Method for Detecting Abnormalities of Mechanical Equipment under Explicit Representation of Structural Information of Multi-Measurement Point Samples

Abstract

The present invention discloses a method for abnormal detection of mechanical equipment under explicit characterization of multi-measurement point sample structure information. The present invention constructs a multi-measurement point state detection sample structure information characterization module to statistically analyze the correlation between the detection samples of each measurement point, and adopts a multi-head attention mechanism to adaptively weight and integrate the results of multiple correlation analyses, based on which the structure information of the multi-measurement point state detection samples is explicitly characterized. Then, a spatial graph convolutional neural network and a multi-channel decoder are constructed and combined with the multi-measurement point state detection sample structure information characterization module as an abnormal detection model to calculate the reconstruction loss of the multi-measurement point monitoring data, and an abnormal detection threshold is calculated based on the exponential weighted moving average method, ultimately realizing the abnormal detection of mechanical equipment under the condition of multi-measurement point fault-free training samples. The present invention provides an effective solution for the abnormal detection of mechanical equipment under the condition of multi-measurement point fault-free training samples based on a supervised learning framework.

Description

Technical Field

[0001] The present invention relates to the technical field of abnormal detection of mechanical equipment, and particularly to a method for abnormal detection of mechanical equipment under explicit characterization of multi-measurement point sample structure information. Background Art

[0002] Abnormal detection of mechanical equipment is a method of analyzing and processing equipment monitoring data by using appropriate signal processing techniques, extracting state features, and formulating an abnormal recognition strategy to judge whether the mechanical equipment is operating normally. In the actual application process, a single sensor is easily affected by uncertain factors such as complex background noise in the environment. To improve the reliability of the detection results, sensors are often arranged at multiple measurement points to collect multi-measurement point monitoring data of mechanical equipment for analysis and processing. However, due to the highly non-linear correlation and significant difference in importance among the monitoring data at each measurement point under the coupling effect of multi-component vibration responses of complex mechanical equipment, it is difficult to reasonably utilize the multi-measurement point monitoring data, and the effects of feature extraction and fusion are difficult to achieve the best, which instead reduces the accuracy of abnormal detection. Therefore, it is necessary to carry out research on a method for abnormal detection of mechanical equipment applicable to multi-measurement point monitoring data.

[0003] Traditional abnormal detection methods (such as principal component analysis, complex spectrum technology, etc.) rely on prior knowledge to manually select feature indicators when processing multi-measurement point monitoring data, and have poor effects in the process of feature extraction and fusion application of multi-measurement point monitoring data, with low accuracy of abnormal detection results. In recent years, intelligent learning methods represented by deep neural networks use a data-driven approach to build models, which can adaptively extract and fuse features by training model parameters, and have been applied to the fields of data fusion and abnormal detection. However, existing intelligent learning methods only perform feature extraction on the data at each measurement point in turn during feature extraction, and then fuse the obtained features, ignoring the potential connections between the data at different measurement points during the feature extraction stage, resulting in the fusion features being difficult to completely represent the health state information contained in the equipment monitoring data. Therefore, it is necessary to study an abnormal detection method that can fully mine the health state information of mechanical equipment in multi-measurement point monitoring data. Summary of the Invention

[0004] The purpose of the present invention is to provide a mechanical equipment anomaly detection method under the explicit representation of multi-measurement point sample structure information to solve the problems existing in the prior art. The present invention constructs a multi-measurement point state detection sample structure information representation module to count the correlation between the detection samples of each measuring point, and adopts a multi-head attention mechanism to adaptively weighted integrate the results of multiple correlation analysis, and on this basis explicitly represents the structure information of the multi-measurement point state detection samples, and then constructs a spatial domain graph convolutional neural network and a multi-channel decoder and combines them with the multi-measurement point state detection sample structure information representation module as an anomaly detection model to calculate the reconstruction loss of multi-measurement point monitoring data, and calculates the anomaly detection threshold based on the exponentially weighted moving average method, and finally realizes the mechanical equipment anomaly detection under the condition of multi-measurement point fault-free training samples.

