Alignment method of predictive maintenance large model

Abstract

The invention provides an alignment method for a predictive maintenance large model. The method comprises the following steps: S1, acquiring text data and monitoring signal data related to the operation condition of industrial equipment; s2, extracting feature vectors from the text data and the monitoring signal data through a text feature extractor and a signal feature extractor respectively; s3, calculating the difference between the feature vectors; s4, repeating the steps S1-S3, and training the signal feature extractor by taking the minimum difference as a target according to the difference generated in each repeating process; s5, extracting a feature vector through the trained signal feature extractor; s6, inputting the feature vector into the large model, and outputting an identification result; s7, calculating the recognition accuracy of the large model based on the text data and the recognition result; and S8, repeating the steps S1-S7 until the recognition accuracy of the large model is greater than a preset threshold. According to the method, the multi-modal data can be effectively aligned and fused, so that efficient predictive maintenance is realized through the model.

Description

Technical Field

[0001] The present application relates to the technical field of predictive maintenance of industrial equipment, and in particular to an alignment method and device for a large predictive maintenance model. Background Art

[0002] In the context of the Industrial 4.0 era, predictive maintenance has received widespread attention as an important means to ensure the stable operation of industrial equipment and improve production efficiency. Predictive maintenance is a strategy that monitors the status of equipment and predicts possible failures, thereby performing maintenance in advance to prevent equipment failures and downtime. By collecting and analyzing the sensor data of the equipment, such as temperature, vibration, current, etc., the working status and performance of the equipment can be monitored. By analyzing and modeling this data, possible equipment failures can be predicted and corresponding maintenance measures can be taken before the failure occurs. This predictive maintenance method can reduce the losses and costs caused by equipment failures and improve the reliability and efficiency of equipment.

[0003] Traditional predictive maintenance methods mainly rely on expert experience and a single data source for fault diagnosis and life prediction. However, these methods face huge challenges in dealing with complex equipment and multi-source heterogeneous data, and it is difficult to accurately capture the subtle signals of equipment failure, resulting in a significant reduction in the accuracy of fault diagnosis and the reliability of prediction.

[0004] In recent years, with the rapid development of artificial intelligence technology, especially the successful application of deep learning in multimodal data processing, a new technical path has been provided for predictive maintenance. By integrating data from different modalities (such as text, signals, etc.), multimodal large models can monitor equipment status from multiple dimensions and significantly improve the accuracy and efficiency of predictive maintenance. However, how to effectively align and fuse these multimodal data so that the model can make full use of the information of various data has become a key technical challenge to achieve efficient predictive maintenance. Summary of the invention

[0005] The present application aims to solve one of the technical problems in the related art at least to some extent.

[0006] To this end, the first purpose of this application is to propose a method for aligning a large model of predictive maintenance, which can align the feature vectors of equipment status signals and the feature vectors of text that can be understood by humans in the semantic space, thereby realizing a large model of predictive maintenance, and finally processing and analyzing the signals through natural language descriptions and performing predictive maintenance.

[0007] The second object of the present application is to provide a computer device.

[0008] To achieve the above-mentioned purpose, the first embodiment of the present application proposes a method for aligning a large predictive maintenance model, including:

[0009] Step S1: acquiring text data and monitoring signal data related to the operation status of industrial equipment, wherein the text data and the monitoring signal have a predetermined relationship;

[0010] Step S2: extracting a first feature vector from the text data by a text feature extractor, and extracting a second feature vector from the monitoring signal data by a signal feature extractor;

[0011] Step S3: Calculate the difference between the first eigenvector and the second eigenvector;

[0012] Step S4: repeating steps S1-S3, training the signal feature extractor according to the difference between the vectors after each repetition, minimizing the difference between the feature vectors as the goal during training, fixing the parameters of the text feature extractor, and adjusting the parameters of the signal feature extractor;

[0013] Step S5: extracting a third feature vector from the monitoring signal data by using the trained signal feature extractor;

[0014] Step S6: input the third feature vector into the large model and output the recognition result;

[0015] Step S7: Calculate the recognition accuracy of the large model based on the text data and the recognition result;

[0016] Step S8: Repeat steps S1-S7 until the recognition accuracy of the large model is greater than a preset threshold.

