Detection system and method

Abstract

A detection system and method are provided. The detection system receives a plurality of sensing samples. The detection system performs a filtering operation on the sensing samples. The detection system performs a dimension reduction operation on the sensing samples by a pre-trained dimension reduction model to generate a dimension-reduced feature corresponding to each sensing sample. The detection system inputs the dimension-reduced feature of each sensing sample to a pre-trained clustering model to calculate a likelihood score corresponding to each sensing sample. The detection system determines whether a speed reducer is abnormal according to a standard deviation threshold and the likelihood scores.

Description

Technical Field

[0001] This invention relates to a detection system and method. Specifically, this invention relates to a system and method for detecting whether a speed reducer is malfunctioning. Background Technology

[0002] In today's smart manufacturing, ensuring the normal operation of robotic arms is a crucial issue. Statistics show that speed reducers are among the components most prone to malfunctions in robotic arms. Therefore, a method for quickly identifying speed reducer abnormalities is needed to prevent these malfunctions from impacting the robotic arm's production efficiency.

[0003] Current analytical methods for determining whether a robotic arm is malfunctioning require the long-term accumulation of a large amount of data from the robotic arm, such as maximum signal values, overall OA (overall value), and RMS (response time and strength). Furthermore, it necessitates collecting both normal and abnormal data from the robotic arm, identifying data trends, and having experts define the upper and lower limits of abnormal values ​​before the data can be used in practical applications. However, the statistical information from current analytical methods is easily affected by environmental conditions or unexpected events (e.g., vibrations from collisions with other machines). Therefore, it requires accumulating a large amount of normal and abnormal data before it can be applied, and human experience is needed to define the upper and lower limits of abnormal values, making it difficult to quickly determine whether a robotic arm is malfunctioning.

[0004] Furthermore, traditional robotic arm monitoring methods collect a large amount of normal and abnormal data values ​​and judge whether there is an anomaly by comparing trends. Due to the complexity of data and signal characteristics, they can usually only monitor rotary robotic arms that operate at a fixed speed, and lack detection methods for non-rotary robotic arms with variable speed (i.e., non-uniform speed) (such as: Selective Compliance Articulated Robot Arm (SCARA)).

[0005] Therefore, how to efficiently and automatically determine whether the reducer of a robotic arm is malfunctioning is a goal that the industry urgently needs to strive for. Summary of the Invention

[0006] One object of the present invention is to provide a detection system. The system includes a memory and a processor electrically connected to the memory. The memory stores a pre-trained dimensionality reduction model and a pre-trained clustering model, wherein the pre-trained dimensionality reduction model performs a dimensionality reduction operation based on multiple normal samples to calculate a likelihood score. The processor receives multiple sensed samples. The processor performs a filtering operation on the sensed samples. The processor performs the dimensionality reduction operation on the sensed samples using the pre-trained dimensionality reduction model to generate a dimensionality reduction feature corresponding to each sensed sample. The processor inputs the dimensionality reduction feature of each sensed sample into the pre-trained clustering model to calculate the likelihood score corresponding to each sensed sample. The processor determines whether a speed reducer is malfunctioning based on a standard deviation threshold and the likelihood scores.

[0007] Another object of the present invention is to provide a detection method. The method is used in a detection system comprising a memory and a processor. The memory stores a pre-trained dimensionality reduction model and a pre-trained clustering model, wherein the pre-trained dimensionality reduction model is used to perform a dimensionality reduction operation, and the pre-trained clustering model is generated based on multiple normal samples to calculate a likelihood score. The detection method is executed by the processor and includes the following steps: receiving multiple sensed samples; performing a filtering operation on the sensed samples; performing the dimensionality reduction operation on the sensed samples using the pre-trained dimensionality reduction model to generate a dimensionality reduction feature corresponding to each sensed sample; inputting the dimensionality reduction feature of each sensed sample into the pre-trained clustering model to calculate the likelihood score corresponding to each sensed sample; and determining whether a speed reducer is malfunctioning based on a standard deviation threshold and the likelihood scores.

