Gearbox variable operating condition fault detection method based on normal sample aggregation contrast learning network

Abstract

A method for detecting gearbox faults under variable operating conditions using a normal sample aggregation contrastive learning network is described. This method uses a vibration sensor to collect vibration signals from the gearbox under varying speed conditions while operating in different health states. Samples in the training set are continuously subjected to offset and identity-preserving transformations to generate sets of virtual normal samples and virtual abnormal samples. A deep contrastive learning network is established, and a contrastive loss based on the generated virtual samples is calculated based on the aggregation of normal samples and the separation of abnormal samples. The network is trained by minimizing the contrastive loss. The trained memory sample set is input into the trained network, and the minimum cosine distance between each sample and the remaining memory sample set is used as the anomaly score to determine the detection threshold. Finally, the anomaly score of the test sample is calculated, and the health state of the gearbox is determined based on the threshold. This method implements fault detection for gearboxes under variable operating conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical equipment condition monitoring, and in particular to a gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network. Background Art

[0002] Monitoring the health status of mechanical equipment and providing abnormal alarms are crucial for maintaining the safe operation of the equipment. Existing intelligent monitoring methods can accurately detect faults under a single operating condition. However, the continuous fluctuations in operating conditions under variable operating conditions will lead to large differences in the collected signals, resulting in a distribution shift between the test samples and the training samples. This can cause the monitoring model trained under fixed operating conditions to generate false alarms or false alarms when faced with vibration signals under variable operating conditions. At the same time, the representations extracted by existing detection models have poor fault discrimination and may be sensitive to operating condition information such as speed but not to the fault state. This limits the versatility of fault detection models in environments with variable speeds and frequent operating condition switching. Self-supervised contrastive learning has excellent representation learning capabilities and provides a new approach to solving the above problems. However, traditional contrastive learning objectives exclude different normal samples, making it difficult to aggregate the representations of normal samples to generate representative normal representations under variable operating conditions.

[0003] The above information disclosed in this Background section is only for enhancement of understanding of the background of the invention and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0004] The present invention provides a gearbox variable operating condition fault detection method based on a normal sample aggregation contrast learning network. The method continuously offsets and replaces the original signal to generate virtual samples, constructs a deep contrast learning framework, designs a contrast loss function with the goal of virtual normal sample aggregation and virtual abnormal sample separation, and trains the network. In the latent space, the minimum cosine distance between the test sample and the memory sample is used as the anomaly score. The detection threshold is determined based on the quantile of the anomaly score of the memory sample. Finally, the fault result is determined by comparing the anomaly score with the detection threshold.

[0005] A gearbox variable operating condition fault detection method based on a normal sample aggregation contrast learning network includes:

[0006] a. Use a vibration sensor to collect vibration signals from the gearbox under varying speed conditions. The signals are then segmented into fixed-length samples and normalized to construct training, memory, and test sample sets.

[0007] b. Continuously perform offset transformation and identity-preserving transformation on the samples in the training set to generate a set of virtual normal samples and virtual abnormal samples;

[0008] c. Establish a deep contrastive learning network, which includes an encoder, a linear layer, and a normalization layer. The encoder is selected as a one-dimensional ResNet18. The linear layer and the normalization layer map the encoder output to a 128-dimensional hidden layer hypersphere space. The activation function is set to the ReLU function. After the virtual normal samples and virtual abnormal samples are input into the network, a hidden layer feature with a dimension of 128 is obtained. Based on the hidden layer features, a contrast loss based on normal sample aggregation and abnormal sample separation is calculated. The network is trained using the gradient descent method. The loss converges after 100 cycles of iterative training, and the model parameters are saved.

[0009] d. Input the memorized sample set into the trained deep contrastive learning network and calculate the minimum cosine distance between each sample and the remaining sample set in the hidden layer hypersphere space as the anomaly score, thereby determining the detection threshold;

[0010] e. Input the samples in the test sample set into the trained deep contrastive learning network, calculate the minimum cosine distance between the test samples in the test sample set and all the memory samples in the memory sample set as the anomaly score, and determine whether a fault exists based on the comparison between the anomaly score and the detection threshold.

