Fault data analysis method and device of smoke alarm and smoke alarm

Abstract

The invention provides a fault data analysis method and device for a smoke alarm and the smoke alarm, and the method comprises the steps: obtaining original data collected by the smoke alarm, and extracting the multi-dimensional features of the original data; inputting the multi-dimensional features into a pre-trained deep neural network, and processing to obtain multi-dimensional reduced features; wherein the deep neural network performs coding compression on the multi-dimensional features through an auto-encoder architecture, and then performs decoding reduction to obtain multi-dimensional reduction features; calculating the difference between the multi-dimensional reduction feature and a standard feature to obtain a difference result; and obtaining fault data in the original data collected by the smoke alarm according to the difference result. According to the invention, deep analysis is carried out on the collected data by integrating multiple factors, fault data are accurately identified, and the reliability and troubleshooting efficiency of the smoke alarm are improved, so that stable operation of a fire safety monitoring system is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a fault data analysis method and device for a smoke alarm, and a smoke alarm. Background Art

[0002] During the long-term operation of smoke alarms, due to the influence of various factors, failures may occur, resulting in false alarms or missed alarms, which seriously affect their reliability and effectiveness.

[0003] On the one hand, the environment in which smoke alarms are located is complex and changeable. Environmental factors such as temperature, humidity, dust, and electromagnetic interference may interfere with the performance of its sensors, thereby causing deviations in the collected data. On the other hand, as the use time increases, problems such as aging and damage of the electronic components inside the smoke alarms will gradually emerge, which will also be reflected in the abnormalities of the collected data.

[0004] Traditional smoke alarm fault detection methods are often simple and limited, usually based on a single threshold judgment. This simple threshold judgment method is difficult to distinguish between real fire danger signals and erroneous data caused by the alarm's own faults, which can easily lead to misjudgment. In actual applications, it may cause unnecessary panic or fail to provide timely warning when a fire occurs.

[0005] That is, the current smoke alarms are unable to identify fault data caused by their own faults. Summary of the invention

[0006] The main purpose of the present invention is to provide a method and device for analyzing fault data of a smoke alarm, and a smoke alarm, aiming to overcome the defect that the current smoke alarm cannot identify fault data.

[0007] To achieve the above object, the present invention provides a method for analyzing fault data of a smoke alarm, comprising the following steps:

[0008] Obtaining raw data collected by the smoke alarm and extracting multi-dimensional features of the raw data;

[0009] Input the multi-dimensional features into a pre-trained deep neural network, and process them to obtain multi-dimensional restored features; wherein the deep neural network encodes and compresses the multi-dimensional features through an autoencoder architecture, and then decodes and restores them to obtain multi-dimensional restored features;

[0010] Calculating the difference between the multi-dimensional restored features and the standard features to obtain a difference result;

[0011] According to the difference result, the fault data in the original data collected by the smoke alarm is obtained.

[0012] Furthermore, the raw data includes smoke concentration data, data acquisition timestamp data, temperature and humidity data, and sensor status data.

[0013] Furthermore, the multi-dimensional features are input into a pre-trained deep neural network to obtain multi-dimensional restored features, including:

[0014] The multi-dimensional features are input into the encoding layer of the autoencoder architecture. The encoding layer adopts a multi-layer fully connected neural network structure. The number of neurons in each layer decreases successively. The features are transformed by a nonlinear activation function, and the high-dimensional multi-dimensional features are mapped to a low-dimensional encoding space to obtain encoded data;

[0015] The encoded data is input into the decoding layer of the autoencoder architecture. The decoding layer adopts a multi-layer fully connected neural network structure. The number of neurons in each layer increases successively and is symmetrical with the structure of the encoding layer. The encoded data is transformed by a nonlinear activation function opposite to that of the encoding layer, and the low-dimensional encoded data is restored to high-dimensional multi-dimensional restored features.

[0016] Furthermore, the difference between the multi-dimensional restored features and the standard features is calculated to obtain a difference result, including:

[0017] Performing principal component analysis on the multi-dimensional restored features and the standard features respectively, extracting their respective principal component feature vectors, and obtaining a principal component matrix of the multi-dimensional restored features and a principal component matrix of the standard features;

[0018] The cosine similarity between the multi-dimensional restored feature principal component matrix and the standard feature principal component matrix is ​​calculated to obtain a similarity value sequence of each dimension as the difference result.

[0019] Furthermore, the extracting multi-dimensional features of the original data includes:

[0020] Based on wavelet packet decomposition, the original data is decomposed into multi-layer wavelet packet subspaces, and the energy entropy of the signal in each subspace is calculated as the energy entropy feature, and the correlation coefficient between the signal in each subspace and the original data is calculated as the correlation coefficient feature;

[0021] Based on a local linear embedding algorithm, nonlinear dimensionality reduction is performed on the original data, and in the low-dimensional space after dimensionality reduction, distance features between data points are calculated, and combined with the Euclidean distance features of the original data in the high-dimensional space to obtain a distance feature combination;

[0022] Using the original data as an observation sequence, based on a hidden Markov model, the hidden state transition probability matrix characteristics and the observation probability distribution characteristics under each hidden state are obtained;

[0023] The energy entropy feature, correlation coefficient feature, distance feature combination, hidden state transition probability matrix feature and observation probability distribution feature are used as multi-dimensional features of the original data.

[0024] Furthermore, after obtaining the fault data in the original data collected by the smoke alarm according to the difference result, the method further comprises:

[0025] Based on the fault data, obtaining the faulty component of the smoke alarm;

[0026] Obtaining characteristic information of a supplier of the faulty component;

[0027] Obtaining product information of the faulty component and manufacturer information of the smoke alarm;

[0028] Generate an after-sales authorization code based on the feature information, product information and manufacturer information;

[0029] After-sales maintenance information is sent to the supplier; the after-sales maintenance information carries the after-sales authority code, which is used by the supplier to authenticate the authority of the after-sales maintenance information.

[0030] Furthermore, based on the feature information, product information and manufacturer information, an after-sales authorization code is generated, including:

[0031] Convert the characteristic information, product information and manufacturer information into preset matrix forms respectively; wherein the characteristic information is constructed into a multi-dimensional sparse matrix, the product information is constructed into a triangular matrix, and the manufacturer information is constructed into a symmetric matrix;

[0032] Performing matrix multiplication operation on the transformed matrix to obtain a composite matrix, and calculating the eigenvalues ​​and eigenvectors of the composite matrix;

[0033] With the eigenvalue as the horizontal coordinate and the modulus of the eigenvector as the vertical coordinate, a plurality of plane coordinate points are constructed, and each plane coordinate point is connected to form a continuous and smooth curve;

[0034] Performing Fourier series expansion on the curve, arranging the coefficients of the preset items in order into a digital sequence;

[0035] The after-sales authorization code is generated based on the digital sequence.