[0005] In order to achieve the above object, the present invention adopts the following technical scheme:

[0006] The method for detecting abnormality of mechanical equipment under the explicit representation of multi-measurement point sample structure information includes the following steps:

[0007] Step 1: Take the one-dimensional vibration signal collected by the sensors of multiple measuring points of each component of the detected object as the target signal, obtain multi-measurement point state detection samples by preprocessing the collected one-dimensional vibration signal, take the normal multi-measurement point state detection samples collected under the normal state of the mechanical equipment as the training set, and take other multi-measurement point state detection samples collected under other states as the test set;

[0008] Step 2: Based on the self-attention layer, a multi-point state detection sample structure information representation module is constructed. By counting the correlations between the detection samples of each measurement point, the autocorrelation matrix of the multi-point state detection samples is calculated, and the results of multiple correlation analyses are adaptively weighted and integrated using a multi-head attention mechanism. Based on this, the structural information of the multi-point state detection samples is explicitly represented. Then, an explicit representation of the structural information of the multi-point state detection samples generated by the spatial domain graph convolutional neural network is constructed, and the effective fusion of the state detection samples of each measurement point is achieved.

[0009] Step 3: Use the multi-point state detection sample structure information representation module and the spatial domain graph convolutional neural network constructed in step 2 as encoders, and combine them with the constructed multi-channel decoder as an anomaly detection model. Train the anomaly detection model based on the training set composed of normal multi-point state detection samples collected under the normal state of the mechanical equipment after division, and update the anomaly detection model parameters by weight decay adaptive moment estimation, so as to mine the nonlinear mapping relationship between the fused state detection samples of each measuring point and the state detection samples of each measuring point obtained after preprocessing of the corresponding measuring point;

[0010] Step 4: Calculate the reconstruction loss of the training set samples using the anomaly detection model obtained through training in Step 3, calculate the anomaly detection threshold based on the method of exponential weighted moving average, and infer the anomaly detection results of the test set samples according to the anomaly detection threshold.

[0011] Further, in Step 1, the collected one-dimensional vibration signals are preprocessed through a sliding window and Min-Max data normalization:

[0012] First, use a sliding window to intercept sequence segments of the one-dimensional vibration signal, and divide the collected one-dimensional vibration signal into samples X of length L; then, use Min-Max data normalization to process the divided samples. The calculation formula of Min-Max data normalization is as follows:

[0013]

[0014] In the formula, is the normalized sample, X max is the maximum amplitude in the sample, X min is the minimum amplitude in the sample.

[0015] Further, in Step 2, the correlation between the detection samples of each measurement point is statistically analyzed based on the multi-head self-attention layer. Specifically:

[0016] First, use the self-attention network module to extract the sample features of each measurement point. The calculation formula of the self-attention network module is as follows:

[0017]

[0018] In the formula, F i is the extracted feature vector, i is the measurement point number, N is the total number of measurement points, W linear is the trainable weight used by the self-attention network module, b linear is the trainable bias used by the self-attention network module, is the i-th normalized sample, ο represents the operation result of obtaining W linear and ;

[0019] After that, use the Pearson correlation coefficient to statistically analyze the correlation of the sample features of different measurement points. The calculation formula of the Pearson correlation coefficient is as follows:

[0020]

[0021] In the formula, L′ is the feature length, j is the data point index, and respectively represent the two feature vectors for which the Pearson correlation coefficient is to be calculated and The value at the j-th point and respectively represent the average values of two feature vectors and and is the Pearson correlation coefficient obtained after calculation.

[0022] Furthermore, in step 2, the multi-head attention mechanism is adopted to integrate the results of multiple correlation analyses; first, multiple self-attention network modules are initialized simultaneously to extract the sample features of each measurement point multiple times, and the Pearson correlation coefficients of multiple matched specific measurement point sample pairs are calculated. The calculation results are as follows:

[0023]

[0024] In the formula, k is the number of self-attention layers when processing the matched specific measurement point sample pairs using the multi-head attention mechanism, and i p , i q are the serial numbers of the two samples in the processed sample pair, are the Pearson correlation coefficients obtained after calculation by multiple self-attention network modules with different initializations respectively, is the vector composed of the above calculated multiple Pearson correlation coefficients spliced together;

[0025] After that, a linear transformation layer is used to reduce the dimension of the spliced Pearson correlation coefficient vector to the same dimension as that processed by the single-head attention mechanism, so as to realize the adaptive weighted integration of multiple correlation analyses. The calculation results are as follows:

[0026]

[0027] In the formula, W′ linear is the trainable weight used by the linear transformation layer, b′ linear is the trainable bias used by the linear transformation layer, ο represents the operation result of obtaining W′ linear and through matrix multiplication, is the calculation result with the same dimension as that processed by the single-head attention mechanism after dimension reduction.