[0011] In the gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network, in step a, the sample length after time series segmentation is 1024 data points, the amplitude is normalized to the range of [-1, 1], 2700 samples are selected from the normal samples as the training sample set, 900 samples are selected as the memory sample set, and then 900 normal samples are selected to form a test sample set together with three types of fault types, 900 fault samples of each type.

[0012] In the gearbox variable operating condition fault detection method of the normal sample aggregation contrast learning network, in step b, the offset transformation is smoothly achieved by randomly disrupting and reorganizing the arrangement of the signal. First, the original signal x is evenly divided into N segments. , disrupt and reorganize the order of each segment to form the offset transformed signal ,in .

[0013] In the gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network, in step b, the identity preservation transformation is performed twice on the original signal and the signal after the offset transformation, respectively. The original signal will generate a pair of virtual normal samples, and the signal after the offset transformation will generate a pair of virtual abnormal samples.

[0014] In the gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network, the identity-preserving transformation includes the following steps:

[0015] Step 1: Add a random sample from a Gaussian distribution with a mean of 0 and a standard deviation of 0.01 to each sample point of the signal;

[0016] Step 2: Multiply each sample point by a value randomly sampled from a Gaussian distribution with a mean of 1 and a standard deviation of 0.01.

[0017] Step 3. Finally, resample the signal with a ratio of 0.8 and fill it with zeros to return it to its original size.

[0018] In the gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network, in step c, the contrast loss consists of a normal aggregation term and an abnormal separation term, and the overall loss is:

[0019] ,

[0020] Where B is the batch size, and are the normal aggregation term and abnormal separation term of the i-th sample,

[0021] The normality aggregation term includes the contrast loss of each virtual normal sample, and the loss form is:

[0022]

[0023] in It is a unit pulse function, and its value is 1 if and only if the expression in the brackets holds, otherwise it is 0. and are respectively a set of virtual normal samples and a set of virtual abnormal samples, and are two virtual samples generated by the i-th sample through identity-preserving transformation, It is the classic SimCLR contrast loss, and its loss form is:

[0024]

[0025] in, Represent the input signal, the positive sample set of the input signal, and the negative sample set of the input signal, respectively. is the cardinality of the set, is the hidden layer representation of the input, is a temperature hyperparameter used to adjust the loss. T represents the vector transpose operator.

[0026] The anomaly separation term includes the contrast loss of each virtual anomaly sample, and the loss form is:

[0027] .

[0028] in It is a unit pulse function, and its value is 1 if and only if the expression in the brackets holds, otherwise it is 0. and are respectively a set of virtual normal samples and a set of virtual abnormal samples, and are two virtual samples generated by the i-th sample through identity-preserving transformation, It is the classic SimCLR contrast loss.

[0029] In the gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network, in step d, the abnormal score of the memory sample is the minimum cosine distance between the sample and the rest of the memory sample set, and is calculated as follows:

[0030] ,

[0031] Among them, x is the sample to be calculated, is a set of memory samples, represents the l2 norm of the vector, is the hidden layer representation of the input, Indicates the maximum value of the calculated results.

[0032] In the gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network, in step d, the detection threshold is the 95th percentile of the abnormality scores of all memory samples.

[0033] In the gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network, in step e, the calculation method of the test sample anomaly score is the same as the calculation method of the memory sample anomaly score.

[0034] In the gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network, for a continuously running variable speed gearbox, if the anomaly score calculated by inputting the collected samples into the deep contrast learning network exceeds the detection threshold for three consecutive times, it is considered that the variable speed gearbox has a fault.

[0035] Compared with existing technologies, this invention offers the following advantages: It significantly improves the representation capability of vibration signals under different operating conditions; generates virtual samples and designs a contrast loss with the goal of aggregating virtual normal samples and separating virtual abnormal samples, learning an ideal representation space for anomaly detection and clustering normal representations under different operating conditions. Finally, it determines anomaly scores through distance calculation, improving the accuracy and versatility of gearbox anomaly detection under varying operating conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are intended only to illustrate preferred embodiments and are not to be construed as limiting the present invention. It should be understood that the drawings described below are merely examples of the present invention, and that those skilled in the art will be able to derive other drawings from these drawings without inventive effort. Throughout the drawings, identical reference numerals are used to denote identical components.

