Intelligent mechanical fault rapid diagnosis method

Abstract

The invention discloses an intelligent mechanical fault rapid diagnosis method. The method comprises the steps of multi-source data acquisition, data optimization processing, mechanical abnormity rapid detection, mechanical fault rapid identification, model performance optimization and mechanical fault intelligent diagnosis. The invention relates to the technical field of mechanical fault data processing, in particular to an intelligent mechanical fault rapid diagnosis method, which innovatively proposes a mechanical fault rapid diagnosis technology for abnormal trigger key sequence extraction, and improves the sensitivity and discrimination precision of mechanical fault features. A random convolution disturbance mechanism is introduced, and a multi-equipment adaptability enhanced loss function is designed, so that rapid, accurate and intelligent diagnosis of mechanical faults under complex working conditions and multi-equipment conditions is realized; an optimization algorithm is improved by adopting a disturbance guide extension strategy, a jump drive fine tuning mechanism and a multi-channel fine tuning candidate strategy, a model performance optimal parameter combination is obtained, and the stability and recognition precision of a mechanical fault recognition model are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical fault data processing, in particular to an intelligent mechanical fault rapid diagnosis method. BACKGROUND

[0002] An intelligent mechanical fault rapid diagnosis method refers to using data acquisition combined with machine learning and pattern recognition and other means to analyze large-scale, multi-source and heterogeneous data generated during the operation of mechanical equipment in real time, so as to realize automatic extraction of fault features, rapid identification and positioning of fault types, and an intelligent diagnosis method that improves the intelligent level and response efficiency of fault diagnosis.

[0003] However, in the traditional mechanical fault diagnosis method, all mechanical operation data are uniformly analyzed, which leads to the technical problems of slow diagnosis response speed, inability to quickly locate faults, and insufficient accuracy of fault identification results, and further results in mechanical fault identification lag and increased misjudgment rate under complex working conditions; the existing mechanical fault identification model has the technical problems of limited fault feature expression capability and insufficient fault identification capability of the model under multi-device conditions, which leads to the decline of fault identification accuracy between different devices; the existing mechanical fault identification model has the technical problem of unreasonable model parameter setting, which leads to unstable fault identification results and insufficient accuracy. SUMMARY

[0004] In view of the above, in order to overcome the defects of the prior art, the application provides an intelligent mechanical fault rapid diagnosis method, which aims at the technical problems that the traditional mechanical fault diagnosis method uniformly analyzes all mechanical operation data, resulting in slow diagnosis response speed, inability to quickly locate faults, and insufficient fault recognition result accuracy, and further causing mechanical fault recognition lag and increasing misjudgment rate under complex working conditions. The present application innovatively proposes an abnormal trigger key sequence extraction mechanical fault rapid diagnosis technology, which first performs mechanical anomaly rapid identification, triggers abnormal timing data extraction of key characteristic fault behaviors only when it is determined to be abnormal, and then performs intelligent fault type identification, effectively reducing the interference of normal data on fault diagnosis, significantly reducing the consumption of computing resources, improving the sensitivity and discrimination accuracy of mechanical fault characteristics, and realizing efficient, accurate and intelligent rapid fault diagnosis of mechanical equipment under complex working conditions. In view of the technical problems that the existing mechanical fault recognition model has limited fault feature expression ability and insufficient model fault recognition ability under multi-device conditions, resulting in a decrease in fault recognition accuracy between different devices, the present application innovatively introduces a random convolution disturbance mechanism and designs a multi-device adaptability enhanced loss function, enhances the deep expression ability of the model to key fault features and the consistency discrimination ability across devices, effectively improves the extraction accuracy of fault features and the adaptability under multi-device conditions, and realizes rapid, accurate and intelligent diagnosis of mechanical faults under complex working conditions and multi-device conditions. In view of the technical problems that the existing mechanical fault recognition model has unreasonable model parameter setting, resulting in unstable and insufficient fault recognition results, the present application adopts a disturbance guided expansion strategy, a jump driven fine tuning mechanism and a multi-channel fine tuning candidate strategy to improve the optimization algorithm, globally optimizes the key parameters in the mechanical fault recognition model, significantly improves the search ability of the model in the hyperparameter space, obtains the optimal parameter combination, thereby enhancing the stability and recognition accuracy of the mechanical fault recognition model, and realizes the comprehensive improvement of the performance of the mechanical fault recognition model, meeting the requirements of rapid and accurate intelligent diagnosis under complex working conditions and variable operating environments.

[0005] The technical solutions adopted by the application are as follows: The application provides an intelligent mechanical fault rapid diagnosis method, which comprises the following steps: Step S1: multi-source data acquisition; Step S2: data optimization processing; Step S3: mechanical anomaly rapid detection; Step S4: mechanical fault rapid identification; Step S5: model performance optimization; Step S6: intelligent diagnosis of mechanical faults.

[0006] Further, in step S1, the multi-source data collection, specifically by laying out multiple types of sensors during the operation of the mechanical equipment to collect data information, obtains mechanical fault diagnosis raw data; the mechanical fault diagnosis raw data includes historical mechanical operation data and real-time mechanical operation data, both of which include mechanical working data, mechanical physical sensor data and mechanical equipment maintenance data; the historical mechanical operation data further includes mechanical operation state and mechanical fault type; the mechanical operation state includes normal state and abnormal state.

[0007] Further, in step S2, the data optimization processing, specifically including the following steps: Step S21: data cleaning processing, specifically missing value processing, abnormal value processing and repeated data processing on the raw data; Step S22: data normalization processing, specifically using the Min-Max normalization method to standardize the cleaned data; Step S23: data time sequence segmentation processing, specifically setting sliding time window parameters based on the mechanical equipment operation beat and sensor sampling frequency, and using sliding window mechanism to perform step-by-step segmentation processing on the data, dividing the data according to window size and step length to obtain non-overlapping time segment sequence data.