[0008] The detection technology (including at least an apparatus and a method) provided by this invention collects normal sample data (each normal sample being the value of a robotic arm performing a complete variable-speed non-rotational motion within a standard cycle) using sensors. A pre-trained dimensionality reduction model and a pre-trained clustering model are first established. During the actual judgment phase, multiple sensing samples collected from the reducer under test are received and filtered. Next, the pre-trained dimensionality reduction model is used to reduce the dimensionality of these sensing samples, generating dimensionality reduction features corresponding to each sample. Subsequently, the dimensionality reduction features of each sensing sample are input into the pre-trained clustering model to calculate the likelihood score for each sample. Finally, based on a standard deviation threshold and the likelihood scores, it is determined whether the reducer under test is malfunctioning. The detection technology provided by this invention, through the above operation, quickly determines whether the reducer under test is malfunctioning, solving the problems of existing technologies.

[0009] The following detailed description of the technology and embodiments of the present invention, in conjunction with the accompanying drawings, will enable those skilled in the art to understand the technical features of the claimed invention. Attached Figure Description

[0010] Figure 1 This is a schematic diagram depicting the architecture of the detection system according to the first embodiment of the present invention;

[0011] Figure 2A This is a schematic diagram depicting a set of normal samples in the first embodiment;

[0012] Figure 2B This is a schematic diagram depicting a normal sample after filtering in the first embodiment;

[0013] Figure 3 This is a schematic diagram depicting a normal sample after dimensionality reduction in the first embodiment;

[0014] Figure 4 This is a partial flowchart depicting the detection method of the second embodiment;

[0015] Figure 5 This is a partial flowchart depicting the detection method of the second embodiment; and

[0016] Figure 6 This is a partial flowchart depicting the detection method of the second embodiment.

[0017] [Symbol Explanation]

[0018] 1: Detection System

[0019] 11: Memory

[0020] 13: Send / Receive Interface

[0021] 15: Processor

[0022] 111: Pre-trained dimensionality reduction model

[0023] 113: Pre-trained clustering model

[0024] 133: Sensing Samples

[0025] S401-S409: Steps

[0026] S501-S505: Steps

[0027] S601-S607: Steps Detailed Implementation

[0028] The detection system and method provided by the present invention will be explained below through embodiments. However, these embodiments are not intended to limit the implementation of the present invention to any environment, application, or manner described in these embodiments. Therefore, the description of the embodiments is for illustrative purposes only and is not intended to limit the scope of the present invention. It should be understood that in the following embodiments and drawings, elements not directly related to the present invention have been omitted and are not shown, and the dimensions of each element and the dimensional proportions between elements are merely illustrative and are not intended to limit the scope of the present invention.

[0029] The first embodiment of the present invention is a detection system 1, the schematic diagram of which is depicted in Figure 1 In this embodiment, the detection system 1 includes a memory 11, a transceiver interface 13, and a processor 15, with the processor 15 electrically connected to the memory 11 and the transceiver interface 13. The memory 11 may be a memory, a Universal Serial Bus (USB) disk, a hard disk, an optical disk, a USB flash drive, or any other storage medium or circuit known to those skilled in the art and having the same function. The transceiver interface 13 is an interface capable of receiving and transmitting data, or other interfaces capable of receiving and transmitting data known to those skilled in the art. The transceiver interface 13 can receive data from sources such as external devices, external web pages, external applications, etc. The processor 15 may be various processing units, a central processing unit (CPU), a microprocessor, or other computing devices known to those skilled in the art. In some embodiments, the detection system 1 may be, but is not limited to, a mobile electronic device, a desktop computer, a portable computer, etc.

