Intelligent home appliance control device and method based on device recognition

Abstract

The application relates to the technical field of Internet of Things equipment state analysis and intelligent network communication, in particular to an intelligent household appliance control device and method based on equipment identification, which comprises an equipment device and a control device, the equipment device comprises a central control node, an environment sound pickup assembly and a household appliance execution node. The control device obtains a voice sound wave signal through a data acquisition module and converts the voice sound wave signal into a digital audio stream signal, a text decoding module performs acoustic decoding on the signal to output an unstructured text stream. A memory calling module extracts historical use memory parameters of a specific node according to the text stream, a tensor deduction module performs tensor fusion calculation on the text stream and the memory parameters, outputs a structured text control instruction containing a target equipment identifier and an action parameter, and a scheduling and issuing module converts the instruction into a control code and issues the control code to corresponding nodes to execute state switching. The application solves the problem that a control system cannot analyze fuzzy semantics and lacks historical use data, leading to the rigid problem of household appliance linkage scheduling.

Description

Technical Field

[0001] This invention relates to the field of IoT device status analysis and intelligent network communication technology, and in particular to an intelligent home appliance control device and method based on device identification. Background Technology

[0002] Smart terminal devices are gradually becoming the entry point for home control networks. Current smart home LANs primarily rely on local microprocessors for device interaction management, with appliance control logic largely built upon basic conditional triggering mechanisms. The local terminal receives user voice input and converts it into basic digital audio signals. The network module then transmits these audio signals to designated home appliance nodes, enabling the appliances to perform start / stop actions based on the control node signals. Multiple home appliances within the smart home system establish network connections via wireless communication technology; a centralized control panel manages network device information, allowing users to issue centralized commands to home appliances.

[0003] Existing intelligent network control architectures suffer from significant technical flaws. Due to the computational bottleneck of local microprocessors, they cannot process unstructured natural language frequencies, requiring users to accurately memorize specific control codes to trigger corresponding hardware relay actions. Furthermore, data sharing between different home appliances is impossible, resulting in isolated data silos. The control system cannot parse ambiguous semantics and lacks historical usage data, causing appliances to rely on specific phrases for triggering and leading to rigid device linkage scheduling. Taking the application scenario of multi-device linkage and fuzzy voice interaction technology in smart homes as an example, when the microphone array acquires natural language sound waves including "feeling a little cold," the local microprocessor system cannot match the instruction structure in the fixed dictionary and is limited by computational power to extract deep action parameters and intentions from the natural language. The lack of user data memory parameters in the control center prevents the control panel from remembering historical operation parameters for specific appliances. Smart air conditioners cannot receive precise control signals from the control module, and window servo mechanisms also cannot receive linkage closing commands, resulting in a mechanical linkage process that ultimately requires tedious resetting of panel values ​​every time appliances are turned on. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent home appliance control device and method based on device identification. This invention solves the technical problem that home appliances rely on specific sentence triggers and the device linkage scheduling is rigid because the control system cannot parse fuzzy semantics and lacks historical usage data.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: In a first aspect, the present invention provides a smart home appliance control device based on device identification, comprising a device device and a control device; The equipment includes a central control node, an environmental sound pickup component, and multiple home appliance execution nodes. The environmental sound pickup component is installed inside the central control node, and each of the multiple home appliance execution nodes establishes a communication link with the central control node through a local area network. The control device establishes a communication connection with the equipment. The control device includes: The data acquisition module, installed within the environmental sound pickup component, is used to acquire speech sound wave signals and convert them into digital audio stream signals. The text decoding module is used to receive digital audio stream signals, perform acoustic decoding on the digital audio stream signals, and output unstructured text streams; The memory retrieval module is used to receive unstructured text streams, send historical record addressing requests to a data storage node based on the unstructured text streams, and extract historical usage memory parameters corresponding to a specific node among multiple home appliance execution nodes. The tensor inference module is used to receive unstructured text streams and historical usage memory parameters, perform tensor fusion calculations on the unstructured text streams and historical usage memory parameters, and output structured text control commands. The structured text control commands include a target device identifier and an action parameter. The scheduling and distribution module is used to receive structured text control commands, convert the structured text control commands into a format to generate control codes, and distribute the control codes to specific nodes. The specific nodes correspond to the target device identifier, so that the specific nodes can switch states according to the control codes.

[0006] Furthermore, the text decoding module of the smart home appliance control device based on device identification described in this invention includes: The acoustic receiving module is used to receive digital audio stream signals and perform frame-segmentation processing on the digital audio stream signals to generate a frequency domain feature matrix. The noise reduction and filtering module is used to receive the frequency domain feature matrix, remove the environmental noise matrix from the frequency domain feature matrix, and output the noise-reduced frequency feature matrix; The text mapping module receives the noise-reduced frequency feature matrix, performs character sequence matching between the noise-reduced frequency feature matrix and a preset vocabulary, and outputs an unstructured text stream.