[0037] In the attached figure:

[0038] Figure 1 This is a flowchart of a gearbox variable operating condition fault detection method based on a normal sample aggregation contrast learning network provided by an embodiment of the present disclosure;

[0039] Figure 2 This is a schematic diagram of three types of fault states of a planetary gearbox provided by an embodiment of the present disclosure;

[0040] FIG3 (a) to FIG3 (d) are schematic diagrams showing measured vibration signals and speed changes of a planetary gearbox under different health states according to an embodiment of the present disclosure;

[0041] Figure 4 is a schematic diagram of a performance comparison of a gearbox fault detection method provided by an embodiment of the present disclosure;

[0042] Figure 5 1 is a schematic diagram of abnormal scores of samples in different health states under variable speeds provided by an embodiment of the present disclosure.

[0043] The present invention will be further explained below with reference to the accompanying drawings and embodiments. DETAILED DESCRIPTION

[0044] Specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although specific embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0045] It should be noted that certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different nouns to refer to the same component. This specification and claims do not use the difference in nouns as a way to distinguish components, but use the difference in the functions of the components as the criterion for distinction. As mentioned throughout the specification and claims, "including" or "comprising" is an open term, so it should be interpreted as "including but not limited to". The subsequent description of the specification is a preferred embodiment of the present invention, but the description is based on the general principles of the specification and is not intended to limit the scope of the invention. The scope of protection of the present invention shall be as defined in the attached claims.

[0046] To facilitate understanding of the embodiments of the present invention, further explanation will be given below using specific embodiments as examples in conjunction with the accompanying drawings, and the accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0047] like Figures 1 to 5 As shown in FIG, the gearbox variable operating condition fault detection method based on the normal sample aggregation contrast learning network includes the following steps:

[0048] a. Use a vibration sensor to collect vibration signals from the gearbox under varying speed conditions. The signals are then segmented into fixed-length samples and normalized to construct training, memory, and test sample sets.

[0049] b. Continuously perform offset transformation and identity-preserving transformation on the samples in the training set to generate a set of virtual normal samples and virtual abnormal samples;

[0050] c. Establish a deep contrastive learning network, which includes an encoder, a linear layer, and a normalization layer, wherein the encoder is selected as a one-dimensional ResNet18, the linear layer and the normalization layer map the encoder output to a 128-dimensional hidden layer hypersphere space, and the activation function is set to the ReLU function. After the virtual normal samples and the virtual abnormal samples are input into the network, a hidden layer feature with a dimension of 128 is obtained. Based on the hidden layer features, a contrast loss based on normal sample aggregation and abnormal sample separation is calculated. The network is trained using the gradient descent method. After 100 cycles of iterative training, the loss converges and the model parameters are saved. The deep contrastive learning network model training includes the following steps:

[0051] A one-dimensional ResNet18 network is used as the encoder, and virtual normal samples and virtual abnormal samples with a sample length of 1024 are input into the encoder to obtain features with a dimension of 512. Subsequently, a linear layer and a normalization layer with a dimension of 128 are passed, and the activation function is set to the ReLU function to obtain the hidden layer features of the virtual normal samples and virtual abnormal samples.

[0052] The overall contrast loss is calculated based on the hidden layer features. The overall contrast loss is used as the objective function. The gradient descent method is used to train the network. After 100 cycles of iterative training, the loss converges and the model parameters are saved.

[0053] d. Input the memorized sample set into the trained deep contrastive learning network and calculate the minimum cosine distance between each sample and the remaining sample set in the hidden layer hypersphere space as the anomaly score, thereby determining the detection threshold;

[0054] e. Input the samples in the test sample set into the trained deep contrastive learning network, calculate the minimum cosine distance between the test samples in the test sample set and all the memory samples in the memory sample set as the anomaly score, and determine whether a fault exists based on the comparison between the anomaly score and the detection threshold.