[0028] Furthermore, the adaptive weighted integration result of multiple correlation analyses in step 2 explicitly represents the structural information of the multi-measurement point state detection samples. The explicit representation of the multi-measurement point state detection sample structural information consists of a finite non-empty set of vertices and a set of edges between the vertices. The preprocessed samples of each measurement point are used as the vertices of the explicit representation of the multi-measurement point state detection sample structural information, and the operation result after reducing the dimension of the spliced Pearson correlation coefficient vector using the linear transformation layer As the edge of the explicit representation of the multi-measurement point state detection sample structure information, where the vertex set V of the explicit representation of the multi-measurement point state detection sample structure information is represented as follows:

[0029]

[0030] In the formula, i1, i2, … i N is the serial number of the multi-measurement point state detection sample, respectively represent the preprocessed samples obtained after normalization of the multi-measurement point state detection samples corresponding to the serial numbers;

[0031] The edge set E of the explicit representation of the multi-measurement point state detection sample structure information is represented as follows:

[0032]

[0033] In the formula, E i,j represents the value of the i-th row and j-th column of the edge set E, i1, i2, … i N is the serial number of the multi-measurement point state detection sample, is the Pearson correlation coefficient calculated after adaptive weighted integration of the sample pair composed of the samples with the corresponding serial numbers;

[0034] After generating the edge set E, in order to prevent the values at specific positions in the edge set E from being too large or too small and thus affecting the subsequent calculation results, the Softmax calculation is performed on each row of the matrix to convert the values of each row into a probability distribution with the sum of all elements' weights being 1. The Softmax calculation formula is as follows:

[0035]

[0036] In the formula, E i,j represents the value of the i-th row and j-th column of the edge set E.

[0037] Furthermore, in step 2, the explicit representation of the multi-measurement point state detection sample structure information generated by the spatial graph convolutional neural network is processed; first, in order to avoid the abnormal samples in the signal acquisition process from interfering with the calculation results, it is considered that the sample pairs with smaller values in the edge set E are abnormal, and a shielding rule is set to shield the edge set E. The shielding rule is as follows:

[0038]

[0039] In the formula, ξ is the threshold of the shielding rule, and E′ i,j is the value of the i-th row and j-th column of the edge set E′ after shielding processing;

[0040] After that, the convolutional neural network layer is used to process the vertices of the explicit representation of the multi-measurement point state detection sample structure information To further extract data features and achieve effective fusion of the information of vertex set V by combining with the edge set E after masking processing, the fused vertex set V′ is represented as follows:

[0041]

[0042] In the formula, E′ 1,1 , E′ 1,2 , … E′ N,N are the values in the edge set E′ after masking processing, conv(·) represents the convolution operation, and ο represents obtaining the operation result in the way of matrix multiplication.

[0043] Furthermore, in step 3, the multi - measurement - point state detection sample structure information representation module and the spatial - domain graph convolutional neural network constructed in step 2 are used as the encoder, and combined with the constructed multi - channel decoder as the anomaly detection model. Specifically:[[]]

[0044] First, for each data in the fused vertex set V′, a specific - channel decoder module is constructed to raise the dimension to be the same as the data at each position in the vertex set V that explicitly represents the multi - measurement - point state detection sample structure information; then, the difference between the reconstructed data and the pre - processed data is calculated using linear transformation and cosine distance, and this is used as the loss function of the model. The calculation process is as follows:

[0045]

[0046] In the formula, i is the sample serial number, N is the total number of samples, cos(·) represents calculating the cosine distance, Linear(·) represents performing linear transformation on the data, ο represents obtaining the operation result in the way of matrix multiplication, (·) T represents obtaining the transposed result and here it is to transpose the row vector into a column vector, ||·||2 represents calculating the 2 - norm of the vector, and Loss is the loss function of the calculated model.

[0047] Furthermore, in step 3, the parameters of the anomaly detection model are updated in the way of weight - decay adaptive moment estimation. The parameter update formula of weight - decay adaptive moment estimation is as follows:

[0048]

[0049]

[0050]

[0051] In the formula, m t is the momentum of parameter update, v tThe second-order momentum for parameter update, t is the generation of parameter update, β1, β2, λ are parameters for adjusting the parameter update process, and θ t-1 and θ t respectively represent the model parameters after the (t - 1)-th update and the t-th update, represents calculating the gradient of the model parameters, Loss is the calculated model loss function, ε is a smoothing term to prevent the denominator from being zero, and η is the model learning rate.