[0036] The present invention also provides a fault data analysis device for a smoke alarm, comprising:

[0037] An acquisition unit, used to acquire raw data collected by the smoke alarm and extract multi-dimensional features of the raw data;

[0038] A processing unit is used to input the multi-dimensional features into a pre-trained deep neural network to obtain multi-dimensional restored features; wherein the deep neural network encodes and compresses the multi-dimensional features through an autoencoder architecture, and then decodes and restores the multi-dimensional restored features;

[0039] A calculation unit, used to calculate the difference between the multi-dimensional restored features and the standard features to obtain a difference result;

[0040] The analysis unit is used to obtain the fault data in the original data collected by the smoke alarm according to the difference result.

[0041] The present invention further provides a smoke alarm, comprising a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps of any one of the above methods when executing the computer program.

[0042] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0043] The fault data analysis method and device of the smoke alarm provided by the present invention, as well as the smoke alarm, include: obtaining the original data collected by the smoke alarm, extracting the multi-dimensional features of the original data; inputting the multi-dimensional features into a pre-trained deep neural network, processing to obtain multi-dimensional restored features; wherein the deep neural network encodes and compresses the multi-dimensional features through an autoencoder architecture, and then decodes and restores to obtain multi-dimensional restored features; calculating the difference between the multi-dimensional restored features and the standard features to obtain a difference result; according to the difference result, obtaining the fault data in the original data collected by the smoke alarm. In the present invention, the collected data is deeply analyzed by comprehensively considering various factors, the fault data is accurately identified, and the reliability and troubleshooting efficiency of the smoke alarm are improved to ensure the stable operation of the fire safety monitoring system. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 1 is a schematic diagram of the steps of a method for analyzing fault data of a smoke alarm according to an embodiment of the present invention;

[0045] Figure 2 This is a structural block diagram of a fault data analysis device for a smoke alarm according to an embodiment of the present invention;

[0046] Figure 3 The figure is a schematic block diagram of the structure of a smoke alarm according to an embodiment of the present invention.

[0047] The implementation, functional features and advantages of the present invention will be further described in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0049] Reference Figure 1 In one embodiment of the present invention, a method for analyzing fault data of a smoke alarm is provided, comprising the following steps:

[0050] Step S1, obtaining raw data collected by the smoke alarm and extracting multi-dimensional features of the raw data;

[0051] Step S2, inputting the multi-dimensional features into a pre-trained deep neural network, processing to obtain multi-dimensional restored features; wherein the deep neural network encodes and compresses the multi-dimensional features through an autoencoder architecture, and then decodes and restores to obtain multi-dimensional restored features;

[0052] Step S3, calculating the difference between the multi-dimensional restored features and the standard features to obtain a difference result;

[0053] Step S4, obtaining fault data in the original data collected by the smoke alarm according to the difference result.

[0054] In this embodiment, as described in step S1 above, the smoke alarm will continuously collect relevant raw data during operation, and these raw data are the basis for subsequent analysis. For example, the raw data may include a numerical sequence of smoke concentration changes over time, ambient temperature, humidity data, and the sensor's own working state parameters (such as voltage, current, etc.).

[0055] The purpose of extracting multi-dimensional features is to more comprehensively characterize the characteristics of the original data. In one embodiment, for smoke concentration data, the features that can be extracted include the concentration change rate (such as the increase or decrease in smoke concentration per unit time), the fluctuation frequency of concentration data (reflecting the stability of smoke concentration or the frequency of change), the concentration mean and variance in different time periods and other statistical features. For environmental data, extreme values, means, and change trends of temperature and humidity can also be used as features. The extraction of the above multi-dimensional features helps to examine the data collected by the smoke alarm from multiple angles in order to more accurately discover potential fault information.

[0056] As described in step S2 above, the pre-trained deep neural network autoencoder has a unique architecture and function. The autoencoder consists of an encoding layer and a decoding layer. During the encoding process, high-dimensional multi-dimensional feature data is input into the encoding layer, and the encoding layer gradually maps the data to a low-dimensional encoding space through the connection of multiple layers of neurons and a specific activation function to achieve data compression. The above process can extract key information or potential patterns in the data and remove some redundant information.

[0057] The low-dimensional encoded data obtained after encoding is then input into the decoding layer. The structure of the decoding layer corresponds to that of the encoding layer but in the opposite order. Through a series of conversion operations, the low-dimensional encoded data is gradually restored back to the high-dimensional data space to obtain multi-dimensional restored features. Through this process of encoding, compression, decoding and restoration, the neural network can learn the intrinsic structure and feature representation of the data, and ideally, the restored multi-dimensional restored features should have a high similarity with the original multi-dimensional features. If there is a large difference between the two, it indicates that the original data contains abnormal or fault-related information.

[0058] As described in step S3 above, the standard features are determined based on the multi-dimensional feature statistical model of the data collected by the smoke alarm under normal working conditions. The above standard features can be obtained by analyzing and summarizing a large amount of normal operating data, such as the reasonable range of the smoke concentration change rate under normal circumstances, the normal correlation pattern between environmental data and smoke concentration data, and other standard values ​​or standard ranges of multi-dimensional features. When calculating the difference between the multi-dimensional restored features and the standard features, the difference measure can be calculated for each dimension separately, such as the absolute value of the difference, the relative proportion difference, etc. It is also possible to use a specific algorithm or model to calculate the overall degree of difference based on the differences in multiple dimensions, such as by weighted summation, assigning different weights to dimensions of different importance, and obtaining a difference result that comprehensively reflects the degree to which the multi-dimensional restored features deviate from the standard features. This difference result will serve as the key basis for determining whether there is a fault in the original data.

[0059] As described in step S4 above, when the difference result exceeds a preset threshold or meets a specific fault judgment rule, it can be determined that the corresponding original data contains fault information. For example, if the difference in the smoke concentration change rate exceeds a certain threshold, it indicates that the smoke sensor is abnormal, such as sensitivity changes, being blocked or damaged; if the correlation characteristics between the environmental data and the smoke concentration data are significantly different, it indicates that environmental factors have caused abnormal interference to the smoke alarm or that the internal data processing module has a fault. In this way, fault data can be accurately identified from the original data, thereby providing an important basis for subsequent fault location, alarm, and maintenance processing, thereby improving the reliability and stability of the smoke alarm system.

[0060] In this embodiment, the collected data is deeply analyzed by comprehensively considering various factors, fault data is accurately identified, the reliability of smoke alarms and the efficiency of troubleshooting are improved, so as to ensure the stable operation of the fire safety monitoring system.

[0061] In one embodiment, the raw data includes smoke concentration data, data collection timestamp data, temperature and humidity data, and sensor status data.

