spatial encoding module encodes the structural relationships between nodes in the attention mechanism in the following way: For each pair of nodes (v) in the graph structure G i , v j ), calculate the shortest path distance between them, denoted as . If there is no path between the two nodes, then... Set to the preset disconnection flag value; The shortest path distance As an index, the corresponding spatial bias scalar is looked up from a learnable scalar bias table b. ; At computing node v i With v j When considering attention scores, the spatial bias scalar is used. The attention score A is introduced as an additive term. ij The calculation method is as follows: ； Among them, h i and h j They are nodes v i and v j Feature vectors (for the first layer of the network, h) i The generalized node features defined in the foregoing embodiments ), W Q and W K Let be the learnable query transformation matrix and key transformation matrix, and d be the dimension of the feature vector. This is the scaling factor.

9. The intelligent diagnostic method for harmonic reducers according to claim 8, characterized in that, The edge encoding module encodes the edge attribute information of the graph into the attention mechanism in the following way: For each pair of nodes (v) in the graph structure G i , v j Determine the shortest path P between them. ij ; Along the shortest path P ij For each edge e on the path n Take its edge attribute features , with a corresponding learnable edge embedding vector Perform the dot product operation; Calculate the average of these dot product results to obtain the edge-encoded node pair (v i , v j The contribution value c) ij The calculation formula is as follows: ； Where N is the shortest path P ij The number of sides on a vector indicates the dot product operation. The contribution value c ij As an additive term, it is added to the attention score that already contains spatial encoding, thus forming the final attention score that integrates node features, structural relationships, and edge attributes. ; 。 10. The intelligent diagnostic method for harmonic reducers according to claim 1, characterized in that, The one-dimensional temporal vibration training dataset is constructed in the following manner: S11. In an environment including a harmonic reducer fault diagnosis test bench, obtain vibration signals of the harmonic reducer under various preset fault states and various different load-speed combinations. S12. The collected vibration signal is segmented to form multiple one-dimensional time-series vibration data samples with fixed lengths; S13. Randomly divide all samples into a training set, a validation set, and a test set; wherein the training set is used for training in the model building phase, and the test set is used for performance evaluation in the fault diagnosis phase.

Images