[0052] Furthermore, for the reconstruction loss of the training set samples obtained after model training in step 4, an anomaly detection threshold is calculated based on the method of exponential weighted moving average. The calculation formula of the anomaly detection threshold is as follows:

[0053] T1 = L1

[0054] T i = λ′L i +(1 - λ′)T i-1

[0055] T′ Last = βT Last +(1 - β)L Max

[0056] In the formula, L1, L i respectively represent the loss function values obtained by calculating the samples that appear for the first time and the i-th time in chronological order after sample division through the anomaly detection model. L Max is the maximum value of the loss function values obtained by calculating all samples through the anomaly detection model. T1, T i-1 , T i , T Last are the anomaly detection thresholds accumulated by calculating the loss function values of the samples that appear for the first time, the (i - 1)-th time, the i-th time, and the last time. T′ Last is the finally obtained anomaly detection threshold after weighted adjustment. λ′ is the statistical weight of the current reconstruction loss in the moving weighted average process of the anomaly detection threshold, and β is the statistical weight used to determine the final threshold.

[0057] Compared with the prior art, the present invention has the following beneficial technical effects:

[0058] 1) The present invention proposes a multi-measurement point sample correlation analysis method based on the multi-head attention mechanism, realizes the adaptive weighted integration of multiple Pearson correlation coefficients, and improves the stability of the correlation analysis result by excluding the interference of model random initialization on the sample feature extraction result.

[0059] 2) The present invention proposes a multi-measurement point monitoring data fusion method that uses explicit representation of multi-measurement point sample structure information and a spatial graph convolutional neural network. A generation method for explicit representation of multi-measurement point sample structure information based on a finite non-empty vertex set and an edge set between vertices is formulated. At the same time, a shielding strategy for the edge set is formulated to make the mining of the health state information of mechanical equipment in multi-measurement point monitoring data more sufficient.

[0060] 3) The mechanical equipment anomaly detection model constructed by the present invention can capture the non-linear mapping relationship between the fused state detection samples of each measurement point and the state detection samples of each measurement point obtained after preprocessing of the corresponding measurement points at multiple levels of the multi-measurement point state detection sample structure information representation module, the spatial graph convolutional neural network module, and the multi-channel decoder module through the method of weight decay adaptive moment estimation. And an anomaly detection threshold is calculated based on the method of exponential weighted moving average to realize the detection of the abnormal state of mechanical equipment under the condition of multi-measurement point fault-free training samples, which has certain practical application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The accompanying drawings in the specification are used to provide a further understanding of the present invention, and constitute a part of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0062] Figure 1 is a flowchart of the method of the present invention;

[0063] Figure 2 is a structural schematic diagram of the prediction model of the method of the present invention;

[0064] Figure 3 is the mechanical equipment anomaly detection result of the method of the present invention under the condition of multi-measurement point fault-free training samples. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

[0066] It should be noted that the terms "first", "second", etc. in the specification, claims and above-mentioned drawings of the present invention are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention described here can be implemented in an order other than those illustrated or described here. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0067] A method for detecting mechanical equipment anomalies under explicit characterization of multi-measurement point sample structure information, as shown in Figure 1, includes the following steps:

[0068] Step 1: Using the one-dimensional vibration signals collected by sensors at multiple measurement points of each component of the object to be detected as the target signals, preprocess the multi-measurement point one-dimensional vibration signals collected through a sliding window and Min-Max data normalization. Use the sliding window to intercept the sequence segments of the one-dimensional vibration signals, thereby dividing the original data into samples X with a length of L; then, use Min-Max data normalization to process the divided samples. The calculation formula for Min-Max data normalization is as follows:

[0069]

[0070] In the formula, is the normalized sample, X max is the maximum amplitude in the sample, X min is the minimum amplitude in the sample.

[0071] Take the detection samples collected under the normal state of the mechanical equipment as the training set, and the samples collected under other states as the test set.

[0072] Step 2: Use the self-attention network module to extract the sample features of each measurement point. The calculation formula of the self-attention network module is as follows:

[0073]

[0074] In the formula, F i is the extracted feature vector, i is the measurement point number, N is the total number of measurement points, W linear is the trainable weight used by the self-attention network module, b linear is the trainable bias used by the self-attention network module, is the i-th normalized sample, ο represents obtaining W in the way of matrix multiplicationlinear The operation result with

[0075] After that, the Pearson correlation coefficient is used to statistically analyze the correlation of sample features at different measurement points. The calculation formula of the Pearson correlation coefficient is as follows:

[0076]

[0077] In the formula, L′ is the feature length, j is the data point index, and respectively represent two feature vectors for which the Pearson correlation coefficient is to be calculated and at the j-th point, and represent two feature vectors and of the average value, is the Pearson correlation coefficient obtained after calculation.