[0062] In one embodiment, the multi-dimensional features are input into a pre-trained deep neural network, and processed to obtain multi-dimensional restored features, including:

[0063] The multi-dimensional features are input into the encoding layer of the autoencoder architecture. The encoding layer adopts a multi-layer fully connected neural network structure. The number of neurons in each layer decreases successively. The features are transformed by a nonlinear activation function, and the high-dimensional multi-dimensional features are mapped to a low-dimensional encoding space to obtain encoded data;

[0064] The encoded data is input into the decoding layer of the autoencoder architecture. The decoding layer adopts a multi-layer fully connected neural network structure. The number of neurons in each layer increases successively and is symmetrical with the structure of the encoding layer. The encoded data is transformed by a nonlinear activation function opposite to that of the encoding layer, and the low-dimensional encoded data is restored to high-dimensional multi-dimensional restored features.

[0065] In this embodiment, the neurons in each layer of the fully connected neural network are interconnected with all the neurons in the next layer. This structure can fully capture the complex relationship between different dimensions of the input data. A multi-layer structure is used in the encoding layer. As the number of layers increases, deep feature extraction and information integration can be gradually performed on multi-dimensional features. For example, for multi-dimensional features such as smoke concentration changes and environmental parameters extracted from the data collected by the smoke alarm, the first layer of neurons can receive and preliminarily process these features, and then pass them to the next layer. Each layer will perform more abstract feature representation learning based on the output of the previous layer.

[0066] The multi-dimensional features initially input are in a higher dimension and contain relatively rich but also redundant information. As the number of neurons gradually decreases, the data will be mapped to a lower-dimensional space, the so-called encoding space. For example, the number of neurons corresponding to the multi-dimensional features initially input is hundreds. After multiple layers of reduction, it is finally compressed to low-dimensional encoding data represented by only a dozen neurons. This process is to refine the original data, retain the key and representative information, and remove the details that are not important to the overall feature expression, which will help to more efficiently mine the potential structure and feature patterns of the data in the future.

[0067] Without activation functions, multi-layer fully connected neural networks are equivalent to the superposition of multiple linear transformations, and ultimately can only represent linear relationships, but cannot handle complex nonlinear data relationships in reality. Commonly used nonlinear activation functions, such as ReLU (Rectified Linear Unit), Sigmoid or Tanh, can introduce nonlinear factors into neural networks. Taking ReLU as an example, it will change input values ​​less than 0 to 0, and input values ​​greater than 0 remain unchanged, which makes the neuron output nonlinear. In the encoding layer, when multi-dimensional feature data passes through each layer of neurons, through the action of these nonlinear activation functions, the data can be nonlinearly transformed, and then more complex and more realistic feature representations can be learned, so that the neural network has the ability to model the nonlinear relationships contained in the multi-dimensional features, better complete the task of mapping high-dimensional features to low-dimensional coding space, and obtain effective coding data.

[0068] The decoding layer also uses a multi-layer fully connected neural network structure, which is to process the encoded data in a step-by-step manner. The symmetrical structural design with the encoding layer means that the number of neurons in each layer of the decoding layer changes in the opposite trend to the encoding layer, with the encoding layer gradually decreasing and the decoding layer increasing in sequence. The above symmetrical structure helps to ensure that the data can be reasonably restored from the low-dimensional encoding state to the high-dimensional state at the structural level, which is in line with the overall working logic and data processing flow of the autoencoder.

[0069] The restoration process from low-dimensional to high-dimensional requires gradually increasing the amount of information and dimensional representation of the data. The increase in the number of neurons in each layer is to continuously supplement and enrich the data features, so that the restored features can be as close as possible to the multi-dimensional features of the original input. For example, the decoding layer starts with the initially received low-dimensional encoded data. As the number of layers advances and the number of neurons increases, it continuously adds more details and dimensional information to the data, and finally achieves restoration to the original high-dimensional feature space. In the decoding process, a nonlinear activation function opposite to that of the encoding layer is used. This is to be able to reversely restore the features that conform to the distribution characteristics of the original data when restoring the data. Since the encoding layer transforms and compresses the multi-dimensional features through a specific nonlinear activation function, the corresponding and reverse nonlinear activation function is used during decoding, which helps to more accurately restore the encoded data to a state close to the original input features.

[0070] By using the collaborative working method of the encoding layer and the decoding layer mentioned above, and utilizing the deep neural network autoencoder architecture, the multi-dimensional features can be encoded, compressed, and then decoded and restored to obtain multi-dimensional restored features, providing a key data foundation for further fault data analysis.

[0071] In one embodiment, calculating the difference between the multi-dimensional restored features and the standard features to obtain the difference result includes:

[0072] Performing principal component analysis on the multi-dimensional restored features and the standard features respectively, extracting their respective principal component feature vectors, and obtaining a principal component matrix of the multi-dimensional restored features and a principal component matrix of the standard features;

[0073] The cosine similarity between the multi-dimensional restored feature principal component matrix and the standard feature principal component matrix is ​​calculated to obtain a similarity value sequence of each dimension as the difference result.

[0074] In this embodiment, the principal component analysis is a commonly used data dimension reduction and feature extraction method. It converts multiple variables in the original data into a set of new unrelated variables through linear transformation, which are called principal components. The principal components are arranged from large to small according to the variance. The first principal component can explain the variance of the original data to the greatest extent, the second principal component is second, and so on. When processing multi-dimensional data, the first few principal components can often contain most of the information of the original data.

[0075] For multi-dimensional reduction features and standard features, their covariance matrix must be calculated first. The covariance matrix reflects the correlation between the dimensions. Then, the principal components are determined by solving the eigenvalues ​​and eigenvectors of the covariance matrix. The eigenvalue represents the variance explained by the corresponding eigenvector, and the eigenvector is the direction of the principal component. Sort the corresponding eigenvectors from large to small according to the eigenvalue, and select the first few eigenvectors with larger eigenvalues ​​as the principal component eigenvectors. These principal component eigenvectors form the principal component matrix.

[0076] Cosine similarity is a measure of the cosine value of the angle between two vectors. In vector space, it can well reflect the degree of directional consistency of two vectors. The value range of cosine similarity is between -1 and 1. The closer the value is to 1, the more similar the two vectors are and the more consistent their directions are; the closer the value is to -1, the more opposite the directions of the two vectors are; the closer the value is to 0, the more there is almost no linear correlation between the two vectors.

[0077] Each pair of corresponding principal component vectors in the multi-dimensional restored feature principal component matrix and the standard feature principal component matrix (by column or row, depending on the organization of the matrix) is regarded as two vectors, and the cosine similarity between them is calculated. For example, if the multi-dimensional restored feature principal component matrix and the standard feature principal component matrix are both 3×10 matrices, then for the two vectors in each column (representing a principal component dimension), the cosine similarity is calculated, so that a sequence of 10 similarity values ​​(assuming that the original data has 10 dimensions) can be obtained. The above similarity value sequence is used as the difference result between the multi-dimensional restored feature and the standard feature. If the cosine similarity of a certain dimension is low, it means that there is a large difference between the multi-dimensional restored feature and the standard feature in the feature direction represented by this principal component, so it can be inferred that there may be a fault or abnormality in the part related to this dimension of the original data. The above method of combining principal component analysis with cosine similarity can effectively quantify the difference between the multi-dimensional restored feature and the standard feature on the basis of considering the correlation between the dimensions of the data, and provide an important basis for subsequent fault judgment.