[0007] Furthermore, the smart home appliance control device based on device identification described in this invention includes a memory retrieval module comprising: The identifier extraction module is used to receive unstructured text streams, extract named entities from the unstructured text streams, and output user identity identifiers. The sequence addressing module is used to receive the user identity identifier, search within the data storage node based on the user identity identifier, and output the historical interaction text sequence associated with the user identity identifier. The parameter separation module is used to receive the historical interaction text sequence, separate the set value for a specific node from the historical interaction text sequence, and determine the set value as the historical usage memory parameter.

[0008] Furthermore, the tensor inference module of the intelligent home appliance control device based on device identification described in this invention includes: The intent conversion module is used to receive unstructured text streams and perform matrix mapping on the unstructured text streams to generate text vector matrices in the intent vector space. The memory conversion module is used to receive historical usage memory parameters and perform matrix mapping on the historical usage memory parameters to generate a preference vector matrix in the preference vector space. The status assessment module is used to receive load feedback data from a specific node and convert the load feedback data into a device status matrix. The tensor operation module receives a text vector matrix, a preference vector matrix, and a device state matrix. It calculates the dot product of the text vector matrix and the preference vector matrix to output a basic intent matrix. It then performs matrix multiplication on the basic intent matrix and the device state matrix to output a fused intent matrix. The instruction generation module is used to receive the fusion intent matrix and map the fusion intent matrix according to a preset mapping table to generate structured text control instructions.

[0009] Furthermore, the intelligent home appliance control device based on device identification described in this invention includes a scheduling and distribution module comprising: The redundancy filtering module is used to receive structured text control commands, remove invalid control characters from the structured text control commands, and output command strings; The time allocation module is used to receive instruction strings, calculate the execution time delay based on the instruction strings, add corresponding time markers to the instruction strings, and generate instruction strings with markers. The encoding generation module is used to receive the instruction string with identification and convert the instruction string with identification into control code according to the preset communication protocol. The protocol encapsulation module is used to receive control codes and encapsulate the control codes into Ethernet control data frames; The network push module is used to receive Ethernet control data frames and send them to specific nodes via the local area network.

[0010] Furthermore, the intelligent home appliance control device based on device identification described in this invention is divided into a local computing node and a cloud computing node. Local computing nodes are integrated within the central control node, and local computing nodes include data acquisition modules; The cloud computing node is deployed at a remote location on a network. The cloud computing node includes a text decoding module, a memory retrieval module, a tensor inference module, and a scheduling and distribution module. The local computing node establishes a data connection with the cloud computing node through a network transmission path. The local computing node sends digital audio stream signals to the cloud computing node, and the cloud computing node sends control codes to the local computing node.

[0011] Furthermore, the local computing node of the intelligent home appliance control device based on device identification described in this invention further includes: The audio monitoring module is used to collect an ambient sound audio segment; The mode determination module is used to receive ambient sound audio segments, compare the ambient sound audio segments with the preset wake-up acoustic spectrum, and output the wake-up determination result. The activation control module is used to receive the wake-up determination result. When the wake-up determination result is consistent, it sends an activation command to the data acquisition module and establishes a network transmission path.

[0012] Furthermore, the local computing node of the intelligent home appliance control device based on device identification described in this invention further includes: The delay detection module is used to collect communication delay time values ​​within the network transmission path; The signal extraction module is used to receive the communication delay time value. When the communication delay time value is greater than a preset delay time threshold, the digital audio stream signal is extracted. The local matching module is used to receive the extracted digital audio stream signal, compare the extracted digital audio stream signal with the preset vocabulary matrix stored locally, generate local backup control commands, and send the local backup control commands to specific nodes.

[0013] Furthermore, the intelligent home appliance control device based on device identification according to the present invention further includes: The data unpacking node is installed inside multiple home appliance execution nodes. It is used to receive a control data frame, which includes control codes. The control data frame is unpacked to separate the control codes. Level conversion node is used to receive control codes and convert them into analog voltage signals; A state adjustment node is used to receive analog voltage signals and control the displacement of an internal actuator within a specific node based on the voltage parameters of the analog voltage signals.