[0078] At the same time, multiple self-attention network modules are initialized to extract the sample features of each measurement point multiple times, and the Pearson correlation coefficients of the sample features at different measurement points are calculated. The calculation results are as follows:

[0079]

[0080] In the formula, k is the number of heads of the self-attention layer when using the multi-head attention mechanism to process the matched specific measurement point sample pairs, i p ,i q are the serial numbers of the two samples in the processed sample pair, are the Pearson correlation coefficients obtained after calculation by multiple self-attention network modules with different initializations respectively, is a vector composed of the above calculated multiple Pearson correlation coefficients spliced together.

[0081] After that, the linear transformation layer is used to reduce the dimension of the Pearson correlation coefficient vector obtained after splicing to the same dimension as that processed by the single-head attention mechanism, so as to realize the adaptive weighted integration of multiple correlation analyses. The calculation results are as follows:

[0082]

[0083] In the formula, W′ linear is the trainable weight used by the linear transformation layer, b′ linear is the trainable bias used by the linear transformation layer, ο represents the operation result of obtaining W′ linear and in the way of matrix multiplication, is the calculation result of the same dimension as that processed by the single-head attention mechanism after dimension reduction.​

[0084] The structural information of multi - measurement - point state detection samples is explicitly represented by the adaptive weighted integration results of multiple correlation analyses. The explicit representation of the structural information of multi - measurement - point state detection samples consists of a finite non - empty set of vertices and a set of edges between the vertices. Each pre - processed measurement - point sample is used as the vertex of the explicit representation of the structural information of multi - measurement - point state detection samples, and the operation result after reducing the dimension of the concatenated Pearson correlation coefficient vector using a linear transformation layer is used as the edge of the explicit representation of the structural information of multi - measurement - point state detection samples. Among them, the vertex set V of the explicit representation of the structural information of multi - measurement - point state detection samples is represented as follows:

[0085]

[0086] In the formula, i1, i2, … i N is the serial number of the multi - measurement - point state detection sample, respectively represent the pre - processed samples obtained after normalization of the multi - measurement - point state detection samples corresponding to the serial numbers.

[0087] The edge set E of the explicit representation of the structural information of multi - measurement - point state detection samples is represented as follows:

[0088]

[0089] In the formula, E i,j represents the value in the i - th row and j - th column of the edge set E, and i1, i2, … i N is the serial number of the multi - measurement - point state detection sample, is the Pearson correlation coefficient calculated after adaptive weighted integration for the sample pair composed of the samples with the corresponding serial numbers;

[0090] After generating the edge set E, in order to prevent the values at specific positions in the edge set E from being too large or too small and thus affecting the subsequent calculation results, the Softmax calculation is performed on each row of the matrix to convert the values of each row into a probability distribution with the sum of all elements' weights being 1. The Softmax calculation formula is as follows:

[0091]

[0092] In the formula, E i,j represents the value in the i - th row and j - th column of the edge set E.

[0093] Construct a spatial - domain graph convolutional neural network to process the explicit representation of the structural information of the generated multi - measurement - point state detection samples; First, in order to avoid the interference of abnormal samples in the signal acquisition process on the calculation results, it is considered that the sample pairs with smaller values in the edge set E are abnormal, and a shielding rule is set to shield the edge set E. The shielding rule is as follows:

[0094]

[0095] where ξ is the threshold of the shielding rule, and E′ i,j is the value at the i-th row and j-th column of the edge set E′ after shielding processing. Then, a convolutional neural network layer is used to process the vertices explicitly representing the structural information of the multi-measurement point state detection samples to further extract data features, and the vertex set V information is effectively fused by combining with the edge set E after shielding processing. The fused vertex set V′ is expressed as follows:

[0096]

[0097] where E′ 1,1 , E′ 1,2 , … E′ N,N are the values in the edge set E′ after shielding processing, conv(·) represents the convolution operation, and ο represents obtaining the operation result in the way of matrix multiplication.

[0098] Step 3: Use the multi-measurement point state detection sample structure information characterization module and the spatial domain graph convolutional neural network constructed in Step 2 as the encoder, and combine it with the constructed multi-channel decoder as the anomaly detection model. The structure of the constructed anomaly detection model is as Figure 2 shown. First, for each data in the fused vertex set V′, a specific channel decoder module is constructed to increase the dimension to the same dimension as the data at each position in the vertex set V that explicitly represents the structural information of the multi-measurement point state detection samples; then, the difference between the reconstructed data and the preprocessed data is calculated by using linear transformation and cosine distance, and this is used as the loss function of the model. The calculation process is expressed as follows:

[0099]

[0100] where i is the sample serial number, N is the total number of samples, cos(·) represents calculating the cosine distance, Linear(·) represents linearly transforming the data, ο represents obtaining the operation result in the way of matrix multiplication, and (·) T represents obtaining the transposed result and here it is converting the row vector to a column vector, ||·||2 represents calculating the 2-norm of the vector, and Loss is the loss function of the calculated model.