[0017] The alignment method of the predictive maintenance large model in the embodiment of the present application can make full use of the information in different data sources and improve the accuracy and efficiency of predictive maintenance by effectively aligning and fusing the multimodal data; the embodiment of the present application significantly improves the accuracy of fault diagnosis and life prediction by effectively aligning the feature vector of the equipment status signal and the text feature vector that can be understood by humans in the semantic space; the embodiment of the present application adopts a self-supervised comparative learning strategy, which reduces the dependence on a large amount of labeled data, significantly reduces the manpower cost and time cost of model training, and thus reduces the model training cost. In addition, the proposed alignment method does not depend on a specific type of equipment or data, has good versatility, and can be widely used in the predictive maintenance of various industrial equipment. The signal is processed and analyzed through natural language description, making the maintenance decision more intuitive and easy to understand, providing a powerful decision support tool for maintenance personnel, and promoting the realization of decision support.

[0018] Optionally, in one embodiment of the present application, the text data includes text data in a first form and text data in a second form, and acquiring text data related to the operating status of the industrial equipment includes:

[0019] Acquire original text data describing fault characteristics and equipment information generated during the operation of industrial equipment, and convert the original text data into constants to obtain text data in a first form;

[0020] Obtain a predictive maintenance field text alignment template, and fill the template with the text data in the first form to obtain text data in the second form;

[0021] Extracting a first feature vector from the text data by a text feature extractor includes:

[0022] extracting a first feature vector from the text data in the second form by a text feature extractor;

[0023] Calculate the recognition accuracy of the large model based on text data and recognition results, including:

[0024] The recognition accuracy of the large model is calculated based on the text data in the first form and the recognition result.

[0025] Optionally, in one embodiment of the present application, before extracting the second feature vector from the monitoring signal data by the signal feature extractor, the method further includes:

[0026] The multiple monitoring signals in the monitoring signal data are regarded as multivariate time series, and each monitoring signal is set as a separate channel;

[0027] Split the time series of each channel and adjust the size of each time series;

[0028] Extracting a second feature vector from the monitoring signal data by a signal feature extractor includes:

[0029] The time series is input into the signal feature extractor and the second feature vector is output.

[0030] Optionally, in one embodiment of the present application, the time series is input into a signal feature extractor, and a second feature vector is output, including:

[0031] Input the time series into the signal feature extraction layer and output the segmented time series;

[0032] The segmented time series is input into the multi-layer signal additional feature extraction layer, processed by each layer of the signal additional feature extraction layer, and the second feature vector is output.

[0033] Optionally, in one embodiment of the present application, the time series is input into a signal feature extraction layer, and the segmented time series is output, including:

[0034] Transform the time series to obtain a first transformation result matrix, and construct a first feature matrix based on the first transformation result matrix;

[0035] Splitting the first transformation result matrix, constructing a split transformation result matrix, and constructing a second feature matrix based on the split transformation result matrix;

[0036] The second feature matrix is ​​concatenated to obtain a multi-level feature matrix, and the multi-level feature matrix is ​​used as the segmented time series.

[0037] Optionally, in one embodiment of the present application, the segmented time series is input into an additional feature extraction layer of the signal, and a feature vector is output, including:

[0038] Transforming the segmented time series to obtain a second transformation result matrix, and merging the first transformation result matrix and the second transformation result matrix to obtain a merged transformation result matrix;

[0039] Calculate additional multi-level feature matrices based on the merged transformation result matrix;

[0040] Input the additional multi-level feature matrix into the feedforward neural network, output the processed feature matrix, and add the segmented time series after the processed feature matrix to obtain the processed time series;

[0041] The processed time series is layer normalized to obtain the feature vector.

[0042] Optionally, in one embodiment of the present application, the first transformation result matrix is:

[0043] Q=XW Q ,K=XW K ,V=XW V

[0044] Among them, X is the time series, W Q , W K , W V is the learnable weight matrix, Q, K, V are the transformation result matrices;

[0045] The first characteristic matrix is:

[0046]

[0047] Among them, d k is the dimension of K, K T is the transpose of K;

[0048] The multi-level feature matrix is:

[0049] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0050] Among them, Head i =Attention(Q i ,K i ,V i ), Attention(Q i ,K i ,V i ) is the second characteristic matrix, Q i ,K i ,V i is the transformation result matrix of the segmentation, W O is a predetermined parameter matrix, and Concat() is a concatenation operation;

[0051] The second transformation result matrix is:

[0052]

[0053] Among them, Q extra , K extra 、V extra is the second transformation result matrix, For additional learnable parameters, randomly initialized from a normal distribution;