[0008] Further, in step S3, the mechanical anomaly rapid detection, specifically including the following steps: Step S31: constructing and training an anomaly recognition model, specifically constructing a mechanical equipment anomaly determination model based on the Isolation Forest algorithm, and using historical mechanical operation data to train the model to obtain a trained mechanical anomaly recognition model; Step S32: mechanical equipment anomaly recognition, specifically inputting real-time mechanical operation data into the trained mechanical anomaly recognition model, the model calculating an anomaly score for the input time series data and comparing it with a set anomaly recognition threshold, if the anomaly score exceeds the anomaly recognition threshold, the mechanical operation state of the time window is determined to be abnormal, obtaining the mechanical anomaly recognition result, the mechanical anomaly recognition result including the mechanical operation state, the corresponding anomaly score and the anomaly time label of the time window; Step S33: Abnormal timing data dynamic extraction, specifically, according to the mechanical abnormality identification result, if the mechanical operation state is an abnormal state, then the time window corresponding to the abnormal time mark is taken as the starting point of the time window expansion, the forward time window is dynamically expanded based on the abnormal score expansion strategy, the abnormal period optimization data is extracted from the continuous mechanical fault diagnosis optimization data, the abnormal period optimization data is input into the LSTM neural network, the context analysis of the whole sequence is performed by using the timing modeling capability of the LSTM neural network, the key sub-period timing data most capable of representing the fault behavior in the optimization data is automatically identified and extracted, and the mechanical equipment abnormal timing data is obtained, if the mechanical operation state is a normal state, the abnormal timing data dynamic extraction process is not triggered, and the subsequent mechanical fault rapid identification step is not performed.

[0009] Further, in step S4, the mechanical fault rapid identification specifically includes the following steps: Step S41: Abnormal timing signal feature construction, specifically, the mechanical equipment abnormal timing data is processed by continuous wavelet transform, the time-frequency domain features thereof are extracted, the two-dimensional time-frequency thereof is obtained, and the time-frequency graph is stacked according to the channel dimension, and converted into a three-dimensional structured tensor ; Step S42: Construction of a mechanical fault identification model, specifically including the following steps: Step S421: Key fault feature adaptive extraction, specifically, a deep convolutional neural network with residual connection structure is constructed as a fault feature extraction network, the three-dimensional structured tensor is taken as the initial input data, convolution operation, batch normalization processing, nonlinear activation operation and cross-layer residual connection operation are sequentially performed, and multi-level key fault features are extracted layer by layer, and a random convolution disturbance mechanism is introduced in each convolution layer in the fault feature extraction network; finally, the output feature vector of the last residual block in the fault feature extraction network is taken as the key fault feature; the formula is as follows: ; ; In the formula, represents the feature map output by the lth convolution layer, represents the feature map output by the (l-1) th convolution layer, represents the bias term parameter of the lth convolution layer, represents a random disturbance distribution function of the lth convolution kernel parameter, represents a convolution operation, a standard parameter of the lth convolution kernel, represents a disturbed convolution kernel parameter, represents an initial disturbance intensity, denotes an attenuation coefficient, denotes a current training moment, denotes an identity matrix, used to construct a covariance matrix, denotes a Gaussian distribution, denotes a ReLU activation function, denotes a Hadamard product; Step S422: intelligent classification of mechanical fault types, specifically, inputting key fault features into a fault type classification layer for classification calculation to obtain a fault type classification result; Step S423: multi-device fault feature discrimination processing, specifically, constructing a multi-device fault feature discrimination module using a shallow multilayer perceptron neural network, inputting key fault features into the multi-device fault feature discrimination module, and outputting a domain label probability representing the discrimination result of whether the input features correspond to a known device or an unknown device; Step S43: designing a multi-device adaptability enhancement loss function, specifically, the loss function is composed of four sub-terms, i.e., a classification loss term, a feature consistency loss term, a maximum entropy regularization term, and a minimum conditional entropy regularization term, and the formula is as follows: ; ; In the formula, denotes a multi-device adaptability enhancement loss function value, denotes a cross-entropy loss function, denotes a feature consistency loss function, denotes a maximum entropy regularization term function, denotes a minimum conditional entropy regularization function, denotes a known device sample, denotes an unknown device sample, denotes a real fault type, denotes a number of known device samples, denotes an i-th known device sample, denotes a number of unknown device samples, denotes a v-th unknown device sample, denotes a fault feature extraction network, denotes a multi-device fault feature discrimination module, 、 and respectively denote weight coefficients of each loss term; Step S44: model training, specifically, using historical mechanical operation data as training samples, and jointly optimizing parameters in the model through a back propagation algorithm according to the designed multi-device adaptability enhancement loss function, continuously updating during the training process until the loss function converges, and obtaining a trained mechanical fault recognition model.

[0010] Further, in step S5, the model performance is optimized, specifically including the following steps: Step S51: Constructing an optimized individual parameter, specifically encoding the mechanical fault identification model parameter as a search individual position vector; Step S52: Obtaining the model performance optimal parameter combination, specifically performing global optimization on the mechanical fault identification model parameter through the improved optimization algorithm to obtain the performance optimal parameter combination of the mechanical fault identification model; specifically including the following steps: Step S521: Initializing the search population individual, specifically generating N search individual position vectors in the parameter space through random initialization, each individual encoding representing a candidate parameter combination, forming an initial search population; Step S522: Calculating the individual fitness value, specifically calculating the search individual fitness value in the population ; the mechanical fault identification model performance established based on the search individual position is used as the individual fitness value, Step S523: Self-adaptive exploration stage position update, specifically based on the disturbance guided expansion strategy, combined with the dynamically decaying exploration factor, introducing a disturbance direction update operation to the current position of the individual, the formula used is as follows: ; In the formula, represents the jth dimension position of the ith individual in the d+1th generation population, represents the dth generation population exploration factor, d represents the current iteration number, j represents the index of the dimension, and r represents a uniformly distributed random integer in the range of [1, N], represents a uniformly distributed random number in the range of [0, 1], represents the random dimension position of the random individual in the dth generation population, represents the random dimension position of the ith individual in the dth generation population, represents a random dimension; Step S524: Local search stage position update, specifically introducing a jump-driven fine-tuning mechanism to guide the individual to perform a disturbance jump update around the current optimal solution; the formula used is as follows: ; In the formula, , and all represent uniformly distributed random numbers in the range of [0, 1], represents the jth dimension position of the random individual in the dth generation population, represents the current iteration optimal individual position, represents the jump step length; represents the jth dimension position of the ith individual in the dth generation population, represents the maximum number of iterations; Step S525: fine-tuning of the individual position, specifically, generating a temporary candidate position vector according to a multi-channel fine-tuning candidate strategy, and determining whether to update according to the fitness function, to obtain the final individual position of the next iteration; the formula used is as follows: ; In the formula, represents the temporary candidate position vector, and both represent random numbers uniformly distributed in the range of [0, 1], represents a normally distributed random variable, represents the average fitness value of the current population solution, represents the global optimal individual position, represents a random individual position; Step S526: obtaining the optimal individual position, specifically, after each iteration, the fitness values of all individuals in the current population are evaluated, and if the fitness value of a certain individual position is better than the current global optimal individual position, the global optimal individual position is updated with the individual; Step S527: termination of search iteration, specifically, when the search individual fitness value is higher than the fitness threshold and the maximum number of iterations is reached, the search is terminated and the global optimal individual position is obtained, and the global optimal individual position specifically refers to the performance optimal parameter combination of the mechanical fault identification model; Step S53: performance optimization of the mechanical fault identification model, specifically, based on the performance optimal parameter combination, the parameters of the trained mechanical fault identification model are adjusted to obtain a performance optimal mechanical fault identification model.