[0101] The parameters of the anomaly detection model are updated by using the method of weight decay adaptive moment estimation. The parameter update formula of weight decay adaptive moment estimation is as follows:

[0102]

[0103]

[0104]

[0105] Where m t is the momentum for parameter update, v t is the second-order momentum for parameter update, t is the generation of parameter update, β1, β2, λ are parameters for adjusting the parameter update process, θ t-1 and θ t respectively represent the model parameters after the (t - 1)-th update and after the t-th update, represents calculating the gradient of the model parameters, Loss is the model loss function calculated in Claim 7, ε is a smoothing term to prevent the denominator from being zero, and η is the model learning rate.

[0106] Step 4: For the reconstruction loss of the training set samples obtained after model training, calculate the anomaly detection threshold based on the exponentially weighted moving average method. The calculation formula for the anomaly detection threshold is as follows:

[0107] T1 = L1

[0108] T i = λ′L i +(1 - λ′)T i-1

[0109] T′ Last = βT Last +(1 - β)L Max

[0110] Where L1, L i respectively represent the loss function values obtained by calculating the samples that appear for the first time and the i-th time in chronological order after sample division using the anomaly detection model. L Max is the maximum value of the loss function values obtained by calculating all samples using the anomaly detection model. T1, T i-1 , T i , T Last are the anomaly detection thresholds accumulated by calculating the loss function values of the samples that appear for the first time, the (i - 1)-th time, the i-th time, and the last time. T′ Last is the finally obtained anomaly detection threshold after weighted adjustment. λ′ is the statistical quantity weight of the current reconstruction loss in the moving weighted average process of the anomaly detection threshold, and β is the statistical quantity weight used to determine the final threshold.

[0111] Verify this method based on a dataset of the main shaft support bearings of an escalator. The accuracy of anomaly detection of this method is 99.904% under the condition of training only using the monitoring data collected in the healthy state of the escalator. The anomaly detection results are as Figure 3As shown, this demonstrates the effectiveness of the proposed method in establishing the non-linear mapping relationship between the monitoring data of mechanical equipment and the equipment health status, as well as predicting the changes in the mechanical equipment health status based on the distribution relationship of the reconstruction loss and detection threshold reconstructed from the multi-point monitoring data.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the scope of its protection. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: after reading the present invention, various changes, modifications or equivalent substitutions can still be made to the specific implementation manners of the invention, but these changes, modifications or equivalent substitutions are all within the scope of the protection of the pending claims of the invention.

Claims

1. An abnormal detection method for mechanical equipment under explicit characterization of multi - measurement - point sample structure information, characterized in that, The following steps are involved: Step 1: Take the one-dimensional vibration signal collected by the sensors of multiple measuring points of each component of the detected object as the target signal, obtain multi-measurement point state detection samples by preprocessing the collected one-dimensional vibration signal, take the normal multi-measurement point state detection samples collected under the normal state of the mechanical equipment as the training set, and take other multi-measurement point state detection samples collected under other states as the test set; Step 2: Based on the self-attention layer, a multi-point state detection sample structure information representation module is constructed. By counting the correlations between the detection samples of each measurement point, the autocorrelation matrix of the multi-point state detection samples is calculated, and the results of multiple correlation analyses are adaptively weighted and integrated using a multi-head attention mechanism. Based on this, the structural information of the multi-point state detection samples is explicitly represented. Then, an explicit representation of the structural information of the multi-point state detection samples generated by the spatial domain graph convolutional neural network is constructed, and the effective fusion of the state detection samples of each measurement point is achieved. Step 3: Use the multi-point state detection sample structure information representation module and the spatial domain graph convolutional neural network constructed in step 2 as encoders, and combine them with the constructed multi-channel decoder as an anomaly detection model. Train the anomaly detection model based on the training set composed of normal multi-point state detection samples collected under the normal state of the mechanical equipment after division, and update the anomaly detection model parameters by weight decay adaptive moment estimation, so as to mine the nonlinear mapping relationship between the fused state detection samples of each measuring point and the state detection samples of each measuring point obtained after preprocessing of the corresponding measuring point; Step 4: Use the anomaly detection model trained in step 3 to calculate the reconstruction loss of the training set samples, calculate the anomaly detection threshold based on the exponentially weighted moving average method, and infer the anomaly detection results of the test set samples based on the anomaly detection threshold.