[0078] In one embodiment, extracting the multi-dimensional features of the original data includes:

[0079] Based on wavelet packet decomposition, the original data is decomposed into multi-layer wavelet packet subspaces, and the energy entropy of the signal in each subspace is calculated as the energy entropy feature, and the correlation coefficient between the signal in each subspace and the original data is calculated as the correlation coefficient feature;

[0080] Based on a local linear embedding algorithm, nonlinear dimensionality reduction is performed on the original data, and in the low-dimensional space after dimensionality reduction, distance features between data points are calculated, and combined with the Euclidean distance features of the original data in the high-dimensional space to obtain a distance feature combination;

[0081] Using the original data as an observation sequence, based on a hidden Markov model, the hidden state transition probability matrix characteristics and the observation probability distribution characteristics under each hidden state are obtained;

[0082] The energy entropy feature, correlation coefficient feature, distance feature combination, hidden state transition probability matrix feature and observation probability distribution feature are used as multi-dimensional features of the original data.

[0083] In this embodiment, wavelet packet decomposition is a multi-resolution analysis of the signal, which is a further extension of the wavelet transform. It is based on wavelet decomposition, and the high-frequency part that is not subdivided by wavelet decomposition is further decomposed, so that the characteristics of the signal in different frequency bands can be analyzed more finely. For the original data collected by the smoke alarm (such as data on smoke concentration changes over time, etc.), by selecting appropriate wavelet basis functions (such as Haar wavelets, Daubechies wavelets, etc.), according to a certain number of decomposition layers, the original data is gradually decomposed into multi-layer wavelet packet subspaces. Each layer of decomposition will further divide a certain frequency band of the previous layer into narrower sub-bands, so that subspaces of different frequency ranges can be obtained, which can more carefully present the performance of the original data in each frequency component.

[0084] In each wavelet packet subspace, the energy entropy of the signal is calculated. The energy entropy reflects the complexity and uncertainty of the energy distribution of the signal in the subspace. In the specific calculation, the energy of the signal in the subspace is calculated first, usually by summing the square of the signal amplitude to determine its energy value, and then based on the calculation principle of information entropy, the energy value is normalized, and then the energy entropy is calculated according to the calculation formula of information entropy (such as Shannon entropy formula). The energy entropy characteristics of different subspaces can reflect the difference in disorder or regularity of the energy distribution of the original data in different frequency bands. The above differences are very helpful for judging whether the original data has abnormal frequency characteristics changes due to faults.

[0085] Calculate the correlation coefficient between the signal in each subspace and the original data. Here, the statistical calculation method of the correlation coefficient is generally used, such as the commonly used Pearson correlation coefficient. It measures the degree of linear correlation between two variables (i.e., the signal in the subspace and the original data as a whole), and the value range is between -1 and 1. By calculating this correlation coefficient, we can understand the degree of correlation between the signal in each subspace and the original complete data, reflecting the relative importance and changes of different frequency components in the overall structure of the original data. If the correlation between the signal in a subspace and the original data shows obvious abnormal changes, it indicates that the original data is affected by the fault factor in the corresponding frequency part.

[0086] The above-mentioned local linear embedding algorithm is a nonlinear dimensionality reduction method, which is based on the idea of ​​manifold learning and believes that high-dimensional data is often distributed on a low-dimensional manifold. For the original data of the smoke alarm (which may contain multiple dimensions of environmental and sensor-related data, etc.), find the local linear relationship between each data point and its neighboring points in the high-dimensional data space, and then map the high-dimensional data to the low-dimensional space by maintaining these local linear relationships, so that the local geometric structure characteristics of the original data can still be better preserved in the low-dimensional space. The advantage of this is that while reducing the data dimension, it avoids losing nonlinear information in the data like some linear dimensionality reduction methods, which is more in line with the complex distribution of actual data.

[0087] After the original data is nonlinearly reduced to a low-dimensional space, the distance features between the data points in this low-dimensional space are calculated. A variety of distance metrics can be used, such as Euclidean distance (straight-line distance between two points), Manhattan distance, etc. At the same time, the Euclidean distance features of the original data in the high-dimensional space must also be considered. The Euclidean distance in the original high-dimensional space reflects the spatial distance of the data under the combined effect of various dimensions when the dimension is not reduced. It is combined with the distance features of the low-dimensional space to form a distance feature combination. For example, the Euclidean distance of two data points in the high-dimensional space and their corresponding Euclidean distances in the low-dimensional space can be combined into a new feature vector. By comparing the changes and differences in the distances in the high-dimensional and low-dimensional spaces, the characteristic changes of the original data in the dimensional transformation process can be more comprehensively grasped, and the spatial structure abnormalities that may be caused by faults can be found, because under normal circumstances, there should be a certain reasonable correlation between the distances in the high-dimensional and low-dimensional spaces, and faults may break this correlation.

[0088] Hidden Markov Model (HMM) is a statistical model used to describe a Markov process with hidden unknown parameters. The raw data of the smoke alarm can be regarded as an observation sequence, and there are some hidden states behind it (for example, the hidden state can represent the different working states of the smoke alarm, such as normal operation state, slightly disturbed state, impending failure state, etc.). There is a certain transition probability relationship between these hidden states, and the probability of generating observation data (i.e., raw data) under each hidden state is also different. By training a large amount of raw data under known normal and possible failure conditions, the hidden Markov model can learn the transition rules between these hidden states and the probability distribution of the observed data under each hidden state.

[0089] The hidden state transition probability matrix represents the probability of transitioning from one hidden state to another. After the model training is completed, by analyzing this matrix, we can understand the probability of transitioning between different hidden working states of the smoke alarm. For example, if the probability of transitioning from the normal state to the fault state suddenly increases at a certain moment, it indicates that a fault is about to occur. The values ​​of each element in this matrix are the characteristics of the hidden state transition probability matrix, which reflects the dynamic trend and possibility characteristics of the internal state change.

[0090] Each hidden state has a corresponding observation probability distribution, that is, the probability of different observation data (such as different smoke concentration values, environmental parameter values, etc.) appearing under a specific hidden working state. By obtaining the observation probability distribution characteristics under each hidden state, we can know the probability distribution law of the original data under different working conditions. For example, the smoke concentration value may obey a normal distribution under normal conditions, while it may present other different distribution forms under fault conditions. By comparing the actual observation data with the observation probability distribution characteristics under these hidden states, we can determine which hidden state the current original data is more consistent with, and then assist in determining whether there is a fault and the possible type of fault.