[0030] The operation of the first embodiment of the present invention will be briefly described first. The processor 15 first receives multiple sensing samples 133 collected from the reducer under test via the transceiver interface 13. Next, to filter out noise, the processor 15 first performs a filtering operation on the sensing samples 133. Then, the processor 15 performs dimensionality reduction on the sensing samples 133 using a pre-trained dimensionality reduction model 111 to generate dimensionality reduction features corresponding to each sensing sample 133. Subsequently, the processor 15 inputs the dimensionality reduction features of each sensing sample 133 into the pre-trained clustering model 113 to calculate the likelihood score corresponding to each sensing sample 133. Finally, the processor 15 determines whether the reducer under test is malfunctioning based on a standard deviation threshold and the likelihood scores. Users can quickly determine whether the reducer under test is malfunctioning based on the operation of the detection system 1 described above. The following paragraphs will describe in detail the implementation details related to the present invention.

[0031] In this embodiment, as Figure 1As shown, memory 11 stores a pre-trained dimensionality reduction model 111 and a pre-trained clustering model 113. The pre-trained dimensionality reduction model 111 is used to perform dimensionality reduction operations, and the pre-trained clustering model 113 is generated based on multiple normal samples and used to calculate the likelihood score. It should be noted that the pre-trained dimensionality reduction model 111 and the pre-trained clustering model 113 can be built by the detection system 1 itself, or they can be directly received from an external device after training. The construction method and content of the pre-trained dimensionality reduction model 111 and the pre-trained clustering model 113 will be described in detail in later paragraphs.

[0032] For ease of understanding, the following paragraphs will be described using the pre-trained dimensionality reduction model 111 and the pre-trained clustering model 113 built by the detection system 1 itself. Those skilled in the art should be able to understand how other external devices can generate the same operation based on the description, which will not be elaborated here.

[0033] First, we will explain how to collect training data for training the pre-trained dimensionality reduction model 111 and the pre-trained clustering model 113. In this embodiment, the detection system 1 primarily uses at least one sensor to sense the speed reducer operating in a variable-speed non-rotational motion, generating various sensing samples, wherein the data length of each sensing sample corresponds to a standard cycle of the variable-speed non-rotational motion. It should be noted that during the training phase, the detection system 1 collects normal samples generated by the sensor for a normal speed reducer. In some embodiments, the at least one sensor may be directly connected to the detection system 1, or the normal samples may be collected by other devices and then transmitted to the detection system 1.

[0034] It should be noted that in traditional methods, measurements are typically taken only for robotic arms performing constant-speed rotary movements (i.e., rotating in the same direction), and both normal and abnormal data need to be collected for trend comparison and judgment. Unlike traditional methods, in this invention, variable-speed non-rotational movements refer to robotic arms (e.g., horizontally articulated robotic arms) operating at varying speeds in at least two directions (e.g., stopping after a half-turn or changing direction). Furthermore, this invention only requires collecting normal samples from the reducer of a normally functioning robotic arm. In addition, a standard cycle in this invention is the time it takes for the robotic arm to perform a complete variable-speed non-rotational movement (e.g., a non-uniform speed movement from left to right with maximum amplitude is a variable-speed non-rotational movement, while the blades of an electric fan are rotary). For example, if a complete variable-speed non-rotational movement with maximum amplitude from left to right takes 3 seconds, then the standard cycle for each normal sample is 3 seconds.

[0035] In some embodiments, the at least one sensor is an accelerometer (not shown) mounted on a base at a radial position of the reducer of a robotic arm to collect multiple vibration signals. In some embodiments, the at least one sensor is a directional microphone (not shown) mounted around the periphery of the reducer of a robotic arm to collect multiple sound signals. The directional microphone is pointed at the reducer to pick up sound.

[0036] For ease of understanding, let's take a directional microphone as an example, using a sensor as an example. Figure 2A This example illustrates a normal sample (i.e., an audio signal) generated after continuous sound sensing by a directional microphone, where the horizontal axis represents the number of data points and the vertical axis represents the amplitude value, recording the amplitude value corresponding to each data point. For example... Figure 2A As shown, the normal sample is an audio signal generated with a sampling rate of 12800 Hz and a standard period of 3 seconds, therefore, the normal sample has a total of 38400 data points. It should be noted that when the sensor is an accelerometer, since it also detects the amplitude signal from the reducer, those skilled in the art should understand the implementation method when the sensor is an accelerometer based on the above description, and will not be elaborated further.