2. The abnormal detection method for mechanical equipment under explicit characterization of multi-measurement point sample structure information according to claim 1, wherein In step 1, the collected one-dimensional vibration signal is preprocessed by sliding window and Min-Max data normalization: First, a sliding window is used to intercept the sequence fragments of the one-dimensional vibration signal, and the collected one-dimensional vibration signal is divided into samples X with a length of L; then, the divided samples are processed using Min-Max data normalization. The calculation formula of Min-Max data normalization is as follows: In the formula, is the normalized sample, and X max is the maximum amplitude in the sample, and X min is the minimum amplitude in the sample.

3. The mechanical equipment anomaly detection method under explicit characterization of multi-measurement point sample structure information according to claim 1, characterized in that, In step 2, the correlation between the detection samples of each measurement point is counted based on the multi-head self-attention layer, specifically: First, the self-attention network module is used to extract the sample features of each measurement point. The calculation formula of the self-attention network module is as follows: Where F i is the extracted feature vector, i is the measuring point number, N is the total number of measuring points, W linear is the trainable weight used in the self-attention network module, b linear is the trainable bias used in the self-attention network module, is the i-th normalized sample, ο represents the operation result of obtaining W linear and by matrix multiplication; Afterwards, the Pearson correlation coefficient is used to calculate the correlation of sample characteristics at different measurement points. The calculation formula of the Pearson correlation coefficient is as follows: where L′ is the characteristic length and j is the data point index, and respectively represent two feature vectors for which the Pearson correlation coefficient is to be calculated and the values at the j-th point, and respectively represent the two feature vectors and the average values, is the Pearson correlation coefficient obtained after calculation.

4. The abnormal detection method for mechanical equipment under explicit characterization of multi-measurement point sample structure information according to claim 3, wherein, In step 2, a multi-head attention mechanism is used to integrate the results of multiple correlation analyses. First, multiple self-attention network modules are initialized simultaneously to extract the sample features of each measurement point multiple times, and the Pearson correlation coefficients of multiple matched specific measurement point sample pairs are calculated. The calculation results are as follows: where k is the number of self-attention layers when processing the matched specific measurement point sample pairs using the multi-head attention mechanism, and i p , i q are the serial numbers of the two samples in the processed sample pair, are the Pearson correlation coefficients obtained after calculation by multiple self-attention network modules with different initializations respectively, is a vector composed of splicing the multiple Pearson correlation coefficients obtained from the above calculations; After that, a linear transformation layer is used to reduce the dimension of the concatenated Pearson correlation coefficient vector to the same dimension as the single-head attention mechanism, so as to achieve adaptive weighted integration of multiple correlation analyses. The calculation results are as follows: Where, W′ linear is the trainable weight used by the linear transformation layer, and b′ linear is the trainable bias used by the linear transformation layer. ο represents the operation result obtained by matrix multiplication of W′ linear and , is the calculation result of the same dimension as that processed by the single-head attention mechanism after dimensionality reduction.

5. The mechanical equipment anomaly detection method under explicit characterization of multi-measurement point sample structure information according to claim 4, characterized in that, In step 2, the adaptive weighted integration results of multiple correlation analyses are used to explicitly represent the structural information of the multi-measurement point state detection samples. The explicit representation of the structural information of the multi-measurement point state detection samples consists of a finite non-empty set of vertices and a set of edges between the vertices. Each preprocessed measurement point sample is used as a vertex for the explicit representation of the structural information of the multi-measurement point state detection samples, and the operation result after reducing the dimension of the concatenated Pearson correlation coefficient vector using a linear transformation layer is used as an edge for the explicit representation of the structural information of the multi-measurement point state detection samples. Among them, the vertex set V of the explicit representation of the structural information of the multi-measurement point state detection samples is represented as follows: where \(i_1, i_2, \ldots, i\) N are the serial numbers of the multi - measurement - point state detection samples, respectively representing the pre - processed samples obtained after normalization of the multi - measurement - point state detection samples corresponding to the serial numbers; The edge set E for explicitly representing the structural information of the multi-measurement point status detection sample is expressed as follows: where E i,j represents the value at the i-th row and j-th column of the edge set E, and i1, i2, … i N are the serial numbers of the multi-measurement point state detection samples, is the Pearson correlation coefficient calculated after adaptive weighted integration for the sample pairs composed of samples with the corresponding serial numbers; After generating the edge set E, to prevent the values at specific positions in the edge set E from being too large or too small and thus affecting the subsequent calculation results, the Softmax calculation is performed on each row of the matrix to convert the values in each row into a probability distribution where the sum of all element weights is 1. The Softmax calculation formula is as follows: where E i,j represents the value at the i-th row and j-th column of the edge set E.