[0011] Further, in step S6, the intelligent diagnosis of the mechanical fault, specifically, first based on the mechanical anomaly identification result, if the mechanical operating state is an abnormal state, an abnormal alarm of the mechanical equipment is triggered immediately, warning information is sent to the equipment manager, and the extracted mechanical equipment abnormal timing data is input into the performance optimal mechanical fault identification model to quickly obtain the real-time fault identification result of the current equipment, and the mechanical equipment fault type is determined according to the real-time fault identification result; if the mechanical operating state is a normal state, no alarm is triggered, and no mechanical fault rapid identification is performed, and the normal monitoring is continued, thereby realizing the rapid diagnosis of the mechanical equipment fault.

[0012] The above-mentioned scheme has the following beneficial effects: (1) In view of the technical problems in the traditional mechanical fault diagnosis method that all mechanical operation data are uniformly analyzed, resulting in slow diagnosis response speed, inability to quickly locate faults, and insufficient fault recognition result accuracy, which further leads to mechanical fault recognition lag and increased misjudgment rate under complex working conditions, the present scheme innovatively proposes a mechanical fault rapid diagnosis technology based on abnormal trigger key sequence extraction, which first performs mechanical abnormality rapid recognition, triggers abnormal timing data extraction of key fault behavior only when abnormality is determined, and then performs intelligent fault type recognition, effectively reducing the interference of normal data on fault diagnosis, significantly reducing the consumption of computing resources, improving the sensitivity and discrimination accuracy of mechanical fault features, and achieving efficient, accurate and intelligent rapid fault diagnosis of mechanical equipment under complex working conditions.

[0013] (2) In view of the technical problems in the existing mechanical fault recognition model that the fault feature expression ability is limited, and the model fault recognition ability is insufficient under multi-device conditions, resulting in a decrease in fault recognition accuracy between different devices, the present scheme innovatively introduces a random convolution disturbance mechanism and designs a multi-device adaptability enhanced loss function to enhance the deep expression ability of the model to key fault features and the consistency discrimination ability across devices, effectively improving the extraction accuracy of fault features and the adaptability under multi-device conditions, and achieving rapid, accurate and intelligent diagnosis of mechanical faults under complex working conditions and multi-device conditions.

[0014] (3) In view of the technical problems in the existing mechanical fault recognition model that the model parameter setting is unreasonable, resulting in unstable and insufficient fault recognition results, the present scheme uses a disturbance guided expansion strategy, a jump driven fine tuning mechanism and a multi-channel fine tuning candidate strategy to improve the optimization algorithm, globally optimizes the key parameters in the mechanical fault recognition model, significantly improves the search ability of the model in the hyperparameter space, obtains the optimal parameter combination, and enhances the stability and recognition accuracy of the mechanical fault recognition model, achieving comprehensive improvement of the performance of the mechanical fault recognition model, and meeting the needs of rapid and accurate intelligent diagnosis under complex working conditions and variable operating environments. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A flowchart of an intelligent mechanical fault rapid diagnosis method provided by the present application is shown in the figure. Figure 2 A flowchart of step S3 is shown in the figure. Figure 3 A flowchart of step S4 is shown in the figure. Figure 4 A flowchart of step S42 is shown in the figure. Figure 5 A flowchart of step S5 is shown in the figure. Figure 6This is a flowchart illustrating step S52; The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0017] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0018] Example 1, see Figure 1 The technical solution adopted by this invention is as follows: This invention provides an intelligent method for rapid diagnosis of mechanical faults, which includes the following steps: Step S1: Multi-source data acquisition, specifically, during the operation of mechanical equipment, data information acquisition operations are performed to obtain raw data for mechanical fault diagnosis; Step S2: Data optimization processing, used to ensure data quality and meet the time-series analysis requirements for mechanical equipment anomaly judgment and fault type identification. Specifically, it involves data cleaning, data standardization, and time-series segmentation of the raw mechanical fault diagnosis data to obtain structured optimized mechanical fault diagnosis data. Step S3: Rapid detection of mechanical anomalies is used to determine whether mechanical equipment is in an abnormal operating state. After detecting an anomaly, key time-series data related to the fault behavior are extracted to provide high-quality input for subsequent fault identification. Specifically, a mechanical anomaly identification model is first built and trained based on the isolated forest algorithm. Then, the model is used to determine the operating status of real-time data. If the determination result is an abnormal state, an adaptive time window expansion is triggered to extract abnormal time period data from the optimized time-series data and input it into the LSTM network for time-series context analysis to extract key abnormal subsequences representing fault behavior, thereby obtaining the mechanical anomaly identification result and the mechanical equipment anomaly time-series data. Step S4: mechanical fault rapid identification, for realizing intelligent identification of fault type under cross-device condition in abnormal state of mechanical equipment; specifically, the abnormal time series data is processed by continuous wavelet transform to generate a three-dimensional time-frequency tensor, and then the tensor is input into a deep neural network with residual connection structure and random disturbance mechanism to extract key fault features, and then a fault type classification layer is used to identify the fault type, and a mechanical fault identification model is constructed by combining multi-device fault feature identification processing; the mechanical fault identification model is trained based on historical data and combined with a designed multi-device adaptive enhanced loss function to obtain a trained mechanical fault identification model suitable for multi-device scenarios; Step S5: model performance optimization, for improving the overall performance of the mechanical fault identification model; specifically, by constructing optimized individual parameters and using an improved optimization algorithm to perform global search on the mechanical fault identification model parameters, the optimal parameter combination is obtained, and the model parameters are adjusted to obtain the optimal performance mechanical fault identification model; Step S6: mechanical fault intelligent diagnosis, specifically, based on the mechanical abnormality recognition result to determine the running state, if abnormal, triggering an alarm and inputting the abnormal time series data into the fault identification model to determine the fault type, if normal, no alarm and identification are triggered, and continue to monitor.