[0037] The following explains how processor 15 generates a pre-trained dimensionality reduction model 111 based on the normal samples collected after the aforementioned operations. First, to filter out unnecessary interference noise in the environment, processor 15 performs this filtering operation on the normal samples. For example, Figure 2B In response to Figure 2A An example diagram showing the result of performing a bandpass filter from 300 Hz to 400 Hz (i.e., retaining only the signal from 300 Hz to 400 Hz). It should be noted that those skilled in the art will understand that the aforementioned noise reduction processing can be implemented not only by the processor 15 in software, but also by a bandpass filter.

[0038] Next, processor 15 inputs these normal samples into an autoencoder model constructed by a convolutional neural network (e.g., a one-dimensional convolutional neural network) to train the autoencoder model for dimensionality reduction. The autoencoder model includes an encoder and a decoder. Finally, processor 15 sets the encoder portion of the trained autoencoder model as the pre-trained dimensionality reduction model 111. The pre-trained dimensionality reduction model 111 can extract features from the input signal information and generate dimensionality-reduced data. For example, using... Figure 2BFor example, processor 15 inputs a set of samples with 38,400 data points (i.e., can be considered as a set of normal samples with 38,400 dimensions) into the pre-trained dimensionality reduction model 111, and generates dimensionality reduction features corresponding to the normal samples with only 32 dimensions, such as... Figure 3 As shown.

[0039] It should be noted that a convolutional neural network may also include other layers such as pooling layers to acquire features and perform dimensionality reduction. However, the focus of this invention is on sample computation and analysis, so only implementation details related to this invention will be described. Specifically, in the encoding operation part of the autoencoder model constructed by the convolutional neural network, pooling layers (e.g., global pooling layers) can be added after each convolutional layer of the convolutional neural network to perform pooling operations on the convolutional results of each layer, thereby reducing the dimensionality of the features generated by each convolutional layer.

[0040] The following explains how processor 15 generates a pre-trained clustering model 113 based on dimensionality reduction features. First, processor 15 inputs the dimensionality reduction features of each normal sample into the clustering model to train it. This clustering model performs a clustering operation on multiple input data points to generate a likelihood score corresponding to each input data point. Then, processor 15 uses the trained clustering model as the pre-trained clustering model 113. In some embodiments, the clustering model can be, but is not limited to, a model that performs a K-means algorithm, a Gaussian Mixture Model, or any known model that performs a clustering algorithm and can generate corresponding cluster likelihood scores for the input data.

[0041] It should be noted that after inputting multiple unknown samples into the pre-trained clustering model 113, multiple likelihood scores corresponding to the unknown samples can be generated. Since the pre-trained clustering model 113 is trained from multiple normal samples, the processor 15 can determine the similarity between the unknown sample and the normal sample (i.e., whether the unknown sample will be classified in the same group as the normal sample) through the likelihood scores of the unknown samples. If the scores are too far apart, it means that the unknown sample may not belong to the signal range of the normal sample and has an abnormal signal.

[0042] The preceding paragraphs have described the methods for generating the pre-trained dimensionality reduction model 111 and the pre-trained clustering model 113. The following paragraphs continue to describe the operation of the first embodiment of the present invention (i.e., the actual judgment stage). Please refer to [link / reference]. Figure 1 .

[0043] In this embodiment, the processor 15 receives multiple sensing samples 133 collected from the reducer under test via the transceiver interface 13. It should be noted that the method of collecting the sensing samples 133 is the same as the method of collecting training data described in the previous paragraph, and will not be repeated here.

[0044] Next, the processor 15 performs a filtering operation on the sensed samples 133. It should be noted that the processor 15 performs the filtering operation in the same manner as described in the previous paragraphs, and will not be repeated here.