[0091] The energy entropy features, correlation coefficient features, distance feature combinations, hidden state transition probability matrix features, and observation probability distribution features extracted by the above different technical means, each feature describes a certain aspect of the original data from a unique perspective. Integrating them together forms a rich and comprehensive multi-dimensional feature set of the original data. The above multi-dimensional features can more comprehensively reflect the overall situation of the data collected by the smoke alarm, whether it is the frequency characteristics of the data, the spatial structure relationship, or the hidden working status association behind it, etc., which provides a more sufficient and in-depth feature basis for the subsequent input of deep neural networks for further fault data analysis, which helps to more accurately judge whether the original data is abnormal due to a fault, and improve the accuracy and reliability of fault detection.

[0092] In one embodiment, after obtaining the fault data in the original data collected by the smoke alarm according to the difference result, the method includes:

[0093] Based on the fault data, obtaining the faulty component of the smoke alarm;

[0094] Obtaining characteristic information of a supplier of the faulty component;

[0095] Obtaining product information of the faulty component and manufacturer information of the smoke alarm;

[0096] Generate an after-sales authorization code based on the feature information, product information and manufacturer information;

[0097] After-sales maintenance information is sent to the supplier; the after-sales maintenance information carries the after-sales authority code, which is used by the supplier to authenticate the authority of the after-sales maintenance information.

[0098] In this embodiment, the smoke alarm is a complex device composed of multiple different components, such as a smoke sensor, a signal processing module, a power module, a communication module, etc. Failures of different components will show different characteristics in the collected raw data. For example, if the smoke concentration data is always abnormal and does not conform to the normal environmental correlation, it may be that the smoke sensor itself has a fault, such as sensor aging, damage, or being blocked by foreign objects; if there is an abnormal interruption or error in data transmission, it may be that there is a problem with the communication module. By analyzing and matching the specific abnormal patterns presented by the fault data, combined with the internal structure of the smoke alarm and the functional characteristics of each component, using pre-established fault models or empirical knowledge, etc., it is possible to accurately determine which specific component has failed, and provide a clear object for subsequent targeted repairs, replacements and other after-sales processing.

[0099] The supplier characteristic information of the faulty component covers many aspects, such as the supplier's name, address, contact information and other basic business information. Obtaining this information helps to have a more comprehensive understanding of the supplier's background and facilitates subsequent communication and coordination with the supplier on after-sales matters.

[0100] The product information of the faulty component includes the component model, specification, batch number, production date, etc. The model and specification determine the specific performance parameters and applicable scope of the component. Components of different models and specifications will have different functional performance and types of failures. The batch number and production date are critical for tracing the component production process and checking whether there are quality problems caused by the production process. For example, if there are multiple similar failures in the same batch of components, it is likely that there are systemic problems in the raw materials, processes, and other links in the production process.

[0101] The manufacturer information of the smoke alarm is equally important, including the manufacturer's name, address, contact information, manufacturer's qualifications, market reputation, etc. Knowing the manufacturer's information makes it easier to communicate and collaborate with the manufacturer to solve the problem together when complex faults occur or when coordination among multiple parties is involved.

[0102] The generation of after-sales permission codes is to ensure the security and standardization of after-sales maintenance information transmission and processing. By comprehensively utilizing the characteristic information of suppliers, the product information of faulty components, and the manufacturer information of smoke alarms, a unique permission code is generated using preset algorithms or coding rules. For example, the information from these different sources can be formatted and sorted first, key data items can be extracted, and then the sorted information can be encrypted using an encryption algorithm (such as a combination of hash encryption, symmetric encryption, or asymmetric encryption), and the encrypted result can be used as an after-sales permission code. Such permission codes are unique and confidential. Only authorized suppliers who have mastered the corresponding decryption or verification rules can authenticate their permissions, ensure that after-sales maintenance information is accurately delivered to the correct supplier, and prevent information from being tampered with or misused, thereby ensuring the reliability and security of the entire after-sales process.

[0103] After determining the faulty component, collecting relevant information and generating an after-sales authorization code, it is necessary to send after-sales maintenance information to the corresponding supplier. The maintenance information describes in detail the fault condition of the smoke alarm, the faulty component related information and the expected after-sales processing method (such as repair, replacement, etc.), and carries the generated after-sales authorization code. After receiving the after-sales maintenance information, the supplier will first use the pre-set authentication mechanism to verify the authorization code, such as by matching its own saved encryption key or verification algorithm rules, to determine whether the information comes from a legally authorized channel and whether the information content is complete and has not been tampered with. Only after the authorization authentication is passed, the supplier will initiate the corresponding after-sales process based on the specific content of the after-sales maintenance information, such as arranging technicians to perform repairs, preparing replacement components, etc., so as to achieve efficient, safe and standardized after-sales maintenance docking.

[0104] Through the above steps, after detecting a smoke alarm failure, subsequent after-sales processing work can be carried out in an orderly manner to ensure that the equipment can resume normal operation as soon as possible, while maintaining the rights and interests of all parties in the after-sales link and the safety and standardization of the entire process.

[0105] In one embodiment, generating an after-sales authorization code based on the feature information, product information, and manufacturer information includes:

[0106] Convert the characteristic information, product information and manufacturer information into preset matrix forms respectively; wherein the characteristic information is constructed into a multi-dimensional sparse matrix, the product information is constructed into a triangular matrix, and the manufacturer information is constructed into a symmetric matrix;

[0107] Performing matrix multiplication operation on the transformed matrix to obtain a composite matrix, and calculating the eigenvalues ​​and eigenvectors of the composite matrix;

[0108] With the eigenvalue as the horizontal coordinate and the modulus of the eigenvector as the vertical coordinate, a plurality of plane coordinate points are constructed, and each plane coordinate point is connected to form a continuous and smooth curve;

[0109] Performing Fourier series expansion on the curve, arranging the coefficients of the preset items in order into a digital sequence;

[0110] The after-sales authorization code is generated based on the digital sequence.

[0111] In this embodiment, the characteristic information usually includes descriptions of many aspects of the supplier, such as attributes of different dimensions, and these attributes may not all exist densely, and some attributes may be vacant or have very few values ​​for specific situations. Constructing it into a multidimensional sparse matrix means using different attributes as the dimensions of the matrix, such as the supplier's credit rating, the number of major quality accidents in the past, the number of cooperative customers, etc. as element indicators of different rows or columns. Since there are a large number of zero values ​​or vacant values ​​in these elements (corresponding to certain attributes, the supplier has no relevant records, etc.), the formed matrix is ​​sparse. The above representation method is convenient for clearly reflecting the relationship between the attributes and the overall characteristic distribution from a mathematical structure. At the same time, the sparse matrix can use its special storage and calculation methods in subsequent mathematical operations to improve efficiency.

[0112] Product information includes component models, specifications, batch numbers, production dates, etc. The construction of this information as a triangular matrix is ​​based on its inherent data structure and logical relationship. For example, the different elements of product information can be arranged in order, and the more important and basic information (such as model) can be placed on the main diagonal of the matrix. Then, according to the strength of the association or logical dependency between the elements, the other elements can be placed in the upper or lower triangular area of ​​the matrix to form a triangular matrix. The triangular matrix has some special mathematical properties, such as the relatively simple evaluation of the determinant, which facilitates subsequent operations such as calculations with other matrices and feature extraction, and this construction method can well reflect the hierarchy and association characteristics between the various parts of the product information.