[0054] The combined transformation result matrix is:

[0055] Q combined =Q⊕Q extra , K combined =K⊕K extra , V combined =V⊕V extra

[0056] Among them, Q combined , K combined 、V combined is the transformed result matrix after merging, ⊕ represents the dimension merging operation;

[0057] The additional multi-level feature matrix is:

[0058] MultiHead(Q combined ,K combined ,V combined )=Concat(head1,...,head h )W O

[0059] Among them, head i =Attention(Q combinedi ,K combinedi ,V combinedi );

[0060] The processed feature matrix is:

[0061] FFN(x)=max(0,xW1+b1)W2+b2

[0062] Among them, x is an additional multi-level feature matrix, W1, W2, b1, b2 are network preset parameters;

[0063] The eigenvectors are:

[0064]

[0065] Among them, μ is the mean, σ 2 is the variance, ∈ is a constant, and x is the processed time series;

[0066] The second eigenvector is:

[0067] e=f L ...f3f2f1(x)

[0068] Among them, the second feature vector e is obtained through the L-layer signal additional feature extraction layer, x is the time series of the input signal feature extractor, and f L (x) indicates that the signal has been processed by the Lth layer of additional feature extraction layer.

[0069] Optionally, in one embodiment of the present application, the difference between the first eigenvector and the second eigenvector is expressed as:

[0070]

[0071] in, is the similarity loss from the first eigenvector to the second eigenvector, is the similarity loss between the second eigenvector and the first eigenvector,

[0072]

[0073]

[0074] in, t i ,e j Represent the matched text data feature vector and monitoring signal feature vector respectively, t i ,e j They represent the unmatched text data feature vector and the monitoring signal feature vector respectively, and τ is a predetermined parameter.

[0075] Optionally, in one embodiment of the present application, training a signal feature extractor according to the difference between the vectors includes:

[0076] Calculate the gradient of the difference between the vectors with respect to the learnable parameters in the signal feature extractor and update the learnable parameters using gradient descent.

[0077] The update process is expressed as:

[0078]

[0079] Among them, m represents the number of iterations, θ m represents the parameters after m iterations, represents the gradient of the optimization objective to the model parameters, and λ represents the update step size of the parameter update that needs to be corrected;

[0080] The recognition accuracy is:

[0081]

[0082] Among them, δ is a redundant parameter.

[0083] To achieve the above-mentioned purpose, the second embodiment of the present invention proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the alignment method of the above-mentioned predictive maintenance large model is implemented.

[0084] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0086] Figure 1 A flowchart of a method for aligning a large predictive maintenance model provided in Example 1 of the present application;

[0087] Figure 2 A flowchart of a method for aligning a large predictive maintenance model provided in Example 2 of the present application. ;

[0088] Figure 3 A flowchart of an alignment optimization method for an alignment method of a large predictive maintenance model provided in Example 3 of the present application;

[0089] Figure 4 A schematic diagram of the verification process of a method for aligning a large predictive maintenance model provided in Example 4 of the present application. DETAILED DESCRIPTION

[0090] Embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0091] The following describes the method and device for aligning the predictive maintenance big model according to the embodiment of the present application with reference to the accompanying drawings.

[0092] Figure 1 A flowchart of a method for aligning a large predictive maintenance model provided in Example 1 of the present application.

[0093] like Figure 1 As shown, the alignment method of the predictive maintenance big model includes the following steps:

[0094] Step S1: acquiring text data and monitoring signal data related to the operation status of industrial equipment, wherein the text data and the monitoring signal have a predetermined relationship;

[0095] Step S2: extracting a first feature vector from the text data by a text feature extractor, and extracting a second feature vector from the monitoring signal data by a signal feature extractor;

[0096] Step S3: Calculate the difference between the first eigenvector and the second eigenvector;

[0097] Step S4: repeating steps S1-S3, training the signal feature extractor according to the difference between the vectors after each repetition, minimizing the difference between the feature vectors as the goal during training, fixing the parameters of the text feature extractor, and adjusting the parameters of the signal feature extractor;

[0098] Step S5: extracting a third feature vector from the monitoring signal data by using the trained signal feature extractor;

[0099] Step S6: input the third feature vector into the large model and output the recognition result;

[0100] Step S7: Calculate the recognition accuracy of the large model based on the text data and the recognition result;

[0101] Step S8: Repeat steps S1-S7 until the recognition accuracy of the large model is greater than a preset threshold.