[0045] Next, the processor 15 performs dimensionality reduction on the sensed samples 133 using a pre-trained dimensionality reduction model 111 to generate dimensionality-reduced features corresponding to each sensed sample 133. Then, the processor 15 inputs the dimensionality-reduced features of each sensed sample 133 into the pre-trained clustering model 113 to calculate the likelihood score for each sensed sample 133. Finally, the processor 15 determines whether the speed reducer under test is malfunctioning based on a standard deviation threshold and the likelihood scores. In one embodiment, when the likelihood score is lower than the standard deviation threshold, the sensed sample corresponding to the likelihood score is considered malfunctioning.

[0046] Alternatively, for example, when the cluster model 113 is a Gaussian mixture model, the standard deviation threshold can be set to plus or minus three standard deviations as the upper and lower bounds. When the likelihood score generated by the Gaussian mixture model based on the sensed sample is within this limit, it indicates that the sensed sample is a normal value. Conversely, when the likelihood score generated by the Gaussian mixture model based on the sensed sample exceeds this limit, it indicates that the sensed sample is an outlier. In some embodiments, the processor 15 further determines whether the gearbox under test is abnormal based on the proportion of samples in the sensed sample 133 that are judged to be outliers. For example, when more than 5% of the likelihood scores of the sensed sample 133 are evaluated as outliers, the processor 15 determines that the gearbox under test is abnormal.

[0047] The detection system 1 provided by this invention collects normal sample data (each normal sample being the value of a robotic arm performing a complete variable-speed non-rotational motion within a standard cycle) through sensors, and first establishes a pre-trained dimensionality reduction model 111 and a pre-trained clustering model 113. During the actual judgment stage, the detection system 1 receives multiple sensing samples 133 collected from the reducer under test and performs filtering on these sensing samples 133. Then, the pre-trained dimensionality reduction model 111 performs dimensionality reduction on these sensing samples 133 to generate dimensionality reduction features corresponding to each sensing sample 133. Subsequently, the dimensionality reduction features of each sensing sample 133 are input into the pre-trained clustering model 113 to calculate the likelihood score corresponding to each sensing sample 133. Finally, based on a standard deviation threshold and the likelihood scores, it is determined whether the reducer under test is malfunctioning. The detection system 1 provided by this invention, through the above operation, quickly determines whether the reducer under test is malfunctioning, solving the problems of the prior art.

[0048] The second embodiment of the present invention is a detection method, the flowchart of which is depicted in Figure 4 The detection method is used in a detection system, such as the detection system 1 described in the first embodiment. This detection system includes a memory, a transceiver interface, and a processor. The memory stores a pre-trained dimensionality reduction model and a pre-trained clustering model. The pre-trained dimensionality reduction model is used to perform a dimensionality reduction operation, and the pre-trained clustering model is generated based on multiple normal samples to calculate a likelihood score, such as the pre-trained dimensionality reduction model 111 and pre-trained clustering model 113 of the first embodiment. The detection method determines whether the reducer is malfunctioning through steps S401 to S409.

[0049] In some embodiments, the detection system further includes at least one sensor for sensing the sensing samples generated by the reducer during a variable-speed non-rotational operation, wherein the data length of each sensing sample corresponds to a standard cycle of the variable-speed non-rotational operation. In some embodiments, the at least one sensor is an accelerometer mounted on a base at a radial position of the reducer of a robotic arm for collecting multiple vibration signals. In some embodiments, the at least one sensor is a directional microphone mounted on a periphery of the reducer of a robotic arm for collecting multiple sound signals.

[0050] Steps S401 to S409 will be explained below. First, in step S401, the detection system receives multiple sensed samples. Next, in step S403, the detection system performs a filtering operation on these sensed samples. Subsequently, in step S405, the detection system performs dimensionality reduction on these sensed samples using the pre-trained dimensionality reduction model to generate a dimensionality reduction feature corresponding to each sensed sample.

[0051] Next, in step S407, the detection system inputs the dimensionality reduction features of each sensed sample into the pre-trained clustering model to calculate the likelihood score for each sensed sample. Finally, in step S409, the detection system determines whether a speed reducer has malfunctioned based on a standard deviation threshold and the likelihood scores.