[0019] Embodiment two, see Figure 1 This embodiment is based on the above embodiment, in step S1, the multi-source data acquisition, specifically, by laying multiple types of sensors during the operation of the mechanical equipment to collect data information, obtaining the mechanical fault diagnosis raw data; the mechanical fault diagnosis raw data includes historical mechanical operation data and real-time mechanical operation data, the historical mechanical operation data and real-time mechanical operation data both include mechanical working data, mechanical physical sensor data and mechanical equipment maintenance data; the historical mechanical operation data further includes mechanical operation state and mechanical fault type; The mechanical working data includes load state, running time and speed state; The mechanical fault type information in the historical mechanical operation data is partly from the labeled fault samples of known devices to provide clear supervision signals, and partly from unlabeled operation samples of unknown devices to support key fault feature migration under cross-device condition; The mechanical physical sensor data includes vibration data, temperature data, electrical data, speed data and sound data; The mechanical maintenance data includes repair records and maintenance logs; The repair record includes repair time, repair fault occurrence time, repair fault type and treatment measures; The mechanical operation state includes normal state and abnormal state.

[0020] Embodiment three, refer to Figure 1 This embodiment is based on the above embodiment, in step S2, the data optimization processing, specifically for mechanical fault diagnosis original data data cleaning processing, data standardization processing and data time sequence segmentation processing, get mechanical fault diagnosis optimization data; including the following steps: Step S21: data cleaning processing, for improving the quality and effectiveness of mechanical fault diagnosis original data, specifically for the original data missing value processing, abnormal value processing and repeated data processing; The missing value processing is specifically through the time series linear interpolation method to fill the missing values in the continuous sampling with the time points before and after the interpolation; The abnormal value processing is specifically using the method based on Z-score to analyze the statistical fluctuation of the data, if the standardized value of the data point exceeds the set threshold, it is determined as an abnormal point and is removed and replaced with the median value; The repeated data processing is specifically using the joint deduplication algorithm based on timestamp and data hash value to quickly deduplicate the data frames with the same collection time and content, and retaining the unique effective record; Step S22: data normalization processing, for solving the difference problem of various data in numerical range and physical dimension, specifically using Min-Max normalization method to standardize the cleaned data; Step S23: data time sequence segmentation processing, for structuring the continuous original collection data according to the mechanical equipment running characteristics and time window requirements, specifically based on the mechanical equipment running beat and sensor sampling frequency, setting the sliding time window parameters, and using the sliding window mechanism to process the data step by step, according to the window size and step length to divide the data, get the non-overlapping time slice sequence data; The sliding time window parameters include window length and step length.

[0021] Embodiment four, refer to Figure 1 And Figure 2 This embodiment is based on the above embodiment, in step S3, the mechanical anomaly rapid detection, for identifying whether the mechanical equipment is in abnormal running state, and automatically extracting the key time sequence data related to the abnormal behavior after judging the abnormality, providing high quality input for subsequent fault identification, specifically including the following steps: Step S31: build and train the abnormality identification model, for establishing a model that can identify whether the mechanical equipment running state is abnormal, specifically based on the isolation forest algorithm to build the mechanical equipment abnormality judgment model, and use the historical mechanical running data to train the model, get the trained mechanical anomaly identification model; Step S32: mechanical equipment anomaly recognition, for real-time anomaly recognition of mechanical equipment running state, rapid judgment of whether the current running deviates from the normal working condition, specifically inputting the real-time mechanical running data into the trained mechanical anomaly recognition model, the model calculating the anomaly score of the input time series data, and comparing it with the set anomaly recognition threshold, if the anomaly score exceeds the anomaly recognition threshold, the mechanical running state of the time window is determined as an abnormal state, and the mechanical anomaly recognition result is obtained, the mechanical anomaly recognition result specifically represents the current time window anomaly determination information, including the mechanical running state, the corresponding anomaly score and the anomaly time mark of the time window; Step S33: dynamic extraction of abnormal time series data, for adaptive extraction of key time series data related to fault behavior based on mechanical anomaly recognition result after equipment anomaly, specifically, according to the mechanical anomaly recognition result, if the mechanical running state is abnormal, the time window corresponding to the abnormal time mark is taken as the starting point of time window expansion, the forward time window is dynamically expanded based on the anomaly score expansion strategy, the abnormal period optimization data is extracted from the continuous mechanical fault diagnosis optimization data, the abnormal period optimization data is input into the LSTM neural network, and the context analysis of the whole sequence is performed by using the time series modeling ability, the key sub-sequence time series data which can best represent the fault behavior in the optimization data is automatically recognized and extracted, and the mechanical equipment abnormal time series data is obtained, if the mechanical running state is normal, the abnormal time series data dynamic extraction process is not triggered, and the subsequent mechanical fault rapid recognition step is not executed; The anomaly score expansion strategy specifically refers to that if the anomaly scores of the continuous multiple time windows gradually increase, the abnormal behavior is considered to be evolving, and the time window range is dynamically expanded forward accordingly to cover the complete abnormal precursor process.