[0014] Secondly, the smart home appliance control method based on device identification provided by the present invention is applied to the smart home appliance control device based on device identification as described above, comprising: Step 1: Acquire the speech sound wave signal and convert it into a digital audio stream signal; Step 2: Receive the digital audio stream signal, perform acoustic decoding on the digital audio stream signal, and output an unstructured text stream; Step 3: Receive an unstructured text stream, send a history addressing request to a data storage node based on the unstructured text stream, and extract the history usage memory parameters corresponding to a specific node among multiple home appliance execution nodes; Step 4: Receive the unstructured text stream and the historical usage memory parameters, perform tensor fusion calculation on the unstructured text stream and the historical usage memory parameters, and output the structured text control command. The structured text control command includes a target device identifier and an action parameter. Step 5: Receive structured text control instructions, convert the structured text control instructions into a format to generate control codes, and send the control codes to specific nodes, where the specific nodes correspond to the target device identifier, so that the specific nodes can switch states according to the control codes.

[0015] Beneficial effects of this invention: The intelligent home appliance control device and method based on device identification provided by this invention establishes a complete hardware closed-loop link for data processing and action execution. During system operation, the text decoding module receives the digital audio stream signal extracted from the user and performs acoustic decoding, thereby outputting an unstructured text stream.

[0016] Upon receiving the unstructured text stream, the memory retrieval module sends an addressing request to the data storage node based on the unstructured text stream to extract the historical usage memory parameters corresponding to a specific node. Subsequently, the tensor inference module synchronously receives the unstructured text stream and the historical usage memory parameters, performs tensor fusion calculation on the unstructured text stream and the historical usage memory parameters, and outputs a structured text control command containing the target device identifier and action parameters.

[0017] Unstructured text streams and historical usage memory parameters undergo matrix fusion and collaborative computation within the computational core of the tensor inference module. This collaborative computation process overcomes the shortcomings of existing control systems, which, due to their inability to parse fuzzy semantics and lack of historical usage data, cause home appliances to be highly dependent on specific fixed sentence triggers.

[0018] The scheduling and dispatching module, which is in the instruction issuing stage, receives structured text control instructions, converts the structured text control instructions into a format to generate control codes, and sends the control codes to a specific node corresponding to the target device identifier. The specific node receives the control codes and precisely controls the displacement of the actuators inside the specific node according to the control codes, thereby completing the state switching of the terminal actuators.

[0019] By leveraging the data flow logic between various digital computing modules and the underlying home appliance execution nodes, the control device can adaptively generate underlying control codes based on converged historical usage memory parameters. Simultaneously, the collaborative data interaction mechanism between the cloud computing nodes and the edge execution hardware constructs a multi-dimensional data sharing channel between multiple home appliance execution nodes, effectively solving the technical problem of rigid and mechanical device function scheduling in multi-device linkage scenarios in smart homes. Attached Figure Description

[0020] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the intelligent home appliance control method based on device identification according to the present invention. Detailed Implementation

[0022] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] In a first aspect, the present invention provides a smart home appliance control device based on device identification, comprising a device device and a control device; The equipment includes a central control node, an environmental sound pickup component, and multiple home appliance execution nodes. The environmental sound pickup component is installed inside the central control node, and each of the multiple home appliance execution nodes establishes a communication link with the central control node through a local area network. The control device establishes a communication connection with the equipment. The control device includes: The data acquisition module, installed within the environmental sound pickup component, is used to acquire speech sound wave signals and convert them into digital audio stream signals. The text decoding module is used to receive digital audio stream signals, perform acoustic decoding on the digital audio stream signals, and output unstructured text streams; The memory retrieval module is used to receive unstructured text streams, send historical record addressing requests to a data storage node based on the unstructured text streams, and extract historical usage memory parameters corresponding to a specific node among multiple home appliance execution nodes. The tensor inference module is used to receive unstructured text streams and historical usage memory parameters, perform tensor fusion calculations on the unstructured text streams and historical usage memory parameters, and output structured text control commands. The structured text control commands include a target device identifier and an action parameter. The scheduling and distribution module is used to receive structured text control commands, convert the structured text control commands into a format to generate control codes, and distribute the control codes to specific nodes. The specific nodes correspond to the target device identifier, so that the specific nodes can switch states according to the control codes.

[0024] The underlying data processing path of the text decoding module begins with the acoustic receiving module. The acoustic receiving module receives the digital audio stream signal, applies a Hamming window to truncate the digital audio stream signal along the time axis, and performs frame segmentation processing according to preset frame length and frame shift parameters; then, it extracts acoustic features through the Fast Fourier Transform algorithm to generate a frequency domain feature matrix.

[0025] The noise reduction and filtering module receives the aforementioned frequency domain feature matrix and uses spectral subtraction to identify and remove stationary noise spectrum sequences from the background environment. By removing the environmental noise matrix from the frequency domain feature matrix, it outputs a noise-reduced frequency feature matrix with improved signal-to-noise ratio.

[0026] The text mapping module receives the noise-reduced frequency feature matrix, inputs it into the Hidden Markov Acoustic Model, and performs character sequence matching with a preset vocabulary. Through audio feature alignment to specific Chinese character sequences, it ultimately outputs an unstructured text stream reflecting the user's intentions in smart home multi-device linkage and fuzzy voice interaction technology application scenarios.