[0102] The alignment method of the predictive maintenance large model in the embodiment of the present application can make full use of the information in different data sources and improve the accuracy and efficiency of predictive maintenance by effectively aligning and fusing the multimodal data; the embodiment of the present application significantly improves the accuracy of fault diagnosis and life prediction by effectively aligning the feature vector of the equipment status signal and the text feature vector that can be understood by humans in the semantic space; the embodiment of the present application adopts a self-supervised comparative learning strategy, which reduces the dependence on a large amount of labeled data, significantly reduces the manpower cost and time cost of model training, and thus reduces the model training cost. In addition, the proposed alignment method does not depend on a specific type of equipment or data, has good versatility, and can be widely used in the predictive maintenance of various industrial equipment. The signal is processed and analyzed through natural language description, making the maintenance decision more intuitive and easy to understand, providing a powerful decision support tool for maintenance personnel, and promoting the realization of decision support.

[0103] See also Figure 2 , Figure 2 A flowchart of a method for aligning a large predictive maintenance model provided in Embodiment 2 of the present application is shown in FIG. The method is executed by an electronic device.

[0104] like Figure 2 As shown, the method includes:

[0105] Constructing a multimodal dataset for large model alignment in predictive maintenance;

[0106] Construct encoders for embedding features of multimodal data;

[0107] Constructing a multimodal contrastive learning optimization strategy for large model alignment in predictive maintenance;

[0108] Towards a multimodal evaluation approach for large model alignment for predictive maintenance.

[0109] See also Figure 3 , Figure 3 A flowchart of an alignment optimization method of an alignment method of a predictive maintenance large model provided in Example 3 of the present application is shown in FIG. The method is executed by an electronic device.

[0110] like Figure 3 As shown, the method includes:

[0111] Step S1: Collect text data related to the operating status of the rotating machinery, such as maintenance records, fault case analysis reports, etc.;

[0112] Step S2: Acquire monitoring signals of the rotating machinery, including but not limited to vibration signals, temperature signals, sound signals, etc. These signals can reflect the current operating status of the rotating machinery; the monitoring signals monitor the signals of the predetermined equipment,

[0113] Specifically, there is a predetermined relationship between the text data collected in step S1 and step S2 and the monitoring signal;

[0114] Step S3: Extracting feature vectors of text data from the text data obtained in step S1 using the Transformer large model;

[0115] Step S4: extracting a feature vector of the monitoring signal from the monitoring signal obtained in step S2 using a neural network;

[0116] Step S5: Compare the differences between the feature vectors in step S3 and step S4;

[0117] Step S6: Optimizing the feature vector of the monitoring signal in step S4 according to the relationship between the difference in step S5 and the predetermined target;

[0118] Furthermore, step S3 specifically includes:

[0119] Step S31: converting text data describing fault characteristics and device information into a unique constant;

[0120] Step S32: Determine the text alignment template in the field of predictive maintenance, fill the template based on the above constants, and construct the text data that needs to be aligned. The text template is as follows:

[0121] {

[0122] "Device information": "Basic information of the device provided by the manufacturer",

[0123] "Main signal feature information": "mean, standard deviation and other features",

[0124] "Device health status": "Current health status of the device predicted based on signal characteristics",

[0125] "Fault type": "Possible fault type inferred based on signal characteristics",

[0126] "Remaining Lifetime": "Remaining lifetime predicted based on current signal characteristics",

[0127] "Health score": "a quantitative score of the current health status of the device",

[0128] "Future signal state prediction": "Future signal state predicted based on current trends and features",

[0129] "Maintenance Recommendation": "Maintenance or overhaul recommendations based on the health status and remaining life of the equipment",

[0130] }

[0131] The specific information in the above template needs to be completed according to the actual situation of the device.

[0132] Step S33: Input the text data obtained from step S32 into the Transformer large model to determine the feature vector of the text data.

[0133] Furthermore, Transformer large models include ChatGLM series, Qwen series and other large models.

[0134] Furthermore, step S4 specifically includes:

[0135] Step S41: treating the obtained multiple monitoring signals as a multivariate time series, and setting each monitoring signal as a separate channel;

[0136] Step S42: Segment the time series of each channel and adjust the size of each time series;

[0137] Step S43: Input the segmented time series obtained in step S42 into the signal feature extraction model to determine the feature vector of the monitoring signal.