[0055] In a preferred embodiment of the gearbox variable operating condition fault detection method of the normal sample aggregation contrast learning network, in step a, the sample length after time series segmentation is 1024 data points, the amplitude is normalized to the range of [-1,1], 2700 samples are selected from the normal samples as the training sample set, 900 samples are selected as the memory sample set, and then 900 normal samples are selected to form a test sample set together with three types of fault types, 900 fault samples of each type.

[0056] In a preferred embodiment of the gearbox variable operating condition fault detection method of a normal sample aggregation contrast learning network, in step b, the offset transformation is smoothly achieved by randomly disrupting and reorganizing the arrangement of the signal. First, the original signal x is evenly divided into N segments. , disrupt and reorganize the order of each segment to form the offset transformed signal ,in .

[0057] In a preferred embodiment of the gearbox variable operating condition fault detection method of the normal sample aggregation contrast learning network, in step b, the identity preservation transformation is performed twice on the original signal and the offset transformed signal respectively, and the original signal will generate a pair of virtual normal samples, and the offset transformed signal will generate a pair of virtual abnormal samples.

[0058] In a preferred embodiment of the gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network, the identity-preserving transformation includes the following steps:

[0059] Step 1: Add a random sample from a Gaussian distribution with a mean of 0 and a standard deviation of 0.01 to each sample point of the signal;

[0060] Step 2: Multiply each sample point by a value randomly sampled from a Gaussian distribution with a mean of 1 and a standard deviation of 0.01.

[0061] Step 3. Finally, resample the signal with a ratio of 0.8 and fill it with zeros to return it to its original size.

[0062] In a preferred embodiment of the gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network, in step c, the contrast loss consists of a normal aggregation term and an abnormal separation term, and the overall loss is:

[0063] ,

[0064] Where B is the batch size, and are the normal aggregation term and abnormal separation term of the i-th sample,

[0065] The normality aggregation term includes the contrast loss of each virtual normal sample, and the loss form is:

[0066]

[0067] in It is a unit pulse function, and its value is 1 if and only if the expression in the brackets holds, otherwise it is 0. and are respectively a set of virtual normal samples and a set of virtual abnormal samples, and are two virtual samples generated by the i-th sample through identity-preserving transformation, It is the classic SimCLR contrast loss, and its loss form is:

[0068]

[0069] in, Represent the input signal, the positive sample set of the input signal, and the negative sample set of the input signal, respectively. is the cardinality of the set, is the hidden layer representation of the input, is a temperature hyperparameter used to adjust the loss. T represents the vector transpose operator.

[0070] The anomaly separation term includes the contrast loss of each virtual anomaly sample, and the loss form is:

[0071] .

[0072] in It is a unit pulse function, and its value is 1 if and only if the expression in the brackets holds, otherwise it is 0. and are respectively a set of virtual normal samples and a set of virtual abnormal samples, and are two virtual samples generated by the i-th sample through identity-preserving transformation, It is the classic SimCLR contrast loss.

[0073] In a preferred embodiment of the gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network, in step d, the abnormal score of the memory sample is the minimum cosine distance between the sample and the rest of the memory sample set, and is calculated as follows:

[0074] ,

[0075] Among them, x is the sample to be calculated, is a set of memory samples, represents the l2 norm of the vector, is the hidden layer representation of the input, Indicates the maximum value of the calculated results.

[0076] In a preferred embodiment of the gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network, in step d, the detection threshold is the 95% quantile of the abnormality scores of all memory samples.

[0077] In a preferred embodiment of the gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network, in step e, the calculation method of the test sample abnormality score is the same as the calculation method of the memory sample abnormality score.

[0078] In a preferred embodiment of the gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network, for a continuously running variable speed gearbox, if the abnormality score calculated by inputting the collected samples into the deep contrast learning network exceeds the detection threshold for three consecutive times, it is considered that the variable speed gearbox has a fault.