[0113] Manufacturer information involves the manufacturer's name, address, contact information, qualification level and other aspects. Since there is often a corresponding relationship between these information, for example, there is a certain correlation between the reputation of the manufacturer and the qualification level, and from a mathematical point of view, the relationship between many attributes is symmetrical (such as the degree of correlation between attribute A and attribute B is the same as the degree of correlation between attribute B and attribute A), it is constructed as a symmetric matrix. Different manufacturer information attributes are used as rows and columns of the matrix, and the elements in the corresponding positions represent the correlation strength or quantitative relationship between the two attributes. The symmetric matrix can concisely and accurately reflect the internal connection between the various parts of the manufacturer information, and the properties of the symmetric matrix can be used for efficient processing in subsequent matrix operations.

[0114] The above-constructed multidimensional sparse matrix, triangular matrix and symmetric matrix are subjected to matrix multiplication. Matrix multiplication has clear mathematical rules. It will comprehensively consider the corresponding relationship between the elements of each matrix for calculation. In this process, it actually combines three different but interrelated sources of information through mathematical operations. The composite matrix contains the comprehensive mathematical characteristics of the fusion of feature information, product information and manufacturer information. It no longer simply reflects a certain type of information, but interweaves the relationship and characteristics of the three.

[0115] The calculation of the eigenvalues ​​and eigenvectors of a composite matrix is ​​based on relevant theories and methods in linear algebra. Eigenvalues ​​and eigenvectors reflect some essential characteristics of the linear transformation represented by the matrix. For this composite matrix that integrates various information, the eigenvalues ​​are obtained by solving its characteristic equation (generally in the form of a determinant), and the non-zero vector corresponding to each eigenvalue is the eigenvector. The size of the eigenvalue indicates the degree of "stretching" or "compression" in the direction of the corresponding eigenvector, while the eigenvector represents the invariance characteristics of the composite matrix in a specific direction. By obtaining these eigenvalues ​​and eigenvectors, the key mathematical features after the fusion of information can be extracted from a deeper level. The above features are more representative and abstract than the original information, which facilitates the subsequent generation of unique and secure after-sales authorization codes.

[0116] The calculated eigenvalue is used as the horizontal coordinate, and the modulus of the eigenvector (i.e., the length of the eigenvector, calculated by taking the square root of the sum of the squares of each element of the vector) is used as the vertical coordinate. In this way, each set of eigenvalues ​​and the corresponding modulus of the eigenvector can determine a plane coordinate point. Since a composite matrix has multiple eigenvalues ​​and corresponding eigenvectors, multiple such plane coordinate points can be constructed. These coordinate points are distributed on the plane, which intuitively reflects the relationship between the eigenvalues ​​and eigenvectors of the composite matrix from a geometric perspective. Moreover, each point contains the key features extracted after the previous fusion of information. These abstract mathematical features are visualized in the form of coordinate points, which is convenient for subsequent operations such as curve construction.

[0117] Connect the plane coordinate points in order (for example, in order from small to large eigenvalues) to form a continuous and smooth curve. The above curve is a geometric integration and continuous expression of the information contained in the previous coordinate points. It is no longer a discrete point, but has overall geometric features such as shape and trend. The shape of the curve will vary greatly due to the original feature information, product information and manufacturer information. It actually becomes a unique graphical representation generated based on this information, and because of its continuity and smoothness, it is easier to further extract the internal laws and characteristics through mathematical analysis methods (such as Fourier series expansion, etc.).

[0118] Fourier series expansion is a mathematical method that decomposes a periodic function (the continuous smooth curve here can be regarded as a special form of periodic function representation, or processed by periodic extension within a certain interval) into a series of linear combinations of sine and cosine functions. Through this expansion, the complex shape and change law contained in the curve can be represented by the coefficients of a series of trigonometric functions of different frequencies. Specifically, for a given curve, according to the formula of Fourier series expansion, the coefficients of the sine and cosine functions of the corresponding frequency components can be calculated. These coefficients reflect the energy distribution or characteristic contribution of the curve at different frequencies, and mathematically perform a deep analysis and quantitative representation of the curve.

[0119] Select the preset coefficients of the Fourier series expansion and arrange them into a digital sequence in a certain order (for example, in the order of coefficients corresponding to low to high frequencies). The above digital sequence integrates the main frequency characteristic information of the curve. Since the curve itself is connected based on the coordinate points constructed by integrating multiple aspects of information, this digital sequence is actually a concise and orderly digital representation of the original characteristic information, product information and manufacturer information after multiple layers of mathematical transformation and refinement. It has uniqueness and certainty (given the same information input and the same processing steps, the same digital sequence will be obtained), which provides a suitable data basis for generating after-sales authorization codes.

[0120] In this embodiment, this digital sequence can be directly used as the after-sales authorization code, or the digital sequence can be further subjected to simple encoding processing, such as using a specific base conversion (such as decimal to hexadecimal, etc.), adding a check digit (used to verify the accuracy and completeness of the authorization code), etc. to generate the final after-sales authorization code. The after-sales authorization code generated by the above-mentioned unique mathematical processing process has high security and uniqueness. It not only integrates key information closely related to the faulty component and related subjects, but also presents it in a way that is difficult to crack and forge. In the process of after-sales maintenance information transmission, it can effectively ensure the security of information and the supplier's accurate authorization authentication of after-sales maintenance information, ensuring the standardized and reliable operation of the entire after-sales process.

[0121] Through the closely coordinated and interlocking mathematical processing of the above steps, information of different dimensions is cleverly converted into after-sales authority codes, providing an innovative and effective protection mechanism for the after-sales management of smoke alarm failures.

[0122] In one embodiment, based on the feature information, product information and manufacturer information, the step of generating an after-sales authorization code includes:

[0123] The feature information, product information and manufacturer information are converted into binary data strings respectively, and then the binary data strings are concatenated head to tail to form a long binary sequence;

[0124] The long binary sequence is used as initial data, and a chaotic mapping function is used for iterative processing to obtain a set of chaotic sequence values; wherein the parameters of the chaotic mapping function are dynamically adjusted according to the unique identification code of the smoke alarm to increase the randomness and uniqueness of the sequence;

[0125] Perform fractal geometry transformation on chaotic sequence values, construct fractal graphics, and extract boundary contour coordinate point sequences of fractal graphics;

[0126] Calculate the Fourier descriptor based on the boundary contour coordinate point sequence, quantize the Fourier descriptor, and obtain a quantized numerical sequence;

[0127] Perform polynomial fitting on the quantized numerical sequence, take the coefficients of the fitting polynomial as the core data, and use the information diffusion principle to diffuse and encrypt the core data to obtain the encrypted after-sales permission code.