[0138] Further, in step S5, the difference between the feature vector of the text data and the feature vector of the monitoring signal satisfies the following expression:

[0139]

[0140] t i ,e i Represent the matched text data feature vector and monitoring signal feature vector respectively;

[0141] t i ,e j Represent the unmatched text data feature vector and monitoring signal feature vector respectively;

[0142] τ is a predetermined parameter;

[0143] Represents the similarity loss from the feature vector of text data to the feature vector of monitoring signal;

[0144] Represents the similarity loss from the monitoring signal feature vector to the text data feature vector.

[0145] Further, in step S5, the difference between the feature vector of the text data and the feature vector of the monitoring signal satisfies the following expression:

[0146]

[0147] Wherein, L represents the difference between the feature vector of the text data and the feature vector of the monitoring signal.

[0148] Further, in step S43, the process of extracting the monitoring signal by the signal feature extraction model includes the following steps:

[0149] Construct the signal feature extraction layer. The steps of the signal feature extraction layer are as follows:

[0150] The obtained segmented time series is transformed in the following way:

[0151] Q=XW Q ,K=XW K ,V=XW V

[0152] Among them, W Q ,W K ,W V is a learnable weight matrix, X is a time series matrix, Q, K, and V are three types of transformation result matrices respectively;

[0153] Use the Q, K, and V matrices to determine the importance of the matrix eigenvectors, and multiply the resulting matrix V by the importance to obtain the feature matrix Attention(Q,K,V)

[0154]

[0155] d k is the dimension of the K matrix,

[0156]

[0157] Split Q, K, V and construct the transformation result matrix Q of the split i , K i 、V i , use the transformation result matrix after segmentation to construct the feature matrix Attention (Q i ,K i ,V i ). The segmented feature matrix Head i =Attention(Q i ,K i ,V i ) to obtain a multi-level feature matrix

[0158] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0159] W ois a predetermined parameter matrix, and Concat(x) is a concatenation operation.

[0160] The predetermined parameters of the signal feature extraction layer are initialized to the existing large model parameters.

[0161] Furthermore, in step S43, the process of extracting the monitoring signal by the signal feature extraction model further includes the following steps:

[0162] Construct the signal additional feature extraction layer f(x). The steps of the signal additional feature extraction layer are as follows:

[0163] The obtained segmented time series X is transformed in the following way:

[0164]

[0165] in, and are additional learnable parameters, randomly initialized from a normal distribution.

[0166] Merge the original and extra parameters. Merge the original Q, K, V with the Q, K, V transformed by the extra parameters:

[0167] Q combined =Q⊕Q extra , K combined =K⊕K extra , V combined =V⊕V extra

[0168] Where ⊕ represents a dimension merging operation.

[0169] Use the combined transformation result matrix Q combined , K combined , V combined Perform calculations on additional multi-level feature matrices:

[0170] MultiHead(Q combined ,K combined ,V cimbined )=Concat(head1,...,head h )W O

[0171] Among them, each head i =Attention(Q combinedi ,K combinedi ,V combinedi )

[0172] Input the additional multi-level feature matrix into the feedforward neural network FFN(x)

[0173] FFN(x)=max(0,xW1+b1)W2+b2

[0174] Among them, W1, W2, b1, b2 are network preset parameters.

[0175] Add the segmented time series to the time series FFN(x)+X after the feedforward neural network FFN

[0176] Normalize the output FFN(x)+X to get the signal feature vector e

[0177]

[0178] where μ is the mean, σ 2 is the variance, ∈ is a constant.

[0179] Furthermore, in step S43, the process of extracting the monitoring signal by the signal feature extraction model further includes the following steps:

[0180] Construct L layers of signal additional feature extraction layers f1(x), f2(x), ..., f L (x). The feature vector e of the monitoring signal can be obtained by passing through multiple layers of signal additional feature extraction layers, that is, e = f L …f2f1(x).

[0181] See also Figure 4 , Figure 4 A schematic diagram of a verification process of a method for aligning a large predictive maintenance model provided in Example 4 of the present application. The method is executed by an electronic device.

[0182] like Figure 4 As shown, the method includes:

[0183] Step S7: inputting the feature vector of the monitoring signal optimized in step S6 into the large model;

[0184] Step S8: Identify the feature vector of the monitoring signal optimized by the large model.

[0185] Step S9: Determine whether to end the optimization based on the comparison between the recognition result of step S8 and the predetermined target.