6. The mechanical equipment anomaly detection method under explicit characterization of multi-measurement point sample structure information according to claim 5, characterized in that In step 2, a spatial graph convolutional neural network is constructed to process the explicit representation of the structural information of the multi-measurement point status detection sample generated; First, to avoid abnormal samples in the signal acquisition process from interfering with the calculation results, it is considered that the sample pairs with smaller values in the edge set E are abnormal, and a shielding rule is set to shield the edge set E. The shielding rule is as follows: where ξ is the threshold of the shielding rule, and E′ i,j is the value at the i-th row and j-th column of the subsequent set E′ after the shielding process; After that, a convolutional neural network layer is used to process the vertices of the explicit representation of the multi-measurement point state detection sample structure information to further extract data features, and combined with the edge set E after masking processing, an effective fusion of the vertex set V information is realized. The fused vertex set V′ is represented as follows: where E′ 1,1 , E′ 1,2 , … E′ N,N are the values in the set E′ after masking processing, conv(·) represents the convolution operation, and ο represents obtaining the operation result in the form of matrix multiplication.

7. The abnormal detection method for mechanical equipment under explicit characterization of multi-measurement point sample structure information according to claim 1, wherein In step 3, the structural information representation module of the multi-measurement point status detection sample and the spatial graph convolutional neural network constructed in step 2 are used as encoders, and combined with the constructed multi-channel decoder as an anomaly detection model. Specifically: First, for each data in the fused vertex set V′, a specific channel decoder module is constructed to increase the dimension to the same dimension as the data at each position in the vertex set V of the explicit representation of the structural information of the multi-measurement point status detection sample; Then, the difference between the reconstructed data and the preprocessed data is calculated using linear transformation and cosine distance, and this is used as the loss function of the model. The calculation process is expressed as follows: Where \(i\) is the sample serial number, \(N\) is the total number of samples, \(\cos(\cdot)\) represents calculating the cosine distance, \(\text{Linear}(\cdot)\) represents performing a linear transformation on the data, \(\omicron\) represents obtaining the operation result in the way of matrix multiplication, and \((\cdot)\) T represents obtaining the transposed result, and here the row vector is transposed into a column vector, \(\|\cdot\|_2\) represents calculating the 2-norm of the vector, and \(\text{Loss}\) is the loss function of the calculated model.

8. The abnormal detection method of mechanical equipment under explicit characterization of multi-measurement point sample structure information according to claim 7, characterized in that, In step 3, the parameters of the anomaly detection model are updated using the method of weight decay adaptive moment estimation. The parameter update formula for weight decay adaptive moment estimation is as follows: where m t is the momentum for parameter update, v t is the second-order momentum for parameter update, t is the number of parameter update epochs, β1, β2, λ are parameters for adjusting the parameter update process, θ t-1 and θ t represent the model parameters after the (t - 1)-th update and the t-th update respectively, denotes the gradient of the model parameters, Loss is the calculated model loss function, ε is a smoothing term to prevent the denominator from being zero, and η is the model learning rate.

9. The mechanical equipment anomaly detection method under explicit characterization of multi-measurement point sample structure information according to claim 1, wherein, In step 4, for the reconstruction loss of the training set samples obtained after model training, the anomaly detection threshold is calculated based on the method of exponential weighted moving average. The calculation formula for the anomaly detection threshold is as follows: T1 = L1 T i = λ'L i + (1 - λ')T i-1 T′ Last = βT Last + (1 - β)L Max where L1 and L i respectively represent the loss function values obtained by calculating the samples that appear for the first time and the i-th time in chronological order after sample division through the anomaly detection model. L Max is the maximum value among the loss function values obtained by calculating all samples through the anomaly detection model. T1, T i-1 , T i , T Last are the anomaly detection thresholds cumulatively obtained by calculating the loss function values of the samples that appear for the first time, the (i - 1)-th time, the i-th time, and the last time. T′ Last is the finally obtained anomaly detection threshold after weighted adjustment. λ′ is the statistical weight of the current reconstruction loss in the moving weighted average process of the anomaly detection threshold, and β is the statistical weight used when determining the final threshold.

Images