[0027] When the memory retrieval module performs the information tracing step, the identifier extraction module first receives the unstructured text stream. It then uses a conditional random field algorithm to extract named entities from the unstructured text stream, removes irrelevant modal particles, and outputs a unique user identifier.

[0028] The sequence addressing module receives the user's identity identifier, performs a primary key search within the data storage nodes of the relational database based on the user's identity identifier, and outputs a historical interaction text sequence that is associated with the user's identity identifier by a time identifier.

[0029] The parameter separation module receives a sequence of historical interaction texts and uses a sliding window algorithm to extract the set values ​​for specific nodes from the sequence. Specifically, the process involves determining the frequency distribution of the extracted set values; if a set value appears five or more times in the historical 30-day interaction records, it is determined that the set value meets the habit convergence criterion. The system then identifies the set values ​​that meet the criteria as historical usage memory parameters. For example, in the application scenario of a smart refrigerator or washing machine, the separated 4-degree Celsius refrigeration temperature value or standard washing time value is designated as a historical usage memory parameter.

[0030] When the tensor inference module processes multidimensional data calculations, the intent conversion module receives an unstructured text stream and uses a word embedding algorithm to perform matrix mapping on the unstructured text stream, generating a text vector matrix in a 512-dimensional intent vector space. Simultaneously, the memory conversion module receives historical usage memory parameters and performs matrix mapping on these parameters to generate a preference vector matrix in a preference vector space.

[0031] The state assessment module receives load feedback data from sensors inside the housing of a specific node device, normalizes the load feedback data, and converts it into a device state matrix. The tensor operation module receives a text vector matrix, a preference vector matrix, and the device state matrix. Within the microprocessor core unit of the tensor operation module, it calculates the dot product of the text vector matrix and the preference vector matrix to output a basic intent matrix, and then performs matrix multiplication between the basic intent matrix and the device state matrix.

[0032] During the matrix calculation process described above, if the absolute value of an element in the equipment status matrix reflecting the compressor or motor's operating load is detected to be greater than 85% of the preset safe load limit, the microprocessor core unit performs a decay multiplication operation on the corresponding action step size in the basic intent matrix, outputting a fused intent matrix that takes into account both user intent and hardware operating boundaries. Finally, the instruction generation module receives the fused intent matrix and maps it according to a preset mapping table to generate structured text control instructions.

[0033] The underlying communication mechanism of the scheduling and distribution module involves instruction processing and network encapsulation steps. The redundancy filtering module receives structured text control instructions, traverses the structured text control instructions using regular expressions, and outputs a compact instruction string by removing whitespace, newline characters, and invalid control characters from the structured text control instructions.

[0034] The time allocation module receives the instruction string, reads the network time protocol synchronization messages within the local area network, and calculates the execution time delay based on the action type involved in the instruction string. The time allocation module then adds a time stamp containing future timing node information to the instruction string, generating an identifiable instruction string.

[0035] The encoding generation module receives the identifiable instruction string and converts it into control code composed of hexadecimal machine code according to the encoding rules of the preset communication protocol. Subsequently, the protocol encapsulation module receives the control code, appends the source and destination media access control addresses to it, and encapsulates the control code into an Ethernet control data frame. The network push module receives the Ethernet control data frame and sends it to a specific node through the local area network's transmission control protocol channel.

[0036] The control device is architecturally divided into local computing nodes and cloud computing nodes, with a clear hardware collaboration mechanism between the two. The local computing nodes are integrated inside the central control node's housing, and their hardware configuration manifests as a smart home control panel circuit board equipped with a microprocessor. The local computing node includes a data acquisition module for performing analog-to-digital signal conversion.

[0037] Cloud computing nodes are deployed at remote locations within the network, and their hardware configuration consists of server clusters with massively parallel processing graphics processors. Each cloud computing node server contains a text decoding module, a memory retrieval module, a tensor inference module, and a scheduling and distribution module, specifically designed to perform computationally intensive neural network operations.

[0038] Local computing nodes and cloud computing nodes establish a full-duplex data connection via network transmission paths. Local computing nodes send digital audio stream signals to cloud computing nodes along the uplink; cloud computing nodes send control codes to local computing nodes along the downlink, completing end-to-end computing power offloading and control code transmission.

[0039] During the wake-up response process, the low-power digital signal processor (DSP) inside the audio monitoring module hardware circuit remains normally powered on. The DSP is used to acquire ambient sound audio segments and store them in a circular buffer.