[0186] Furthermore, the optimization steps in step S6 are as follows:

[0187] The learnable predetermined parameters in steps 6 and 7 are represented by θ.

[0188] According to the difference L between the feature vector in step 5 and the feature vector of the monitoring signal, the gradient of L with respect to the learnable predetermined parameter is calculated.

[0189] Using the gradient descent method to update the learnable predetermined parameters:

[0190]

[0191] m represents the number of iterations, λ represents the step size of the parameter update that needs to be corrected, Represents the gradient of the optimization objective with respect to the model parameters.

[0192] Furthermore, in step S8, the recognition result can be expressed as g(e), where g is a generative artificial intelligence model.

[0193] Furthermore, in step S9, the steps of comparing the recognition result with the predetermined target are as follows:

[0194] Set a predetermined accuracy threshold.

[0195] Obtain the recognition result g(e) and convert the text data into a unique constant in S31, give the redundant parameter δ, and calculate the actual accuracy. The accuracy formula is as follows:

[0196]

[0197] If the accuracy is lower than the predetermined accuracy threshold, continue to update the predetermined parameters. If the accuracy is higher than the predetermined accuracy threshold, stop updating the learnable parameter θ.

[0198] Furthermore, the following analysis results can be obtained for the monitoring signals through the above-mentioned aligned predictive maintenance big model:

[0199] “By analyzing the detection signal of the key transmission component bearing A, we obtained the following:

[0200] "Equipment Information": "Bearing Model XYZ, Manufacturer Company A",

[0201] "Main signal characteristic information":"Mean value: X, standard deviation: X, fault characteristic frequency

[0202] "Equipment health status": "Based on the signal characteristics, it is predicted that the equipment has an inner circle fault",

[0203] "Fault Type":"Bearing Inner Ring Fault",

[0204] "Remaining Life": "Based on the current signal characteristics, the predicted remaining life is about 1000 hours",

[0205] "Health level score": "70 (out of 100)",

[0206] "Future signal status prediction": "If maintenance is not performed, the failure will increase"

[0207] "Maintenance Recommendation": "It is recommended to immediately shut down the machine for inspection and replace the damaged inner ring of the bearing."

[0208] In order to implement the above embodiments, the present invention further proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method described in the above embodiments is implemented.

[0209] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0210] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0211] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0212] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute the instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0213] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0214] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0215] In addition, each functional unit in each embodiment of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0216] The storage medium mentioned above may be a read-only memory, a disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present application. A person of ordinary skill in the art may change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A method for aligning a large predictive maintenance model, characterized in that: The method comprises: Step S1: Acquire text data and monitoring signal data related to the operating status of industrial equipment, wherein the text data and the monitoring signal have a predetermined relationship; Step S2: extracting a first feature vector from the text data by a text feature extractor, and extracting a second feature vector from the monitoring signal data by a signal feature extractor; Step S3: calculating the difference between the first eigenvector and the second eigenvector; Step S4: repeating steps S1-S3, training the signal feature extractor according to the difference between the vectors after each repetition, minimizing the difference between the feature vectors as the goal during training, fixing the parameters of the text feature extractor, and adjusting the parameters of the signal feature extractor; Step S5: extracting a third feature vector from the monitoring signal data by using the trained signal feature extractor; Step S6: input the third feature vector into the large model and output the recognition result; Step S7: Calculating the recognition accuracy of the large model based on the text data and the recognition result; Step S8: Repeat steps S1-S7 until the recognition accuracy of the large model is greater than a preset threshold.

2. The method according to claim 1, characterized in that The text data includes text data in a first form and text data in a second form, and acquiring text data related to the operating status of the industrial equipment includes: Acquire original text data describing fault characteristics and equipment information generated during the operation of industrial equipment, and convert the original text data into constants to obtain text data in a first form; Obtaining a text alignment template in the field of predictive maintenance, and filling the template with the text data in the first form to obtain text data in a second form; The extracting a first feature vector from the text data by a text feature extractor comprises: extracting a first feature vector from the text data in the second form by a text feature extractor; The calculating the recognition accuracy of the large model based on the text data and the recognition result includes: The recognition accuracy of the large model is calculated based on the text data in the first form and the recognition result.