[0022] By performing the above operation, for the technical problems that the traditional mechanical fault diagnosis method uniformly analyzes all mechanical running data, resulting in slow diagnosis response speed, inability to quickly locate faults, and insufficient fault recognition result accuracy, and further causing mechanical fault recognition lag and increasing misjudgment rate in complex working conditions, the present scheme innovatively proposes a mechanical fault rapid diagnosis technology of abnormal trigger key sequence extraction, which performs mechanical anomaly rapid recognition first, triggers the extraction of abnormal time series data which can best represent the fault behavior only when it is determined to be abnormal, and then performs intelligent fault type recognition, effectively reducing the interference of normal data on fault diagnosis, significantly reducing the consumption of computing resources, improving the sensitivity and discrimination accuracy of mechanical fault characteristics, and realizing efficient, accurate and intelligent rapid fault diagnosis of mechanical equipment in complex working condition environment.

[0023] Embodiment five, refer to Figure 1 , Figure 3 and Figure 4The embodiment is based on the above-mentioned embodiment, and in step S4, the mechanical fault rapid identification is used to realize the rapid and intelligent identification of the fault type of the mechanical equipment under the abnormal state under the cross-equipment condition, and specifically includes the following steps: Step S41: Abnormal timing signal feature construction, for converting the mechanical equipment abnormal timing data into a structured input feature, specifically, the mechanical equipment abnormal timing data is processed through continuous wavelet transform, the time-frequency domain feature is extracted, the two-dimensional time-frequency is obtained, and the time-frequency graph is stacked according to the channel dimension, and the three-dimensional structured tensor is converted The shape of the tensor is Wherein H is the frequency dimension, W is the time dimension, and C is the number of signal channels; Step S42: Constructing a mechanical fault identification model, for establishing a model for intelligently identifying the mechanical fault type under the cross-equipment condition, specifically including the following steps: Step S421: Key fault feature adaptive extraction, for extracting deep semantic features highly related to mechanical faults, specifically, a deep convolutional neural network with residual connection structure is constructed as a fault feature extraction network, and the three-dimensional structured tensor As the initial input data, convolution operation, batch normalization processing, nonlinear activation operation and cross-layer residual connection operation are sequentially performed, and multi-level key fault features are extracted layer by layer, and a random convolution disturbance mechanism is introduced in each convolution layer in the fault feature extraction network, that is, a dynamic disturbance term is introduced to the weight parameters of each convolution layer in the training process, thereby constructing a disturbed convolution kernel parameter; finally, the output feature vector of the last residual block in the fault feature extraction network As the key fault feature; the formula is as follows: ; ; In the formula, Indicates the feature map output by the lth convolution layer, Indicates the feature map output by the (l-1) th convolution layer, Indicates the bias parameter of the lth convolution layer, Indicates the random disturbance distribution function of the lth convolution kernel parameter, Indicates the convolution operation, The standard parameter of the lth convolution kernel, Indicates the disturbed convolution kernel parameter, Indicates the initial disturbance intensity, Indicates the attenuation coefficient, Indicates the current training time, Indicates the unit matrix, which is used to construct the covariance matrix, Indicates the Gaussian distribution, denotes a ReLU activation function, denotes a Hadamard product; Step S422: intelligent classification of mechanical fault types, specifically, inputting the key fault features into a fault type classification layer for classification calculation to obtain a fault type classification result; the fault type classification layer includes one fully connected layer and one Softmax activation function layer; the formula used is as follows: ; In the formula, denotes the fault type classification result, denotes a weight matrix of the fault type classification layer, denotes a bias vector of the fault type classification layer; Step S423: multi-device fault feature discrimination processing, used for device domain discrimination training of the extracted key fault features; specifically, a shallow multi-layer perceptron neural network is used to construct a multi-device fault feature discrimination module, the key fault features are input into the multi-device fault feature discrimination module, and an output is a domain label probability, indicating a discrimination result of whether the input features correspond to a known device or an unknown device; Step S43: designing a multi-device adaptability enhancement loss function, used for improving the migration generalization ability of the mechanical fault recognition model under multiple device working conditions, and simultaneously enhancing the discriminability of the fault features, the multi-device feature adaptability, and the fault recognition robustness, so as to meet the demand for fault intelligent recognition accuracy and stability in complex engineering environments; specifically, the loss function is composed of four sub-terms, including a classification loss term, a feature consistency loss term, a maximum entropy regularization term, and a minimum conditional entropy regularization term; the formula used is as follows: ; ; ; ; ; ; In the formula, denotes a multi-device adaptability enhancement loss function value, denotes a cross-entropy loss function, denotes a feature consistency loss function, denotes a maximum entropy regularization term function, used for enhancing the distribution uniformity of unknown device samples and avoiding overconfident prediction of the model on unknown device samples, represents the minimum conditional entropy regularizer, used to minimize the conditional entropy of the class under the joint distribution of known device and unknown device samples, and helps to strengthen the deterministic classification of key features, improve the consistency of the decision boundary between different devices, and enhance the discriminability and stability of the model in a multi-device environment, represents the known device samples, represents the unknown device samples, represents the true fault type, represents the number of known device samples, represents the i-th known device sample, represents the number of unknown device samples, represents the v-th unknown device sample, represents the fault feature extraction network, represents the multi-device fault feature discrimination module, represents the predicted probability of the known device sample on the j-th fault, represents the fault type classification layer, represents the average probability of the sample being predicted on the j-th fault under the joint distribution of known device and unknown device, and K represents the number of fault types. 、 and represent the weight coefficients of each loss term, represents the known sample distribution, represents the unknown sample distribution. Step S44: model training, specifically, historical mechanical operation data is used as training samples, and a multi-device adaptability enhancement loss function is designed to jointly optimize the parameters in the model through a back propagation algorithm. The training process is continuously updated until the loss function converges, and the trained mechanical fault recognition model is obtained.

[0024] By performing the above operations, in view of the technical problems that the existing mechanical fault recognition model has limited fault feature expression capability, and the model's fault recognition capability is insufficient under multi-device conditions, resulting in a decrease in fault recognition accuracy between different devices, the present scheme innovatively introduces a random convolution perturbation mechanism and designs a multi-device adaptability enhancement loss function to enhance the model's deep-level expression capability of key fault features and consistent discrimination capability across devices, effectively improving the extraction accuracy of fault features and adaptability in a multi-device environment, and achieving rapid, accurate and intelligent diagnosis of mechanical faults under complex working conditions and multi-device conditions.