[0040] The mode determination module receives ambient sound audio segments and extracts sound features using a Mel-frequency cepstral coefficient extraction algorithm. The module then compares these ambient sound audio segments with a pre-set wake-up acoustic spectrum in the local flash memory chip. The comparison process employs a dynamic time warping algorithm, outputting a wake-up determination result that reflects the similarity of the acoustic feature distribution.

[0041] The activation control module receives the wake-up determination result and calculates the distance cost value generated by feature comparison. When the distance cost value is less than or equal to the wake-up cost value threshold of 0.45, the wake-up determination result is considered a match. The system sends a high-level enable command to the data acquisition module via a general-purpose input / output pin and wakes up the main network communication chip to establish a network transmission path.

[0042] The local computing node has backup control logic in situations where network data transmission is limited. The latency detection module first sends a Transmission Control Protocol (TCP) test message to a remote location on the network to collect a communication latency value within the network transmission path.

[0043] The signal extraction module receives the communication delay time value and compares it with the pre-stored network outage threshold parameter. When the communication delay time value is greater than or equal to a preset delay time threshold of 200 milliseconds, it is determined that the network transmission path is congested; subsequently, the signal extraction module intercepts and extracts the digital audio stream signal through the local computing node's motherboard data bus.

[0044] The local matching module receives the extracted digital audio stream signal, compares it with a locally stored preset vocabulary matrix, and generates local backup control commands. The local matching module does not process the commands through a cloud computing node; instead, it directly sends the local backup control commands to a specific node connected to the local area network.

[0045] When the device performs hardware action conversion at the end, the data unpacking node is assembled in the housing of multiple home appliance execution node devices. The data unpacking node receives control data frames transmitted via the local area network, reads the header and checksum sequence of the control data frame; then it performs data unpacking processing on the control data frame, strips off the network layer protocol packet header, and separates the core control code.

[0046] The level conversion node receives the control code and converts the hexadecimal control code into a continuous analog voltage signal in the range of zero to five volts through a digital-to-analog converter chip.

[0047] The state adjustment node receives an analog voltage signal and outputs a matching drive current based on the voltage parameter value of the analog voltage signal. The drive current controls the displacement of an actuator within a specific node. For example, in a washing machine water filling scenario, it adjusts the opening stroke parameter of the solenoid water valve actuator, completing the state closed-loop transition from network digital commands to mechanical actions.

[0048] Secondly, please refer to Figure 1 The intelligent home appliance control method based on device identification provided by this invention is applied to the intelligent home appliance control device based on device identification as described above, comprising: Step 1: Acquire the speech sound wave signal and convert it into a digital audio stream signal; Step 2: Receive the digital audio stream signal, perform acoustic decoding on the digital audio stream signal, and output an unstructured text stream; Step 3: Receive an unstructured text stream, send a history addressing request to a data storage node based on the unstructured text stream, and extract the history usage memory parameters corresponding to a specific node among multiple home appliance execution nodes; Step 4: Receive the unstructured text stream and the historical usage memory parameters, perform tensor fusion calculation on the unstructured text stream and the historical usage memory parameters, and output the structured text control command. The structured text control command includes a target device identifier and an action parameter. Step 5: Receive structured text control instructions, convert the structured text control instructions into a format to generate control codes, and send the control codes to specific nodes, where the specific nodes correspond to the target device identifier, so that the specific nodes can switch states according to the control codes.

[0049] In smart home multi-device linkage and fuzzy voice interaction technology applications, the equipment includes a central control node located in the living room, and smart air conditioners and smart windows serving as execution nodes for multiple home appliances. A low-power digital signal processor (DSP) is deployed on the integrated circuit board of the audio monitoring module; the DSP remains normally active, collecting ambient sound audio segments and storing them in a circular data buffer. An acoustic wake-up model is pre-loaded into the memory of the pattern determination module.

[0050] The construction process of the acoustic wake-up model involves pre-collecting 10,000 positive wake-up speech samples and 50,000 negative environmental noise samples to build a lightweight convolutional neural network. The backpropagation algorithm is used to adjust the weights of the convolutional neural network, and after training convergence, feature maps are extracted to generate a pre-defined wake-up acoustic map.

[0051] The pattern determination module receives ambient sound audio segments and uses a Mel-frequency cepstral coefficient extraction algorithm to convert these segments into feature sequences. These feature sequences are then input into the acoustic wake-up model. The acoustic wake-up model employs a dynamic time warping algorithm to compare the ambient sound audio segments with a preset wake-up acoustic spectrum using distance-cost comparison, outputting a wake-up determination result reflecting the similarity of feature distributions.

[0052] The activation control module receives the wake-up determination result. When the distance cost of the wake-up determination result is less than the wake-up cost threshold, the system determines that the wake-up determination result is a match. Subsequently, the activation control module sends a high-level enable command to the data acquisition module via a general-purpose input / output pin and activates the main network communication chip to establish a network transmission path.