[0128] In this embodiment, by combining and transforming various technical means such as information conversion, chaos mapping, fractal geometry, Fourier descriptor and information diffusion encryption in an innovative way, a simple combination of existing technologies is avoided, and after-sales permission codes are generated from a unique mathematical and graphic processing perspective, which greatly improves the creativity and security of permission code generation, making after-sales permission management more reliable and difficult to crack.

[0129] In one embodiment, based on the feature information, product information and manufacturer information, the step of generating an after-sales authorization code includes:

[0130] Encoding the characteristic information, product information and manufacturer information to obtain a coded string;

[0131] Perform feature extraction on the feature information, product information, and manufacturer information, and extract multiple numbers from the corresponding features;

[0132] Based on the extracted numbers, a simulated graph is generated;

[0133] The encoded character string is discretely distributed on the simulation graphic, and a plurality of target characters are found on the simulation graphic, and characters other than the target characters are used as characters to be selected;

[0134] Calculate the position relationship between the candidate character and each target character, and when the candidate character meets the preset position relationship with any target character, use it as the permission character;

[0135] The various permission characters are combined to obtain the after-sales permission code.

[0136] In this embodiment, feature information, product information, and manufacturer information often contain data of various formats and semantics. In order to facilitate subsequent unified mathematical processing and integration of these information of different natures, they need to be encoded. A general encoding method can be used, such as the common ASCII code. Through encoding, these originally diverse information are converted into a unified encoded string composed of numbers and specific symbols, so that subsequent computer systems or corresponding algorithms can process and calculate them efficiently, and also realize the digital representation of information.

[0137] In the above complex information, numbers often carry key quantitative features, and numbers are more convenient for mathematical operations and the construction of regular logical relationships. For characteristic information, such as the number of past service times of the supplier and the year of quality certification obtained, they can be extracted as numbers; the digital part of the production batch number of the component in the product information and the key values ​​in the specification parameters (such as the rated power value, etc.) are all extractable objects; the numbers in the manufacturer information such as the year of establishment of the manufacturer and the number of employees can also be extracted. By carefully selecting and extracting these numbers from the corresponding features of each information, it is to further explore the valuable key elements that can reflect the essence of the information and facilitate the subsequent construction of graphics and generation of permission codes, focus the information from a broader description form to a more operational digital set, and use the uniqueness and certainty of numbers to create conditions for the generation of permission codes.

[0138] There are many ways to construct simulation graphics using these extracted numbers. For example, the numbers can be used as coordinate parameters of the graphics. If there are many numbers extracted, a two-dimensional coordinate system can be used to determine the coordinate points one by one by taking some of the numbers as the values ​​of the horizontal and vertical coordinates in turn, and then draw lines in a certain order (such as connecting adjacent coordinate points in sequence or according to specific drawing algorithm rules) to form a polygon or a more complex curve graphic; the shape, color, filling method and other visual features of the graphics can also be determined according to the size relationship, parity and other attributes of the numbers (for example, even numbers correspond to filling a certain area, odd numbers correspond to not filling, etc.), to construct a complex simulation graphic composed of a combination of rectangles, circles, irregular graphics, etc. The purpose of generating this simulation graphic is to present the digital information in a visual form with a spatial structure, to provide an intuitive and unique carrier for subsequent operations such as distributing characters on the graphics and filtering permission characters based on graphic features, so that the generation process of the permission code is combined with the structural characteristics of the graphics, increasing the complexity and uniqueness of the permission code generation.

[0139] The previously obtained coded string is discretely distributed on the simulated graphics. A variety of mapping rules can be used, such as dividing the graphics equally according to the perimeter, area, etc., and placing the characters in the coded string in these divided positions in sequence; or according to the sub-regions inside the graphics (if it is a complex graphics with multiple segmented regions), the characters are assigned to different sub-regions in sequence. When determining the target characters, some clear rules can be set in advance, such as defining the characters located at special positions such as the vertex position of the graphics, the intersection position of the symmetry axis of the graphics, and the specific color area as target characters. The selection of these target characters is usually based on the key geometric features of the graphics or the position attributes with obvious distinction. They are in relatively important and iconic positions in the entire graphics. The remaining characters that are not in these special positions are used as candidate characters in order to subsequently screen out the permission characters that meet specific conditions based on the positional relationship between the candidate characters and the target characters. Through this character classification based on the graphic position, the graphic structure characteristics are further introduced to participate in the permission code generation process, increasing the complexity and uniqueness of the permission code generation logic.

[0140] The calculation of positional relationship can be considered from multiple dimensions, such as calculating the distance between the candidate character and the target character on the graph (the Euclidean distance, Manhattan distance and other measurement methods can be used, depending on the nature of the graph and specific needs), to determine whether the candidate character falls within a specific radius with the target character as the center; or starting from the angle relationship, to see whether the angle formed by the line connecting the candidate character and the target character and a certain baseline of the graph (such as the coordinate axis, symmetry axis, etc.) is within the preset angle range; it can also be determined from the perspective of regional attribution whether the sub-region where the candidate character is located satisfies the preset regional relationship such as adjacent and contained with the sub-region where the target character is located. When the candidate character meets these pre-set positional relationship conditions with any target character, it is marked as a permission character. Through such a screening method based on the positional relationship of the graph, the spatial structure and layout characteristics constructed by the simulated graph are utilized, so that the permission characters finally screened out are related to the original information encoding and closely depend on the unique structure of the graph, making the permission code more secure and unpredictable, because only the authorized party who knows the entire graph construction, character distribution and positional relationship setting rules can accurately generate and verify the permission code.

[0141] Combine the selected permission characters in order, which can be determined according to the order in which the characters are selected on the simulation graphic, or according to the position order of the characters in the original encoding string. Through this combination, a complete after-sales permission code is finally formed. The generated after-sales permission code integrates the results of multiple steps such as encoding processing of feature information, product information and manufacturer information, feature extraction, and complex screening based on simulation graphics. It has a high degree of uniqueness and confidentiality, and can be used as an effective permission authentication mark in the process of after-sales maintenance information transmission, ensuring that only legitimate suppliers can process after-sales maintenance information with the correct permission authentication mechanism, thus ensuring the information security and standardization of the entire after-sales process.

[0142] Reference Figure 2 In one embodiment of the present invention, a fault data analysis device for a smoke alarm is provided, comprising:

[0143] An acquisition unit, used to acquire raw data collected by the smoke alarm and extract multi-dimensional features of the raw data;

[0144] A processing unit is used to input the multi-dimensional features into a pre-trained deep neural network to obtain multi-dimensional restored features; wherein the deep neural network encodes and compresses the multi-dimensional features through an autoencoder architecture, and then decodes and restores the multi-dimensional restored features;

[0145] A calculation unit, used for calculating the difference between the multi-dimensional restored features and the standard features to obtain a difference result;

[0146] The analysis unit is used to obtain the fault data in the original data collected by the smoke alarm according to the difference result.