[0079] In one embodiment, a method for detecting gearbox variable operating condition faults using a normal sample aggregation contrast learning network includes:

[0080] In the first step, a vibration sensor was used to collect vibration signals from the gearbox under varying speed conditions, including normal state (NS), broken tooth (BT), root crack (RC), and surface groove (SG). The sampling frequency was 10240 Hz, and the sample length was 1024. The signal time series was segmented into samples of length 1024, and the amplitudes were normalized to the range [-1, 1]. 2700 samples were selected from the normal samples as the training set, 900 samples were selected as the memory set, and 900 normal samples were selected to form the test set, along with 900 samples from each of the three fault types.

[0081] In the second step, the samples in the training set are continuously shifted and transformed to generate a set of virtual normal samples and virtual abnormal samples. The shift transformation is smoothly implemented by randomly shuffling and reorganizing the arrangement of the signal. First, the original signal x is evenly divided into N segments. , disrupt and reorganize the order of each segment to form the offset transformed signal ,in Subsequently, two identity-preserving transformations are performed on the original signal and the shifted signal, respectively. These transformations include the following steps: adding a value randomly sampled from a Gaussian distribution with a mean of 0 and a standard deviation of 0.01 to each sample point of the signal; multiplying each sample point by a value randomly sampled from a Gaussian distribution with a mean of 1 and a standard deviation of 0.01; and resampling the signal at a ratio of 0.8 and padding it with zeros to its original size. After continuously performing the shift and identity-preserving transformations, the original signal will generate a pair of virtual normal samples and a pair of virtual abnormal samples, which can be used as comparison examples for subsequent comparative learning.

[0082] In the third step, a deep contrastive learning network is established, consisting of an encoder, a linear layer, and a normalization layer. The encoder is a one-dimensional ResNet18, and the linear and normalization layers map the encoder output to a 128-dimensional hidden hypersphere space. Subsequently, based on the representation of the generated virtual normal samples and virtual abnormal samples in the hidden space, a contrastive loss based on normal sample aggregation and abnormal sample separation is calculated. This loss includes a normal aggregation term and an abnormal separation term, and its overall form is:

[0083]

[0084] Where B is the batch size, and are the normality aggregation term and anomaly separation term for the i-th sample, respectively. In the representation space, the ideal distribution is that normal samples are clustered together, while different types of faults are discretely distributed in different areas. Therefore, we hope that virtual normal samples are pulled together, while virtual anomaly samples are repelled from normal samples and other types of anomaly samples. The normality aggregation term includes the contrast loss of each virtual normal sample, and the loss form is:

[0085]

[0086] The anomaly separation term includes the contrast loss of each virtual anomaly sample, and the loss form is:

[0087]

[0088] in It is a unit pulse function, and its value is 1 if and only if the expression in the brackets holds, otherwise it is 0. and are respectively a set of virtual normal samples and a set of virtual abnormal samples, and are two virtual samples generated by the i-th sample through identity-preserving transformation, It is the classic SimCLR contrast loss, and its loss form is:

[0089]

[0090] in, Represent the input signal, the positive sample set of the input signal, and the negative sample set of the input signal, respectively. is the cardinality of the set, is the hidden layer representation of the input, is a temperature hyperparameter used to adjust the loss. T represents the vector transpose operator. Since the positive sample pairs are in the numerator and the negative sample pairs are in the denominator, minimizing the normality aggregation term can shorten the distance between virtual normal samples, while minimizing the anomaly separation term can increase the distance between each type of virtual anomaly sample and the virtual normal samples and the remaining virtual anomaly samples, thereby forming a good anomaly detection representation space. After calculating the overall contrastive loss, the contrastive learning network is trained by minimizing the contrastive loss. When the overall contrastive loss converges, the network parameters are fixed.

[0091] In the fourth step, the memory sample set is input into the trained network, and the minimum cosine distance between each sample and the remaining sample set is calculated in the hidden space as the anomaly score, which is calculated as follows;

[0092]

[0093] Among them, x is the sample to be calculated, is a set of memory samples, represents the l2 norm of the vector, is the hidden layer representation of the input, Represents the maximum value of the calculated results. Since all the memory samples are normal samples, they follow the characteristic distribution of normal samples and have a small minimum cosine distance from the remaining memory samples. Therefore, after the anomaly score calculation is completed, the detection threshold is selected as the 95th percentile of the anomaly scores of all memory samples.