[0053] The data acquisition module acquires voice signals such as "feeling a bit cold" emitted by the user, and converts these signals into digital audio streams using an analog-to-digital converter. The text decoding module's flash memory chip houses a Hidden Markov Model (HMM). The HMM is constructed by accessing a 5,000-hour open-source Mandarin Chinese speech dataset, combining it with a triphone Gaussian mixture model for forced alignment training, thereby generating an acoustic state transition probability matrix that records changes in phoneme probabilities.

[0054] The acoustic receiving module receives the digital audio stream signal, truncates it along the time axis using a Hamming window, performs frame segmentation according to preset frame length and frame shift parameters, and generates a frequency domain feature matrix using a Fast Fourier Transform algorithm. Subsequently, the noise reduction and filtering module receives the frequency domain feature matrix, uses spectral subtraction to remove stationary noise spectrum sequences, and outputs a noise-reduced frequency feature matrix with improved signal-to-noise ratio by removing the ambient noise matrix from the frequency domain feature matrix.

[0055] The text mapping module receives the noise-reduced frequency feature matrix and inputs it into the Hidden Markov Acoustic Decoding Model (HMM). The HMM uses the Viterbi decoding algorithm to search for the maximum likelihood path in the acoustic state transition probability matrix, matches the noise-reduced frequency feature matrix with a preset vocabulary, and finally outputs an unstructured text stream that reflects the user's ambiguous expression intent.

[0056] The conditional random field entity extraction model is loaded into the computing chip of the memory retrieval module. When constructing the conditional random field entity extraction model, a smart home historical corpus labeled with personnel identity information and home appliance names is used. The feature weight vector is trained by the maximum likelihood estimation method to form a conditional random field entity extraction model with time-series label classification capabilities.

[0057] The identifier extraction module receives an unstructured text stream and inputs it into a Conditional Random Field (CRF) entity extraction model. The CRF entity extraction model calculates the label transition probability of each character combination in the unstructured text stream, performs named entity extraction on the unstructured text stream, and outputs a user identifier with a unique identification code.

[0058] The sequence addressing module receives the user's identity identifier, performs a primary key search in the data storage nodes of the relational database based on the user's identity identifier, and outputs a historical interaction text sequence that is associated with the user's identity identifier by a time identifier.

[0059] The parameter separation module receives a sequence of historical interaction texts and uses a sliding time window algorithm to extract the set values ​​for specific input nodes from the historical interaction text sequence. Specifically, the extraction logic involves statistically analyzing the distribution frequency of the set values ​​in the historical interaction records over the past thirty days. When the distribution frequency is greater than five times, the converged set value is determined as the historical usage memory parameter. In the application scenario of "feeling a little cold," the historical usage memory parameters separated by the parameter separation module are the 26-degree Celsius temperature value of the smart air conditioner and the closed status value of the smart window.

[0060] The Tensor Inference module's graphics processor runs a large language feature representation model. The construction process of the large language feature representation model is based on a transformer network architecture with a self-attention mechanism. It takes a massive human-computer interaction instruction dataset as input, performs unsupervised pre-training using a masked language modeling task, and fine-tunes it through gradient descent to form a large language feature representation model that supports multi-dimensional matrix output.

[0061] The intent transformation module receives an unstructured text stream and inputs it into a large language feature representation model. The multi-head attention layer of the large language feature representation model performs matrix mapping on the unstructured text stream, generating a text vector matrix in a 512-dimensional intent vector space. Simultaneously, the memory transformation module receives historical usage memory parameters and inputs them into the large language feature representation model for matrix mapping, generating a preference vector matrix in a preference vector space.

[0062] The state assessment module receives load feedback data from sensors in the housing of a specific node device, such as the current operating current value of a smart air conditioner compressor, and normalizes the load feedback data to convert it into a device state matrix. The tensor operation module receives a text vector matrix, a preference vector matrix, and a device state matrix. In the operation core, it calculates the dot product of the text vector matrix and the preference vector matrix to output a basic intent matrix, and then performs matrix multiplication between the basic intent matrix and the device state matrix to output a fused intent matrix.

[0063] The instruction generation module receives the fusion intent matrix, maps the fusion intent matrix to characters according to a preset mapping table, and generates structured text control instructions. The structured text control instructions include target device identifiers pointing to the smart air conditioner and smart window, as well as action parameters for adjusting the temperature to 26 degrees Celsius and performing a closing action.

[0064] The underlying communication mechanism of the scheduling and distribution module handles the network command flow. The redundancy filtering module receives structured text control commands, uses regular expressions to traverse the structured text control commands, and removes whitespace characters, newline characters, and invalid control characters from the structured text control commands, thereby outputting a compressed command string.