[0052] In some implementations, such as Figure 5 As shown, the detection method of the present invention may further include steps S501 to S505. In step S501, the filtering operation is performed on the normal samples, wherein each normal sample is generated by at least one sensor detecting a normal speed reducer operating a variable-speed non-rotational motion. In step S503, the normal samples are input into an autoencoder model constructed by a convolutional neural network to train the autoencoder model, wherein the autoencoder model includes an encoding operation part and a decoding operation part. In step S505, the encoding operation part of the trained autoencoder model is set as the pre-trained dimensionality reduction model.

[0053] In some implementations, such as Figure 6 As shown, the detection method of the present invention may further include steps S601 to S607. In step S601, the filtering operation is performed on the normal samples, wherein each normal sample is generated by at least one sensor detecting a normal reducer operating a variable-speed non-rotational motion. In step S603, the dimensionality reduction operation is performed on the normal samples using the pre-trained dimensionality reduction model to generate dimensionality reduction features corresponding to each normal sample. In step S605, the dimensionality reduction features of each normal sample are input into a clustering model to train the clustering model, wherein the clustering model is used to perform a clustering operation on multiple input data to generate a likelihood score corresponding to each input data. In step S607, the trained clustering model is used as the pre-trained clustering model.

[0054] In addition to the steps described above, the second embodiment can also perform all the operations and steps of the detection system 1 described in the first embodiment, have the same function, and achieve the same technical effect. Those skilled in the art to which this invention pertains can directly understand how the second embodiment performs these operations and steps based on the first embodiment described above, has the same function, and achieves the same technical effect, so it will not be described in detail here.

[0055] In summary, the detection technology (including at least an apparatus and a method) provided by this invention collects normal sample data (each normal sample being the value of a robotic arm performing a complete variable-speed non-rotational motion within a standard cycle) using sensors, and first establishes a pre-trained dimensionality reduction model and a pre-trained clustering model. During the actual judgment stage, multiple sensing samples collected from the reducer under test are received and filtered. Then, the pre-trained dimensionality reduction model is used to reduce the dimensionality of these sensing samples to generate dimensionality reduction features corresponding to each sensing sample. Subsequently, the dimensionality reduction features of each sensing sample are input into the pre-trained clustering model to calculate the likelihood score for each sensing sample. Finally, based on a standard deviation threshold and the likelihood scores, it is determined whether the reducer under test is malfunctioning. The detection technology provided by this invention, through the above-described operation, quickly determines whether the reducer under test is malfunctioning, solving the problems of the prior art.

[0056] The above embodiments are merely illustrative of some implementations of the present invention and to explain the technical features of the present invention, and are not intended to limit the scope and range of protection of the present invention. Any changes or equivalent arrangements that can be easily made by those skilled in the art to which this invention pertains are within the scope of the present invention, and the scope of protection of the present invention is determined by the claims.

Claims

1. A detection system, comprising: The memory is used to store the pre-trained dimensionality reduction model and the pre-trained clustering model, wherein the pre-trained dimensionality reduction model is used to perform dimensionality reduction operations, and the pre-trained clustering model is generated based on multiple normal samples and used to calculate the likelihood score. as well as A processor, electrically connected to the memory, is configured to perform the following operations: Receive multiple sensor samples; Filtering is performed on the multiple sensed samples; The pre-trained dimensionality reduction model is used to perform the dimensionality reduction operation on the multiple sensed samples to generate dimensionality reduction features corresponding to each sensed sample. The dimensionality reduction features of each of the sensed samples are input into the pre-trained clustering model to calculate the likelihood score for each of the sensed samples. as well as Based on the standard deviation threshold and the likelihood score, determine whether the reducer is malfunctioning; The system further includes: At least one sensor is used to sense the speed reducer generating various sensing samples during operation of variable speed non-rotational motion, wherein the data length of each sensing sample corresponds to a standard cycle of the variable speed non-rotational motion.