[0025] Embodiment six, refer to Figure 1 、 Figure 5 and Figure 6 , this embodiment is based on the above-mentioned embodiments, in step S5, the model performance optimization, specifically including the following steps: Step S51: constructing an optimized individual parameter for establishing an optimized search space, specifically encoding the mechanical fault identification model parameters into a search individual position vector; the mechanical fault identification model parameters include the convolution kernel size, learning rate, batch size, and loss function weight coefficient; Step S52: obtaining a model performance optimal parameter combination for improving the overall performance of the model, specifically performing global optimization on the mechanical fault identification model parameters through an improved optimization algorithm to obtain the performance optimal parameter combination of the mechanical fault identification model; Step S521: initializing a search population individual, specifically generating N search individual position vectors in the parameter space through random initialization, with each individual encoding representing a candidate parameter combination to form an initial search population; Step S522: calculating an individual fitness value, specifically calculating the search individual fitness value in the population ; the performance of the mechanical fault identification model established based on the search individual position is used as the fitness value of the individual, Step S523: self-adaptive exploration stage position update for guiding the search individual to fully explore the potential optimal solution region in the global range in the early stage, specifically introducing a disturbed direction update operation to the current position of the individual based on the disturbance guided expansion strategy combined with a dynamically decaying exploration factor, and the used formula is as follows: ; ; In the formula, represents the jthdimensional position of the ithindividual in the d+1thgeneration population, represents the dthgeneration population exploration factor, d represents the current iteration number, represents the maximum iteration number, represents the initialization population exploration factor, with a value range of [0, 1], j represents the index of the dimension, and r represents a uniformly distributed random integer in the range of [1, N], represents a uniformly distributed random number in the range of [0, 1], represents a random dimensional position of a random individual in the dthgeneration population, represents a random dimensional position of the ithindividual in the dthgeneration population, represents a random dimension; Step S524: local search stage position update for fine mining in the neighborhood of the current optimal solution in the later stage of the algorithm, specifically introducing a jump-driven fine-tuning mechanism to guide the individual to perform a disturbed jump update around the current optimal solution; the used formula is as follows: ; ; wherein, , and all represent random numbers uniformly distributed in the range of [0, 1], represents the jthdimensional position of a random individual in the dthgeneration population, and represent random numbers subject to a normal distribution, represents the optimal individual position in the current iteration, represents a jump step size; represents the jthdimensional position of the ithindividual in the dthgeneration population; Step S525: fine-tuning of individual position, for fine-tuning of the individual position through mutation to enhance the ability of the model to jump out of local optimum; specifically, generating a temporary candidate position vector according to a multi-channel fine-tuning candidate strategy, and selecting whether to update according to a fitness function; the formula used is as follows: ; ; wherein, represents a temporary candidate position vector, and all represent random numbers uniformly distributed in the range of [0, 1], represents a random variable subject to a normal distribution, represents the average fitness value of the current population solution, represents an individual fitness function, represents the final individual position in the next iteration, represents the global optimal individual position, represents a random individual position; Step S526: obtaining of the optimal individual position, specifically, after the end of each round of iteration, evaluating the fitness values of all individuals in the current population, and if the fitness value of the position of an individual is better than the current global optimal individual position, updating the global optimal individual position with the individual; Step S527: termination of search iteration, specifically, when the fitness value of the search individual is higher than a fitness threshold and the maximum number of iterations is reached, terminating the search and obtaining the global optimal individual position, which specifically refers to the optimal parameter combination of the performance of the mechanical fault identification model; Step S53: performance optimization of the mechanical fault identification model, specifically, based on the optimal parameter combination of the performance, adjusting the parameters of the trained mechanical fault identification model to obtain an optimal performance mechanical fault identification model.

[0026] By performing the above operation, in view of the technical problem that unreasonable model parameter settings exist in the existing mechanical fault identification model, resulting in unstable fault identification results and insufficient precision, the scheme adopts a disturbance-guided expansion strategy, a jump-driven fine-tuning mechanism and a multi-channel fine-tuning candidate strategy to improve the optimization algorithm, globally optimizes the key parameters in the mechanical fault identification model, significantly improves the search capability of the model in the hyperparameter space, obtains the optimal parameter combination in performance, thereby enhancing the stability and identification precision of the mechanical fault identification model, and comprehensively improves the performance of the mechanical fault identification model, meeting the requirements of rapid and accurate intelligent diagnosis under complex working conditions and variable operating environments.

[0027] Embodiment seven, refer to Figure 1 The embodiment is based on the above-mentioned embodiment, in step S6, the intelligent diagnosis of mechanical fault is specifically: first, based on the mechanical abnormality identification result, if the mechanical operation state is an abnormal state, an abnormal alarm of mechanical equipment is triggered immediately, warning information is sent to the equipment manager, and the extracted mechanical equipment abnormal timing data is input into the performance-optimal mechanical fault identification model to quickly obtain the real-time fault identification result of the current equipment, and the mechanical equipment fault type is judged according to the real-time fault identification result; if the mechanical operation state is a normal state, no alarm is triggered, and no mechanical fault rapid identification is performed, and the normal monitoring is continued, thereby realizing the rapid diagnosis of mechanical equipment fault.

[0028] It should be noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0029] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application.

[0030] The above describes the present application and its embodiments, which are not limited, and the drawings only show one of the embodiments of the present application, and the actual structure is not limited thereto. In general, if those skilled in the art are inspired, without departing from the purpose of the present application, without creative design, similar structure and embodiments of the technical solution can be designed, which shall belong to the protection scope of the present application.