[0065] The time allocation module receives the instruction string, reads the network time protocol synchronization messages in the local area network, and calculates the execution time delay required for the smart air conditioner and smart window to complete their actions based on the instruction string. The time allocation module adds a time identifier containing absolute time information to the instruction string, generating an identifiable instruction string to unify the action pace of multiple devices.

[0066] The encoding generation module receives the identifiable instruction string and converts it into control code composed of hexadecimal machine code according to the encoding rules of the preset communication protocol. The protocol encapsulation module receives the control code, appends a media access control address to the control code, and encapsulates the control code into an Ethernet control data frame. The network push module receives the Ethernet control data frame and sends the Ethernet control data frame to a specific node through the network port of the local area network.

[0067] The device's end-point execution performs edge hardware action conversion. Data unpacking nodes are mounted within the metal casings of multiple appliance execution nodes; these nodes receive control data frames and read their headers and checksum sequences. By unpacking the control data frames, the data unpacking nodes strip away the network layer protocol headers, extracting the core control code.

[0068] The level conversion node receives the control code and converts the hexadecimal control code into a continuous analog voltage signal in the range of zero to five volts via a digital-to-analog converter chip. The status adjustment node receives the analog voltage signal and outputs a matched drive current based on the voltage parameter value of the analog voltage signal.

[0069] The drive current controls the displacement of the actuator in a specific node. Specifically, it drives the stepper motor of the smart window to rotate to the closed stroke position and simultaneously adjusts the opening and closing angle of the smart air conditioner expansion valve, ultimately completing the closed-loop conversion from cloud digital commands to mechanical actions.

[0070] When the network transmission path quality is limited, the local computing node triggers the network outage backup control logic. The delay detection module sends a Transmission Control Protocol (TCP) test message to a remote location on the network to collect the communication delay time value in the network transmission path. The signal extraction module receives the communication delay time value and compares it with the pre-stored network outage boundary parameter.

[0071] When the communication delay exceeds a preset delay threshold, the system determines that the network link is congested. In this congested state, the signal extraction module extracts the digital audio stream signal via the motherboard data bus of the local computing node. The local matching module receives the extracted digital audio stream signal and compares it with a preset vocabulary matrix pre-programmed in the local read-only memory using a dynamic time warping algorithm to generate a local backup control command.

[0072] Ultimately, the local matching module bypasses the cloud processing stage and directly sends local backup control commands containing basic start-stop actions to specific nodes, maintaining basic smart home appliance start-stop control capabilities.

Claims

1. A smart home appliance control device based on device identification, characterized in that, This includes equipment and control devices; The equipment includes a central control node, an environmental sound pickup component, and multiple home appliance execution nodes. The environmental sound pickup component is installed inside the central control node, and each of the multiple home appliance execution nodes establishes a communication link with the central control node through a local area network. The control device establishes a communication connection with the equipment. The control device includes: The data acquisition module, installed within the environmental sound pickup component, is used to acquire speech sound wave signals and convert them into digital audio stream signals. The text decoding module is used to receive digital audio stream signals, perform acoustic decoding on the digital audio stream signals, and output unstructured text streams; The memory retrieval module is used to receive unstructured text streams, send historical record addressing requests to a data storage node based on the unstructured text streams, and extract historical usage memory parameters corresponding to a specific node among multiple home appliance execution nodes. The tensor inference module is used to receive unstructured text streams and historical usage memory parameters, perform tensor fusion calculations on the unstructured text streams and historical usage memory parameters, and output structured text control commands. The structured text control commands include a target device identifier and an action parameter. The scheduling and distribution module is used to receive structured text control commands, convert the structured text control commands into a format to generate control codes, and distribute the control codes to specific nodes. The specific nodes correspond to the target device identifier, so that the specific nodes can switch states according to the control codes.

2. The intelligent home appliance control device based on device identification according to claim 1, characterized in that, The text decoding module includes: The acoustic receiving module is used to receive digital audio stream signals and perform frame-segmentation processing on the digital audio stream signals to generate a frequency domain feature matrix. The noise reduction and filtering module is used to receive the frequency domain feature matrix, remove the environmental noise matrix from the frequency domain feature matrix, and output the noise-reduced frequency feature matrix; The text mapping module receives the noise-reduced frequency feature matrix, performs character sequence matching between the noise-reduced frequency feature matrix and a preset vocabulary, and outputs an unstructured text stream.

3. The intelligent home appliance control device based on device identification according to claim 2, characterized in that, The memory retrieval module includes: The identifier extraction module is used to receive unstructured text streams, extract named entities from the unstructured text streams, and output user identity identifiers. The sequence addressing module is used to receive the user identity identifier, search within the data storage node based on the user identity identifier, and output the historical interaction text sequence associated with the user identity identifier. The parameter separation module is used to receive the historical interaction text sequence, separate the set value for a specific node from the historical interaction text sequence, and determine the set value as the historical usage memory parameter.