2. The detection system according to claim 1, wherein the at least one sensor is an accelerometer, the accelerometer being mounted on a base at the radial position of the reducer of the robotic arm to collect multiple vibration signals.

3. The detection system according to claim 1, wherein the at least one sensor is a directional microphone, the directional microphone being installed around the reducer of the robotic arm to collect multiple sound signals.

4. The detection system according to claim 1, wherein the processor is further configured to perform the following operations: The filtering operation is performed on the plurality of normal samples, wherein each normal sample is generated by at least one sensor detecting a normal reducer operating in a variable speed non-rotational motion. The multiple normal samples are input into an autoencoder model constructed by a convolutional neural network to train the autoencoder model, wherein the autoencoder model includes an encoding operation part and a decoding operation part; and The encoding operation part of the trained autoencoder model is set as the pre-trained dimensionality reduction model.

5. The detection system according to claim 1, wherein the processor is further configured to perform the following operations: The filtering operation is performed on the plurality of normal samples, wherein each normal sample is generated by at least one sensor detecting a normal reducer operating in a variable speed non-rotational motion. The dimensionality reduction operation is performed on the multiple normal samples using the pre-trained dimensionality reduction model to generate the dimensionality reduction features corresponding to each of the normal samples; The dimensionality reduction features of each of the normal samples are input into a clustering model to train the clustering model, wherein the clustering model is used to perform clustering operations on multiple input data to generate the likelihood score corresponding to each of the input data. as well as The trained clustering model is used as the pre-trained clustering model.

6. A detection method for a detection system, the detection system comprising a memory and a processor, the memory storing a pre-trained dimensionality reduction model and a pre-trained clustering model, wherein the pre-trained dimensionality reduction model is used to perform dimensionality reduction operations, and the pre-trained clustering model is generated based on multiple normal samples to calculate a likelihood score, the detection method being executed by the processor and comprising the following steps: Receive multiple sensor samples; Filtering is performed on the multiple sensed samples; The pre-trained dimensionality reduction model is used to perform the dimensionality reduction operation on the multiple sensed samples to generate dimensionality reduction features corresponding to each sensed sample. The dimensionality reduction features of each of the sensed samples are input into the pre-trained clustering model to calculate the likelihood score for each of the sensed samples. as well as Based on the standard deviation threshold and the likelihood score, determine whether the reducer is malfunctioning; The detection system further includes at least one sensor for sensing the various sensing samples generated by the reducer during operation of the variable speed non-rotational motion, wherein the data length of each sensing sample corresponds to the standard cycle of the variable speed non-rotational motion.

7. The detection method according to claim 6, wherein the at least one sensor is an accelerometer, the accelerometer being mounted on a base at the radial position of the reducer of the robotic arm to collect multiple vibration signals.

8. The detection method according to claim 6, wherein the at least one sensor is a directional microphone, the directional microphone being installed around the reducer of the robotic arm to collect multiple sound signals.

9. The detection method according to claim 6, further comprising the following steps: The filtering operation is performed on the plurality of normal samples, wherein each normal sample is generated by at least one sensor detecting a normal reducer operating in a variable speed non-rotational motion. The multiple normal samples are input into an autoencoder model constructed by a convolutional neural network to train the autoencoder model, wherein the autoencoder model includes an encoding operation part and a decoding operation part; and The encoding operation part of the trained autoencoder model is set as the pre-trained dimensionality reduction model.

10. The detection method according to claim 6, further comprising the following steps: The filtering operation is performed on the plurality of normal samples, wherein each normal sample is generated by at least one sensor detecting a normal reducer operating in a variable speed non-rotational motion. The dimensionality reduction operation is performed on the multiple normal samples using the pre-trained dimensionality reduction model to generate the dimensionality reduction features corresponding to each of the normal samples; The dimensionality reduction features of each of the normal samples are input into a clustering model to train the clustering model, wherein the clustering model is used to perform clustering operations on multiple input data to generate the likelihood score corresponding to each of the input data. as well as The trained clustering model is used as the pre-trained clustering model.

Images