Claims

1. An intelligent method for rapid diagnosis of mechanical faults, characterized in that: The method comprises the following steps: Step S1: multi-source data acquisition, through data information acquisition operation, to obtain mechanical fault diagnosis original data; Step S2: data optimization processing, data cleaning processing, data standardization processing and data time sequence segmentation processing are performed on the original data to obtain mechanical fault diagnosis optimization data; Step S3: mechanical anomaly rapid detection, a mechanical anomaly recognition model is constructed and trained, the running state of real-time data is determined, if the determination result is an abnormal state, an adaptive time window expansion is triggered, abnormal time period data is extracted from the optimization data, and input into an LSTM network for time sequence analysis, the key abnormal subsequence of the fault behavior is extracted, and the mechanical anomaly recognition result and the mechanical equipment abnormal time sequence data are obtained; Step S4: mechanical fault rapid identification, the mechanical equipment abnormal time sequence data is converted into a three-dimensional time-frequency tensor, and input into a deep neural network with a residual connection structure and a random disturbance mechanism to extract key fault features, the fault type classification layer is used to determine the fault type, and the multi-device fault feature determination processing is combined to construct a mechanical fault identification model, the historical data is used as the basis, and a design multi-device adaptive enhanced loss function is used for model training, and a trained mechanical fault identification model is obtained; Step S5: model performance optimization, by constructing an optimized individual parameter, and using a disturbance guided expansion strategy, a jump driven fine tuning mechanism and a multi-channel fine tuning candidate strategy to improve the optimization algorithm, the mechanical fault identification model parameters are globally searched, the model performance optimal parameter combination is obtained, and the model parameters are adjusted, and the performance optimal mechanical fault identification model is obtained; Step S6: mechanical fault intelligent diagnosis, based on the mechanical anomaly recognition result, the running state is determined, if it is abnormal, an alarm is triggered and the abnormal time sequence data is input into the mechanical fault identification model to determine the fault type, if it is normal, no alarm and identification are triggered, and the monitoring is continued.

2. The intelligent mechanical fault rapid diagnosis method according to claim 1, characterized in that: In step S3, the mechanical anomaly rapid detection is used to judge whether the mechanical equipment is in an abnormal running state, and to extract the key time sequence data related to the fault behavior after detecting the anomaly; Specifically, the following steps are included: Step S31: constructing and training an anomaly recognition model, specifically, constructing a mechanical equipment anomaly determination model based on the isolation forest algorithm, and using historical mechanical operation data for model training to obtain a trained mechanical anomaly recognition model; Step S32: mechanical equipment anomaly recognition, specifically, inputting real-time mechanical operation data into the trained mechanical anomaly recognition model, the model calculates the anomaly score of the input time sequence data, and compares it with the set anomaly recognition threshold, if the anomaly score exceeds the anomaly recognition threshold, the mechanical running state of the time window is determined as an abnormal state, the mechanical anomaly recognition result is obtained, and the mechanical anomaly recognition result includes the mechanical running state, the corresponding anomaly score and the abnormal time mark of the time window; Step S33: Abnormal timing data dynamic extraction, specifically, according to the mechanical abnormality identification result, if the mechanical operation state is an abnormal state, the abnormal time mark corresponding time window is taken as the starting point of time window expansion, the forward time window is dynamically expanded based on the abnormal score expansion strategy, the abnormal period optimization data is extracted from the continuous mechanical fault diagnosis optimization data, the abnormal period optimization data is input into the LSTM neural network, the context analysis of the whole sequence is performed by using the timing modeling capability, the key sub-section timing data in the optimization data that best represents the fault behavior is automatically identified and extracted, and the mechanical equipment abnormal timing data is obtained; if the mechanical operation state is a normal state, the abnormal timing data dynamic extraction process is not triggered, and the subsequent mechanical fault rapid identification step is not performed.

3. The method of claim 1, wherein the method further comprises: In step S4, the mechanical fault rapid identification is used to quickly realize intelligent identification of fault types under cross-device conditions in the abnormal state of the mechanical equipment; specifically, the following steps are included: Step S41: Abnormal timing signal feature construction, specifically, the abnormal timing data of the mechanical equipment is processed by continuous wavelet transform, the time-frequency domain features are extracted, the two-dimensional time-frequency is obtained, and the time-frequency diagram is stacked according to the channel dimension, which is converted into a three-dimensional structured tensor ; Step S42: Constructing a mechanical fault identification model; Step S43: Designing a multi-device adaptability enhanced loss function, specifically, the loss function is composed of four sub-items, including a classification loss item, a feature consistency loss item, a maximum entropy regularization item, and a minimum conditional entropy regularization item, and the formula is as follows: ； ； In the formula, denotes a multi-device adaptive enhanced loss function value, denotes a cross-entropy loss function, denotes a feature consistency loss function, denotes a maximum entropy regularization term function, denotes a minimum conditional entropy regularization function, denotes a known device sample, denotes an unknown device sample, denotes a real fault type, denotes a number of known device samples, denotes an i-th known device sample, denotes a number of unknown device samples, denotes a v-th unknown device sample, denotes a fault feature extraction network, denotes a multi-device fault feature discrimination module, , and respectively denote weight coefficients of each loss term; Step S44: Model training, specifically, the historical mechanical operation data is taken as a training sample, the multi-device adaptability enhanced loss function is designed, the parameters in the model are jointly optimized through a back propagation algorithm, and the training process is continuously iterated and updated until the loss function converges, and the trained mechanical fault identification model is obtained.

4. The intelligent mechanical fault rapid diagnosis method according to claim 1, characterized in that: In step S42: Constructing a mechanical fault identification model, specifically, the following steps are included: Step S421: key fault feature adaptive extraction, specifically, a deep convolutional neural network with residual connection structure is constructed as a fault feature extraction network, and a three-dimensional structured tensor As the initial input data, the convolution operation, batch normalization processing, nonlinear activation operation and cross-layer residual connection operation are sequentially performed, and the key fault features of multiple levels are extracted layer by layer. A random convolution disturbance mechanism is introduced in each convolution layer in the fault feature extraction network; finally, the output feature vector of the last residual block in the fault feature extraction network As the key fault feature; the formula is as follows: ； ； wherein, denotes the feature map output by the l-th convolutional layer, denotes the feature map output by the l-1-th convolutional layer, denotes the bias term parameter of the l-th convolutional layer, denotes the random perturbation distribution function of the l-th layer convolutional kernel parameter, denotes the convolution operation, denotes the standard parameter of the l-th layer convolutional kernel, denotes the perturbed convolutional kernel parameter, denotes the initial perturbation strength, denotes the decay coefficient, denotes the current training time, denotes the identity matrix, used to construct the covariance matrix, denotes the Gaussian distribution, denotes the ReLU activation function, denotes the Hadamard product; Step S422: Intelligent classification of mechanical fault types, specifically, the key fault features are input into the fault type classification layer for classification calculation, and the fault type classification result is obtained; Step S423: Multi-device fault feature discrimination processing, specifically, a shallow multi-layer perceptron neural network is used to construct a multi-device fault feature discrimination module, the key fault features are input into the multi-device fault feature discrimination module, and an domain label probability is output, which represents the discrimination result of the input features corresponding to the sample from the known device or the unknown device.