[0147] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0148] Reference Figure 3 In an embodiment of the present invention, a smoke alarm is also provided. The internal structure of the smoke alarm can be as follows: Figure 3As shown. The smoke alarm includes a processor, a memory, a display screen, an input device, a network interface and a database connected via a system bus. The computer-designed processor is used to provide computing and control capabilities. The memory of the smoke alarm includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the smoke alarm is used to store the corresponding data in this embodiment. The network interface of the smoke alarm is used to communicate with an external terminal via a network connection. The computer program is executed by the processor to implement the above method.

[0149] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a part of the structure related to the solution of the present invention, and does not constitute a limitation on the smoke alarm to which the solution of the present invention is applied.

[0150] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0151] In summary, the fault data analysis method, device and smoke alarm provided in the embodiments of the present invention include: obtaining the original data collected by the smoke alarm, extracting the multi-dimensional features of the original data; inputting the multi-dimensional features into a pre-trained deep neural network, and processing to obtain multi-dimensional restored features; wherein the deep neural network encodes and compresses the multi-dimensional features through an autoencoder architecture, and then decodes and restores to obtain multi-dimensional restored features; calculating the difference between the multi-dimensional restored features and the standard features to obtain a difference result; according to the difference result, obtaining the fault data in the original data collected by the smoke alarm. In the present invention, the collected data is deeply analyzed by comprehensively considering various factors, the fault data is accurately identified, and the reliability of the smoke alarm and the troubleshooting efficiency are improved to ensure the stable operation of the fire safety monitoring system.

[0152] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0153] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0154] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for analyzing fault data of a smoke alarm, characterized in that: The following steps are involved: Obtaining raw data collected by the smoke alarm and extracting multi-dimensional features of the raw data; Input the multi-dimensional features into a pre-trained deep neural network, and process them to obtain multi-dimensional restored features; wherein the deep neural network encodes and compresses the multi-dimensional features through an autoencoder architecture, and then decodes and restores them to obtain multi-dimensional restored features; Calculating the difference between the multi-dimensional restored features and the standard features to obtain a difference result; According to the difference result, the fault data in the original data collected by the smoke alarm is obtained.

2. The method for analyzing fault data of a smoke alarm according to claim 1, characterized in that: The raw data includes smoke concentration data, data acquisition timestamp data, temperature and humidity data, and sensor status data.

3. The method for analyzing fault data of a smoke alarm according to claim 1, characterized in that: The multi-dimensional features are input into a pre-trained deep neural network, and processed to obtain multi-dimensional restored features, including: The multi-dimensional features are input into the encoding layer of the autoencoder architecture. The encoding layer adopts a multi-layer fully connected neural network structure. The number of neurons in each layer decreases successively. The features are transformed by a nonlinear activation function, and the high-dimensional multi-dimensional features are mapped to a low-dimensional encoding space to obtain encoded data; The encoded data is input into the decoding layer of the autoencoder architecture. The decoding layer adopts a multi-layer fully connected neural network structure. The number of neurons in each layer increases successively and is symmetrical with the structure of the encoding layer. The encoded data is transformed by a nonlinear activation function opposite to that of the encoding layer, and the low-dimensional encoded data is restored to high-dimensional multi-dimensional restored features.

4. The method for analyzing fault data of a smoke alarm according to claim 1, characterized in that: Calculating the difference between the multi-dimensional restored features and the standard features to obtain a difference result includes: Performing principal component analysis on the multi-dimensional restored features and the standard features respectively, extracting their respective principal component feature vectors, and obtaining a principal component matrix of the multi-dimensional restored features and a principal component matrix of the standard features; The cosine similarity between the multi-dimensional restored feature principal component matrix and the standard feature principal component matrix is ​​calculated to obtain a similarity value sequence of each dimension as the difference result.

5. The method for analyzing fault data of a smoke alarm according to claim 1, characterized in that: The extracting multi-dimensional features of the original data includes: Based on wavelet packet decomposition, the original data is decomposed into multi-layer wavelet packet subspaces, and the energy entropy of the signal in each subspace is calculated as the energy entropy feature, and the correlation coefficient between the signal in each subspace and the original data is calculated as the correlation coefficient feature; Based on a local linear embedding algorithm, nonlinear dimensionality reduction is performed on the original data, and in the low-dimensional space after dimensionality reduction, distance features between data points are calculated, and combined with the Euclidean distance features of the original data in the high-dimensional space to obtain a distance feature combination; Using the original data as an observation sequence, based on a hidden Markov model, the characteristics of a hidden state transition probability matrix and the characteristics of an observation probability distribution under each hidden state are obtained; The energy entropy feature, correlation coefficient feature, distance feature combination, hidden state transition probability matrix feature and observation probability distribution feature are used as multi-dimensional features of the original data.

6. The method for analyzing fault data of a smoke alarm according to claim 1, characterized in that: After obtaining the fault data in the original data collected by the smoke alarm according to the difference result, the method includes: Based on the fault data, obtaining the faulty component of the smoke alarm; Obtaining characteristic information of a supplier of the faulty component; Obtaining product information of the faulty component and manufacturer information of the smoke alarm; Generate an after-sales authorization code based on the feature information, product information and manufacturer information; After-sales maintenance information is sent to the supplier; the after-sales maintenance information carries the after-sales authority code, which is used by the supplier to authenticate the authority of the after-sales maintenance information.

7. The method for analyzing fault data of a smoke alarm according to claim 6, characterized in that: Based on the feature information, product information and manufacturer information, an after-sales authorization code is generated, including: Convert the characteristic information, product information and manufacturer information into preset matrix forms respectively; wherein the characteristic information is constructed into a multi-dimensional sparse matrix, the product information is constructed into a triangular matrix, and the manufacturer information is constructed into a symmetric matrix; Performing matrix multiplication operation on the transformed matrix to obtain a composite matrix, and calculating the eigenvalues ​​and eigenvectors of the composite matrix; With the eigenvalue as the horizontal coordinate and the modulus of the eigenvector as the vertical coordinate, a plurality of plane coordinate points are constructed, and each plane coordinate point is connected to form a continuous and smooth curve; Performing Fourier series expansion on the curve, arranging the coefficients of the preset items in order into a digital sequence; The after-sales authorization code is generated based on the digital sequence.

8. A fault data analysis device for a smoke alarm, characterized in that: include: An acquisition unit, used to acquire raw data collected by the smoke alarm and extract multi-dimensional features of the raw data; A processing unit is used to input the multi-dimensional features into a pre-trained deep neural network to obtain multi-dimensional restored features; wherein the deep neural network encodes and compresses the multi-dimensional features through an autoencoder architecture, and then decodes and restores the multi-dimensional restored features; A calculation unit, used to calculate the difference between the multi-dimensional restored features and the standard features to obtain a difference result; The analysis unit is used to obtain the fault data in the original data collected by the smoke alarm according to the difference result.

9. A smoke alarm, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images