4. The intelligent home appliance control device based on device identification according to claim 3, characterized in that, The tensor derivation module includes: The intent conversion module is used to receive unstructured text streams and perform matrix mapping on the unstructured text streams to generate text vector matrices in the intent vector space. The memory conversion module is used to receive historical usage memory parameters and perform matrix mapping on the historical usage memory parameters to generate a preference vector matrix in the preference vector space. The status assessment module is used to receive load feedback data from specific nodes and convert the load feedback data into a device status matrix. The tensor operation module receives a text vector matrix, a preference vector matrix, and a device state matrix. It calculates the dot product of the text vector matrix and the preference vector matrix to output a basic intent matrix. It then performs matrix multiplication on the basic intent matrix and the device state matrix to output a fused intent matrix. The instruction generation module is used to receive the fusion intent matrix and map the fusion intent matrix according to a preset mapping table to generate structured text control instructions.

5. The intelligent home appliance control device based on device identification according to claim 4, characterized in that, The scheduling and distribution module includes: The redundancy filtering module is used to receive structured text control commands, remove invalid control characters from the structured text control commands, and output command strings; The time allocation module is used to receive instruction strings, calculate the execution time delay based on the instruction strings, add corresponding time markers to the instruction strings, and generate instruction strings with markers. The encoding generation module is used to receive the instruction string with identification and convert the instruction string with identification into control code according to the preset communication protocol. The protocol encapsulation module is used to receive control codes and encapsulate the control codes into Ethernet control data frames; The network push module is used to receive Ethernet control data frames and send them to specific nodes via the local area network.

6. The intelligent home appliance control device based on device identification according to claim 5, characterized in that, The control device is divided into local computing nodes and cloud computing nodes; Local computing nodes are integrated within the central control node, and local computing nodes include data acquisition modules; The cloud computing node is deployed at a remote location on a network. The cloud computing node includes a text decoding module, a memory retrieval module, a tensor inference module, and a scheduling and distribution module. The local computing node establishes a data connection with the cloud computing node through a network transmission path. The local computing node sends digital audio stream signals to the cloud computing node, and the cloud computing node sends control codes to the local computing node.

7. The intelligent home appliance control device based on device identification according to claim 6, characterized in that, Local compute nodes also include: The audio monitoring module is used to collect an ambient sound audio segment; The mode determination module is used to receive ambient sound audio segments, compare the ambient sound audio segments with the preset wake-up acoustic spectrum, and output the wake-up determination result. The activation control module is used to receive the wake-up determination result. When the wake-up determination result is consistent, it sends an activation command to the data acquisition module and establishes a network transmission path.

8. The intelligent home appliance control device based on device identification according to claim 6, characterized in that, Local compute nodes also include: The delay detection module is used to collect communication delay time values ​​within the network transmission path; The signal extraction module is used to receive the communication delay time value. When the communication delay time value is greater than a preset delay time threshold, the digital audio stream signal is extracted. The local matching module is used to receive the extracted digital audio stream signal, compare the extracted digital audio stream signal with the preset vocabulary matrix stored locally, generate local backup control commands, and send the local backup control commands to specific nodes.

9. The intelligent home appliance control device based on device identification according to claim 1, characterized in that, The equipment also includes: The data unpacking node is installed inside multiple home appliance execution nodes. It is used to receive a control data frame, which includes control codes. The control data frame is unpacked to separate the control codes. Level conversion node is used to receive control codes and convert them into analog voltage signals; A state adjustment node is used to receive analog voltage signals and control the displacement of an internal actuator within a specific node based on the voltage parameters of the analog voltage signals.

10. A smart home appliance control method based on device identification, applied to the smart home appliance control device based on device identification as described in any one of claims 1 to 9, characterized in that, include: Step 1: Acquire the speech sound wave signal and convert it into a digital audio stream signal; Step 2: Receive the digital audio stream signal, perform acoustic decoding on the digital audio stream signal, and output an unstructured text stream; Step 3: Receive an unstructured text stream, send a history addressing request to a data storage node based on the unstructured text stream, and extract the history usage memory parameters corresponding to a specific node among multiple home appliance execution nodes; Step 4: Receive the unstructured text stream and the historical usage memory parameters, perform tensor fusion calculation on the unstructured text stream and the historical usage memory parameters, and output the structured text control command. The structured text control command includes a target device identifier and an action parameter. Step 5: Receive structured text control instructions, convert the structured text control instructions into a format to generate control codes, and send the control codes to specific nodes, where the specific nodes correspond to the target device identifier, so that the specific nodes can switch states according to the control codes.

Images