5. The intelligent mechanical fault rapid diagnosis method according to claim 1, characterized in that: In step S5, the model performance optimization is used to improve the overall performance of the mechanical fault identification model; specifically, the following steps are included: Step S51: Constructing an optimization individual parameter, specifically, the mechanical fault identification model parameters are encoded as search individual position vectors; Step S52: Obtaining a model performance optimal parameter combination, specifically, the improved optimization algorithm is used to globally optimize the mechanical fault identification model parameters, and the performance optimal parameter combination of the mechanical fault identification model is obtained; Step S53: Mechanical fault identification model performance optimization, specifically, based on the performance optimal parameter combination, the parameters of the trained mechanical fault identification model are adjusted, and the performance optimal mechanical fault identification model is obtained.

6. The intelligent mechanical fault rapid diagnosis method according to claim 1, characterized in that: In step S52, the model performance optimal parameter combination is obtained, specifically, the following steps are included: Step S521: initialize the search population individuals, specifically, generate N search individual position vectors in the parameter space through random initialization, each individual encodes a candidate parameter combination, forming an initial search population; Step S522: Calculate the individual fitness value, specifically, calculate the search individual fitness value in the population ; the performance of the mechanical fault identification model established based on the search individual position is taken as the fitness value of the individual Step S523: adaptive exploration stage position update, specifically, based on the perturbation guided expansion strategy, combined with a dynamically decaying exploration factor, introduce a perturbed direction update operation to the current position of the individual, the formula used is as follows: ； In the formula, represents the jth dimensional position of the ith individual in the d+1th generation population, represents the dth generation population exploration factor, d represents the current iteration number, j represents the index of the dimension, and r represents a random integer uniformly distributed in the range [1, N], represents a random number uniformly distributed in the range [0, 1], represents a random dimensional position of a random individual in the dth generation population, represents a random dimensional position of the ith individual in the dth generation population, represents a random dimension; Step S524: local search stage position update, specifically, introduce a jump driven fine tuning mechanism to guide the individual to make a perturbed jump update around the current optimal solution; the formula used is as follows: ； wherein, , and are random numbers uniformly distributed in the range [0, 1], is the jthdimensional position of the random individual in the dthgeneration population, is the optimal individual position in the current iteration, is the jump step size; is the jthdimensional position of the ithindividual in the dthgeneration population, is the maximum number of iterations; Step S525: individual position fine tuning, specifically, according to the multi-channel fine tuning candidate strategy, generate temporary candidate position vectors, and according to the fitness function, select whether to update, to obtain the final individual position of the next iteration; the formula used is as follows: ； wherein, denotes the current time step, denotes the current time step, denote a random number uniformly distributed in the range [0, 1], denotes a normally distributed random variable, denotes the average fitness value of the current population solution, denotes the global optimal individual position, denotes a random individual position; Step S526: obtain the optimal position of the individual, specifically, after each iteration, evaluate the fitness values of all individuals in the current population, if the fitness of a certain individual position is better than that of the current global optimal individual position, update the global optimal individual position with the individual; Step S527: search iteration termination, specifically, when the search individual fitness value When the fitness threshold is higher and the maximum iteration number is reached, the search is terminated and the global optimal individual position is obtained, and the global optimal individual position specifically refers to the optimal parameter combination of the mechanical fault identification model.

7. The intelligent mechanical fault rapid diagnosis method according to claim 1, characterized in that: In step S6, the intelligent mechanical fault diagnosis, specifically, first based on the mechanical abnormality recognition result, if the mechanical operation state is an abnormal state, immediately trigger the mechanical equipment abnormality alarm, send the warning information to the equipment manager, and input the extracted mechanical equipment abnormal timing data into the performance optimal mechanical fault recognition model, quickly obtain the real-time fault recognition result of the current equipment, and judge the mechanical equipment fault type according to the real-time fault recognition result; if the mechanical operation state is normal, do not trigger the alarm, do not perform the mechanical fault rapid recognition, and continue the regular monitoring, thereby realizing the rapid diagnosis of the mechanical equipment fault.

8. The intelligent mechanical fault rapid diagnosis method according to claim 1, characterized in that: In step S1, the multi-source data acquisition, specifically, through the arrangement of multiple types of sensors during the operation of the mechanical equipment, data information is collected to obtain mechanical fault diagnosis raw data; the mechanical fault diagnosis raw data includes historical mechanical operation data and real-time mechanical operation data, the historical mechanical operation data and the real-time mechanical operation data both include mechanical working data, mechanical physical sensor data and mechanical equipment maintenance data; the historical mechanical operation data further includes mechanical operation state and mechanical fault type; the mechanical operation state includes normal state and abnormal state.

9. The intelligent mechanical fault rapid diagnosis method according to claim 1, characterized in that: In step S2, the data optimization processing, specifically, includes the following steps: Step S21: data cleaning processing, specifically, missing value processing, abnormal value processing and repeated data processing are performed on the raw data; Step S22: data normalization processing, specifically, the Min-Max normalization method is used to standardize the cleaned data; Step S23: data time sequence segmentation processing, specifically, based on the mechanical equipment operation beat and the sensor sampling frequency, set the sliding time window parameter, and use the sliding window mechanism to perform step-by-step segmentation processing on the data, divide the data according to the window size and step length to obtain non-overlapping time segment sequence data.

Images