3. The method according to claim 1, characterized in that Before extracting the second feature vector from the monitoring signal data by the signal feature extractor, the method further includes: The multiple monitoring signals in the monitoring signal data are regarded as multivariate time series, and each monitoring signal is set as a separate channel; Split the time series of each channel and adjust the size of each time series; The extracting a second feature vector from the monitoring signal data by a signal feature extractor comprises: The time series is input into the signal feature extractor, and the second feature vector is output.

4. The method according to claim 3, characterized in that The step of inputting the time series into the signal feature extractor and outputting the second feature vector comprises: Input the time series into a signal feature extraction layer, and output the segmented time series; The segmented time series is input into a multi-layer signal additional feature extraction layer, processed by each layer of the signal additional feature extraction layer, and a second feature vector is output.

5. The method according to claim 4, characterized in that The step of inputting the time series into a signal feature extraction layer and outputting a segmented time series comprises: Transform the time series to obtain a first transformation result matrix, and construct a first feature matrix based on the first transformation result matrix; Splitting the first transformation result matrix, constructing a split transformation result matrix, and constructing a second feature matrix based on the split transformation result matrix; The second feature matrices are concatenated to obtain a multi-level feature matrix, and the multi-level feature matrix is ​​used as the segmented time series.

6. The method according to claim 5, characterized in that The segmented time series is input into the additional feature extraction layer of the signal, Output feature vector, including: Transforming the segmented time series to obtain a second transformation result matrix, and merging the first transformation result matrix and the second transformation result matrix to obtain a merged transformation result matrix; Calculate additional multi-level feature matrices based on the merged transformation result matrix; Input the additional multi-level feature matrix into the feedforward neural network, output the processed feature matrix, and add the segmented time series after the processed feature matrix to obtain the processed time series; The processed time series is layer normalized to obtain the feature vector.

7. The method according to claim 6, characterized in that The first transformation result matrix is: Q=XW Q ,K=XW K ,V=XW V Among them, X is the time series, W Q , W K , W V is the learnable weight matrix, Q, K, V are the transformation result matrices; The first characteristic matrix is: Among them, d k is the dimension of K, K T is the transpose of K; The multi-level feature matrix is: MultiHead(Q,K,V)=Concat(head1,...,head h )W O Among them, Head i =Attention(Q i ,K i ,V i ), Attention(Q i ,K i ,V i ) is the second characteristic matrix, Q i ,K i ,V i is the transformation result matrix of the segmentation, W O is a predetermined parameter matrix, and Concat() is a concatenation operation; The second transformation result matrix is: Among them, Q extra , K extra 、V extra is the second transformation result matrix, For additional learnable parameters, randomly initialized by normal distribution; The combined transformation result matrix is: Among them, Q combined , K combined 、V combined is the combined transformation result matrix, Represents a dimension merge operation; The additional multi-level feature matrix is: MultiHead(Q combined ,K combined ,V combined )=Concat(head1,...,head h )W O in, The feature matrix after the processing is: FFN(x)=max(0,xW1+b1)W2+b2 Among them, x is an additional multi-level feature matrix, W1, W2, b1, b2 are network preset parameters; The feature vector is: Among them, μ is the mean, σ 2 is the variance, ∈ is a constant, and x is the processed time series; The second eigenvector is: e=f L ...f3f2f1(x) Among them, the second feature vector e is obtained through the L-layer signal additional feature extraction layer, x is the time series of the input signal feature extractor, and f L (x) indicates that the signal has been processed by the Lth layer of additional feature extraction layer.

8. The method according to claim 1, characterized in that The difference between the first eigenvector and the second eigenvector is expressed as: in, is the similarity loss from the first eigenvector to the second eigenvector, is the similarity loss between the second eigenvector and the first eigenvector, in, t i ,e j Represent the matched text data feature vector and monitoring signal feature vector respectively, t i ,e j They represent the unmatched text data feature vector and the monitoring signal feature vector respectively, and τ is a predetermined parameter.

9. The method according to claim 1, characterized in that The signal feature extractor is trained according to the difference between the vectors, including: Calculate the gradient of the difference between the vectors with respect to the learnable parameters in the signal feature extractor and update the learnable parameters using gradient descent. Wherein, the updating process is expressed as: Among them, m represents the number of iterations, θ m represents the parameters after m iterations, represents the gradient of the optimization objective to the model parameters, and λ represents the update step size of the parameter update that needs to be corrected; The recognition accuracy is: Among them, δ is a redundant parameter.

10. A computer device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method according to any one of claims 1 to 9 is implemented.

Images