[0094] In the fifth step, samples from the test set are fed into the trained network. The minimum cosine distance between the test sample and all memorized samples is calculated as an anomaly score. The presence of a fault is determined by comparing the anomaly score with the detection threshold. To improve the robustness of the detection results, for continuously operating variable-speed gearboxes, a gearbox fault is considered if the calculated anomaly score exceeds the detection threshold three times in a row.

[0095] Figure 2Three types of faulty gears in addition to the normal state in the gearbox fault pre-setting experiment are displayed, including broken teeth, tooth root cracks and surface grooves.

[0096] Figures 3(a) through 3(d) illustrate the vibration signal variations with speed for a planetary gearbox in four healthy states: normal, broken tooth, tooth root crack, and surface groove. It is clear that the amplitude of the vibration signal changes with speed, with higher speeds increasing the amplitude. Therefore, the signal is affected by both speed and health, making it difficult to determine whether a gearbox fault has occurred based solely on the time domain signal. Variable-condition fault detection requires a model to extract fault-related features to produce accurate detection results.

[0097] Figure 4 The figure below is a schematic diagram comparing the performance of the gearbox fault detection method according to an embodiment of the present invention. The evaluation metrics for the detection results include time-to-test ratio (TPR), precision ratio (TPR), precision ratio (TPR), and area under the receiver operating characteristic (AUROC). The experimental results are compared with two models: an autoencoder and a traditional contrastive learning method. The experimental results also show the means and standard deviations of the four detection metrics achieved by each model across five experiments, as shown in Table 1.

[0098]

[0099] In this embodiment, normal sample aggregation comparative learning shortens the distance between normal sample features and learns a more representative normal sample feature distribution under variable working conditions, thereby achieving fault detection. Figure 4 As can be seen from the results in Table 1, the proposed normal sample aggregation contrast learning has achieved the best performance in all detection indicators, and the stability of multiple experimental results is good, which verifies the superiority of the method.

[0100] Figure 5 A diagram shows the anomaly scores for samples in different health states under varying speeds. It can be seen that despite the varying speeds, the threshold determined by the anomaly scores of the memorized samples effectively separates normal and faulty test samples. Normal samples are generally below the threshold, while all types of faulty samples are generally above the threshold, demonstrating that the proposed method can effectively detect gearbox faults under varying operating conditions.

[0101] Although the embodiments of the present invention have been described above with reference to the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments and application fields. The above-mentioned specific embodiments are merely illustrative and instructive, and are not restrictive. A person skilled in the art, guided by this specification and without departing from the scope of protection of the claims of the present invention, may also devise various forms, all of which fall within the scope of protection of the present invention.

Claims

1. A gearbox variable operating condition fault detection method based on a normal sample aggregation contrast learning network, characterized in that: The steps include: a. Use a vibration sensor to collect vibration signals from the gearbox under varying speed conditions. The signals are then segmented into fixed-length samples and normalized to construct training, memory, and test sample sets. b. Continuously perform offset transformation and identity-preserving transformation on the samples in the training sample set to generate a set of virtual normal samples and a set of virtual abnormal samples; c. Establish a deep contrastive learning network, which includes an encoder, a linear layer, and a normalization layer. The encoder is selected as a one-dimensional ResNet18. The linear layer and the normalization layer map the encoder output to a 128-dimensional hidden layer hypersphere space. The activation function is set to the ReLU function. After the virtual normal sample set and the virtual abnormal sample set are input into the deep contrastive learning network, a hidden layer feature with a dimension of 128 is obtained. Based on the hidden layer features, a contrast loss based on normal sample aggregation and abnormal sample separation is calculated. The deep contrastive learning network is trained using the gradient descent method. The loss converges after 100 cycles of iterative training, and the model parameters are saved. d. Input the samples in the memory sample set into the trained deep contrastive learning network, calculate the minimum cosine distance between each memory sample and the remaining memory sample set as the memory sample anomaly score, and then determine the detection threshold; e. Input the samples in the test sample set into the trained deep contrastive learning network, calculate the minimum cosine distance between the test sample in the test sample set and all the memory samples in the memory sample set as the test sample anomaly score, and determine whether a fault exists based on the comparison between the test sample anomaly score and the detection threshold; In step b, the offset transformation is smoothly implemented by randomly scrambling and reorganizing the arrangement of the signal. First, the original signal x is evenly divided into N segments. , disrupt and reorganize the order of each segment to form the offset transformed signal ,in ; In step b, the identity-preserving transformation is performed twice on the original signal and the signal after the shift transformation, respectively. The original signal will generate a pair of virtual normal sample sets, and the signal after the shift transformation will generate a pair of virtual abnormal sample sets.

2. The gearbox variable operating condition fault detection method based on a normal sample aggregation contrast learning network according to claim 1 is characterized in that: In step a, the sample length after time series segmentation is 1024 data points, the amplitude is normalized to the range of [-1, 1], 2700 samples are selected from the normal samples as the training sample set, 900 samples are selected as the memory sample set, and then 900 normal samples are selected to form the test sample set together with 900 fault samples of each of the three types of faults.

3. The gearbox variable operating condition fault detection method based on a normal sample aggregation contrast learning network according to claim 1 is characterized in that: The conversion includes the following steps: Step 1: Add a random sample from a Gaussian distribution with a mean of 0 and a standard deviation of 0.01 to each sample point of the signal; Step 2: Multiply each sample point by a value randomly sampled from a Gaussian distribution with a mean of 1 and a standard deviation of 0.

01. Step 3. Finally, resample the signal with a ratio of 0.8 and fill it with zeros to return it to its original size.

4. The gearbox variable operating condition fault detection method based on a normal sample aggregation contrast learning network according to claim 1 is characterized in that: In step c, the contrast loss consists of a normal aggregation term and an abnormal separation term, and the overall loss is: ， Where B is the batch size, and are the normal aggregation term and abnormal separation term of the i-th sample, The normal aggregation term includes the contrast loss of each set of virtual normal samples, and the loss form is: ， in It is a unit pulse function, and its value is 1 if and only if the expression in the brackets holds, otherwise it is 0. and are respectively a set of virtual normal samples and a set of virtual abnormal samples, and are two virtual samples generated by the i-th sample through identity-preserving transformation, It is the classic SimCLR contrast loss, and its loss form is: ， in, Represent the input signal, the positive sample set of the input signal, and the negative sample set of the input signal, respectively. is the cardinality of the set, is the hidden layer representation of the input, Is a temperature hyperparameter used to adjust the loss, T represents the vector transpose operator, The anomaly separation term includes the contrast loss of each set of virtual anomaly samples, and the loss form is: ， in It is a unit pulse function, and its value is 1 if and only if the expression in the brackets holds, otherwise it is 0. and are respectively a set of virtual normal samples and a set of virtual abnormal samples, and are two virtual samples generated by the i-th sample through identity-preserving transformation, It is the classic SimCLR contrast loss.

5. The gearbox variable operating condition fault detection method based on a normal sample aggregation contrast learning network according to claim 1 is characterized in that: In step d, the anomaly score of the memory sample is the minimum cosine distance between the memory sample and the rest of the memory sample set, and is calculated as follows: ， Among them, x is the memory sample to be calculated, is a set of memory samples, represents the l2 norm of the vector, is the hidden layer representation of the input, Indicates the maximum value of the calculated results.

6. The gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network according to claim 1 is characterized in that: In step d, the detection threshold is the 95th percentile of the abnormality scores of all memory samples.

7. The gearbox variable operating condition fault detection method based on a normal sample aggregation contrast learning network according to claim 1 is characterized in that: In step e, the calculation method of the test sample anomaly score is the same as the calculation method of the memory sample anomaly score.

8. The gearbox variable operating condition fault detection method using a normal sample aggregation contrast learning network according to claim 1 is characterized in that: For a continuously running variable speed gearbox, if the collected samples are input into the deep contrast learning network and the test sample anomaly score calculated exceeds the detection threshold three times in a row, the variable speed gearbox is considered to be faulty.

Images